Machine Learning for the Control and Monitoring

of Electric Machine Drives: Advances and Trends

Shen Zhang, Oliver Wallscheid, and Mario Porrmann

Abstract—This review paper systematically summarizes the

existing literature on utilizing machine learning (ML) techniques
for the control and monitoring of electric machine drives. It
arXiv:2110.05403v2 [eess.SY] 18 Feb 2023

is anticipated that with the rapid progress in learning algo-

rithms and specialized embedded hardware platforms, machine
learning-based data-driven approaches will become standard
tools for the automated high-performance control and monitoring
of electric drives. Additionally, this paper also provides some
outlook toward promoting its widespread application in the
industry with a focus on deploying ML algorithms onto embedded
system-on-chip (SoC) field-programmable gate array (FPGA)
devices.
Index Terms—Machine learning; electric machine drives; deep
learning; reinforcement learning; embedded systems; FPGA. Fig. 1. A vision of future electric machine drives powered by machine
learning.

I. I NTRODUCTION
The motor control community is well-informed on the boom parallel architectures, many existing ML implementations in
of machine learning (ML) after the modern back-propagation electric machine drives were based on slow and sequentially
paper was first published in 1986 [1], which is evident by executed digital signal processors (DSP) prior to the deep
the work that appeared three years later on training a neural learning era, although in some cases multiple DSPs were also
network offline to mimic the behavior of hysteresis current used to increase the execution speed. Embedded platforms
controllers in a three-phase PWM inverter [2]. This work is such as field-programmable gate arrays (FPGA) that excel at
later followed by a series of pioneering efforts in the early parallel processing, were not matured technologies at the time
1990s on general voltage-fed AC machines [3], [4], induction and had limited use.
machines [5]–[15], DC machines [16], [17], synchronous It is worth noting that hardware limitations are still the
machines [18], and switched reluctance machines [19]. In main bottleneck for ML applications in electric machine drives
addition to the broad interest in applying ML to motor drive even today. This remains a major problem particularly in the
control, such technologies, especially concerning classification industrial world due to the high-frequency update rates1 of
or regression techniques, have also found their presence in the ML algorithms required for motor drive online applications in
condition monitoring and fault diagnosis on various types of combination with cost-oriented, computationally-constrained
electric machines [20]–[27]. embedded hardware. This hardware constraint has further
Around that time, the frontier of power electronics gradually impeded the deployment of ML algorithms in machine drives
advanced with the advent of ML models such as neural and resulted in insufficient performance in their identification
networks, which have emerged as the most important area and control. It is envisioned in [30] that with the present trend
for complex system identification, control, and estimation in of FPGA development and as the ML technology matures, “in-
power electronics and motor drives [28]. However, it was telligent control and estimation (particularly based on neural
also concluded that “in spite of the technology advancement, networks) will find increasing acceptance in power electronics,
currently, industrial applications of neural networks in power particularly in the robust control of drives” [31], and they are
electronics appear to be very few” [29]. expected to have widespread applications in the industry [32]–
While ML applications always targeted the fastest avail- [34].
The past decade has marked an incredibly fast-paced and
able hardware platforms, especially focusing on (massively)
innovative period in the history of ML [35]. Spurred by the
Shen Zhang was with the School of Electrical and Computer Engineering, development of ever-more-powerful computing platforms and
Georgia Institute of Technology, Atlanta, GA 30332, USA. He is now with the increased availability of big data, ML has successfully
Joby Aviation, San Carlos, CA 94070, USA (e-mail: [email protected]).
Oliver Wallscheid is with the Department of Automatic Control, Pader-
tackled many previously intractable problems, especially in
born University, 33098 Paderborn, Germany (e-mail: oliver.wallscheid@uni- computer vision with the convolutional neural network (CNN)
paderborn.de).
Mario Porrmann is with the Institute of Computer Science, Os- 1 While ML algorithms for real-time video processing only require to run
nabrüeck University, Wachsbleiche 27, 49090 Osnabrück, Germany (email: at a couple of ten hertz, typical motor control update frequencies are in the
[email protected]). range of ten kilohertz, so multiple orders of magnitude higher.

1
 Unsupervised Clustering extensively applied to the control and monitoring of electric
Learning
Process and interpret Dimension Reduction machine drives, as well as the estimation of models or model
data based only
on the input parameters associated with the drives.
2) Unsupervised Learning: Unsupervised learning is a type
Supervised Regression
Learning of algorithm that processes and interprets data based solely
Machine Develop models Classification
Learning to map input
on input. Classical tasks in unsupervised learning include
and output data
clustering, dimensionality reduction, and anomaly detection. In
Reinforcement Single-Agent
contrast to supervised learning, unsupervised learning models
Learning
Multi-Agent
work on their own to discover the inherent structure of
Learn optimal control
actions to maximize
long-term reward
unlabeled data. Unsupervised learning can also be used as
an auxiliary pre-processing step in order to apply feature
engineering for supervised learning [80].
Fig. 2. Disciplines of machine learning [80].
3) Reinforcement Learning: As a subcategory of ML, rein-
forcement learning (RL) aims to solve a variety of decision-
[36] and in natural language processing with the Transformer making and control problems in a data-driven manner. Specifi-
architecture [37]. ML has also been applied to and is in cally, RL is able to learn in a trial-and-error way and does not
the process of transforming many real-world applications, require explicit human labeling or supervision of each data
including entertainment, healthcare, fraud detection, virtual sample. Instead, it requires a well-defined reward function
assistants, and autonomous vehicles. Hardware platforms in- to obtain reward signals throughout the learning process.
cluding GPUs and FPGA fabric can also achieve very good The core of RL is to learn optimal control actions in an
parallel computing performance with architecture customiza- environment to maximize the long-term cumulative reward
tion [38], which is intrinsically well-suited for the parallel [80], i.e., it can be considered the model-free counterpart of
characteristics inherent in neural networks and hence their model predictive control.
widespread applications in power electronics and motor drives.
However, the entire field of electric machine drives has B. Scope
remained pretty much silent on the resurgence of ML over
the past decade, when compared with its continued success The scope of this paper is to provide a comprehensive
and widespread application in condition monitoring [39]–[45], overview of the pertinent literature that applies ML techniques
design optimization [46]–[65], and manufacturing [66], [67] of to electric machine drives from the 1980s to the state of the
various types of electric machines. It was not until in the last art. Despite the widespread application of classical artificial
few years that research efforts have begun to gradually catch intelligence (AI) techniques in the field of electric machine
up with the trend [68]–[78]. It is anticipated that with the drives, such as expert systems [81], fuzzy logic systems
rapid progress in ML models and embedded systems, the data- [6], [82]–[90], and evolutionary algorithms applied to tuning
driven approach will become increasingly popular for the high- classical controller/estimator parameters [83], [91]–[96], it is
performance control of electric machine drives, as envisioned in the authors’ humble opinion that AI is not used here
in Fig. 1. While most ML models require some complex offline by definition, since the usual procedures are rather simple
training processes, the online inference process can be made compared to the cutting-edge research within AI computer
relatively simple through various pruning and quantization science, and resulting algorithms do not stand up to the usual
methods [38], such that the groups of parameters with insignif- definitions of “intelligence” that mimic “cognitive” functions
icant impact for the artificial neural network’s (ANN) input- such as perception, attention, memory or language processing
output characteristics are removed and the numeric precision [97], as illustrated in Fig. 4. On the other hand, “machine
of the weights is reduced, resulting in reduced model size learning” is defined as the study of (computer) algorithms
and faster computation at the cost of minimal reductions in that can improve automatically through experience and by the
predictive accuracy [79]. use of data, which is considered a more appropriate summary
of the majority of literature included in this review paper.
Therefore, “machine learning” will be used for the remainder
A. Machine Learning Algorithms of this manuscript, even if some authors might have used the
There are three main classes of ML: supervised learning, exact wording of AI in the titles of their papers.
unsupervised learning, and reinforcement learning. A brief Furthermore, it is also shown in Fig. 4 that the general
description of each class and its classic applications are understanding of “deep” ML or deep learning is typically char-
illustrated in Fig. 2 and in the paragraphs below. acterized by using data-driven models of (very) large depths
1) Supervised Learning: Supervised learning is the ML (e.g., many ANN layers with millions of tunable parameters)
task of developing models to map input and output data. such that the manual feature engineering is automatically
The term “supervised” refers to the use of labeled data to handled within the ML pipeline (e.g., using pooling). In
train models for classification or regression problems. Some contrast, classical (shallow) ML comes along with manual
of the most widely used supervised learning algorithms include feature engineering in expert-driven or heuristic pre-processing
ordinary least squares, ANN, support vector machines, etc. As steps and ML models of very limited depth. Because “genuine”
will be discussed in later sections, these algorithms have been deep ML models are dramatically computationally expensive,

2
 Embedded
Control Hardware
Deep Learning (DL)
A class of ML which uses huge, layered models
Modulation Scheme (e.g. large artifical neural networks) to
Higher-Level Control Core Control progressively extract more information from the data.
DC
Machine Learning (ML)
AC A subset of AI involved with the creation of
Inverter algorithms which can modify itself without
Compens. human intervention to produce desired output
by feeding itself through structured data.

Flux Observer Artifical Intelligence (AI)

Any device that perceives its environment and
Motor takes actions that maximize its chance of
successfully achieving its goals. AI is often used to
describe machines that mimic "cognitive" functions
that humans associate with the human mind.

Fig. 4. The broader scope around ML [80].

Fig. 3. Simplified block diagram on ML applications in a generic motor drive
system. Every part of the motor drive control scheme could be ML-based,
while also the entire control framework could be just one large ML model.
The control depiction utilizing rotating dq-coordinates is only of illustrative Due to the widespread popularity of ML and the abundant
purpose as the ML-based method is not limited to this coordinate system. resources regarding its fundamentals, this review paper as-
sumes that readers have a sufficient understanding of the basic
concepts of ML, as outlined in [29], [81], so it can be more
they will remain non-real-time capable for years or even
pivoted to introducing their successful and diverse applications
decades with regard to embedded applications in motor drives.
in electric machine drives.
Therefore, the scope of deep learning will also not be covered
in this paper.
II. ML-BASED I NDUCTION M ACHINE D RIVES
A. Classical PI/PID Controllers Replaced by ML-Based Con-
C. Contribution trollers
The contribution of this paper is to comprehensively sum- While the conventional proportional-integral (PI) or pro-
marize the recent advances in applying ML-based methods to portional–integral–derivative (PID)-type controllers are widely
the control and monitoring of electric machine drives and to used in the industry due to their simple control structure,
identify suitable embedded systems for deploying such ML ease of design, and inexpensive computation cost [83], [98],
algorithms in real time. Essentially, every part of a generic they often cannot provide perfect control performance if the
motor drive system, such as the core current/speed controller, controlled plant is nonlinear and uncertain [99]. Moreover,
the higher-level controller generating the optimal torque, flux, when using a standard field-oriented control (FOC) frame-
or speed commands, the flux estimator, the inverter nonideality work a certain voltage margin is required to ensure proper
compensation, and the modulation scheme, can be substituted decoupling which reduces the achievable power density of a
by ML-based models as shown in Fig. 3. Additionally, the given drive [100]. Therefore, many ML-based controllers are
entire control framework could also be accomplished using designed and implemented as alternatives to the conventional
just one large ML model. These ML-based models should PI/PID controllers to identify and adaptively control induction
normally be executed at the same task frequency as the motor machines.
control software, i.e., in the micro- to millisecond range. The idea of using artificial neural networks (ANN) to con-
Classical field-oriented approaches or model-predictive control trol inverter drives was first proposed in [2], [3], where ANNs
are typically used in this context. However, most ML methods are trained offline to mimic the behavior of hysteresis current
do not rely on a specific control scheme. controllers to generate desired switching patterns. It is found
that such ANN controllers can deliver similar performance
to the original hysteresis controllers, plus certain advantages
D. Outline such as enhanced fault tolerance to the lack of one-phase
The paper is organized as follows: Sections II and III current error input [2]. It is also worthwhile to mention that
introduce specific applications of ML methods developed the scope of fault tolerance is only narrowly defined in [2], as
for induction machines and permanent magnet synchronous ML techniques generally cannot extrapolate to unseen events
machines. As the inverter and sensors are important parts of and, therefore, one would need to spend a lot of resources
any drive system, Section IV discusses state-of-the-art ML to train such an ML-based controller for many potential fault
techniques applied to those drive components. In Section V, scenarios to make it truly fault-tolerant. Additionally, these
the future trend of electric machine drives enabled by state- early works have not made attempts to design an ML-based
of-the-art reinforcement learning algorithms is introduced. controller with better dynamic performance.
Section VI presents an in-depth comparative study on the The first attempt using ANN to identify the induction
potential embedded platforms to host such ML applications machine dynamics and then control its stator currents and rotor
in electric machine drives for optimal cost and performance. speed in an adaptive manner is presented [7]. For both control

3
schemes, observable forms of the electromagnetic model of where vαs , vβs are the stator voltage components, iαs , iβs are
the induction machine are presented, and two systems are the stator current components, and ψαs , ψβs are the reference
introduced to identify the model and the change in rotor rotor flux linkage components all expressed in the stationary
speed using ANNs. Based directly on these two identification reference frame. Lm is the machine mutual inductance, Rs is
models, two ANN controllers are trained to adaptively control the stator resistance, Ls is the stator self-inductance, Lr is the
the stator currents and the motor speed. It is shown in rotor self-inductance, and σ is the leakage coefficient given by
simulation that the response of the ANN-controlled system σ = 1 − L2m / (Ls Lr ).
improves with time as the system learns, and during the last Additionally, the IM current model in the stationary refer-
transient, it actually outperforms a fine-tuned vector control ence frame can be written as
system. dψ̂αr Lm 1
Besides the specific ML-based controller scheme proposed = iαs − ψ̂αr − ωr ψ̂αr ,
dt Tr Tr
in [7], there are also many other variants of such controllers (2)
that offer decent dynamic performance. For example, a two- dψ̂βr Lm 1
= iβs − ψ̂βr + ωr ψ̂βr ,
degree-of-freedom (2DOF) controller is adopted in [8] to dt Tr Tr
regulate the rotor speed and the stator currents as an alternative where Tr is the rotor time constant, ωr is the measured or
to the conventional PI controller. The controller parameters are estimated rotor speed, ψ̂αr and ψ̂βr are the estimated rotor
adaptively tuned in real-time using neural networks, which can flux linkage components in the stationary reference frame.
offer much improved transient performance when compared It is also well-understood that the accuracy of the voltage
with fixed-gain 2DOF controllers. Furthermore, a robust speed model suffers at low frequencies due to the presence of ideal
controller based on the recurrent neural network is developed integration, which is susceptible to the measured input voltage
in [101]. Nevertheless, it pretty much follows the same control bias and uncertainties on the stator resistance. However, its
architecture by having a recurrent neural network identifier and performance at high speeds is much more reliable as the
a recurrent neural network controller. effective voltage drop across the stator resistance becomes
However, it is also reported in [102] that such a control negligible when compared with the back-EMF. The current
scheme involving two distinct neural networks in charge of the model, on the other hand, tends to have good accuracy at
system identification and the control might lead to inadequate lower speeds due to the advantage that such ideal integration
performance in the presence of rapid load changes. Therefore, is not required. However, its dependence on the rotor time
it is recommended that the two separate tasks of system constant Tr , which varies widely due to temperature-incurred
identification and control be combined into a single operation variations of Rr and magnetic saturation-incurred variations
enabled by a single ANN, though no comparison results are of Lr . Therefore, these two models are usually blended into a
provided to justify such claims. In [103], the same authors hybrid model to cover the whole frequency range [31]2 .
further propose using five feedforward ANNs trained in paral- In attempts to overcome these issues, various ML-based
lel, instead of one, to perform such a distinct neural-network- state estimation schemes in the form of flux or speed observers
based estimation and control scheme. A rigorous comparative are proposed for the rotor field-oriented vector control of
study of neural network controllers against PI controllers is induction machines.
presented in [104], where both PI controllers for the d- and
1) ML-Based Flux Observers for the Rotor-Flux-Oriented
q-axis current can be replaced with ANN controllers as shown
Indirect Vector Control: One of the earliest implementations of
in Fig. 3. The simulation results demonstrate that the ANN-
an ML-based flux estimator for the rotor field-oriented indirect
based controllers can provide better current tracking ability
vector control is presented in [9], where a three-layer ANN
than PI controllers with fewer oscillations and low harmonics,
with 20, 10, and 1 neurons is trained for different load torque
and they are also less vulnerable to detuning effects caused by
transient response cases using the stator current ids , iqs in the
the variation of rotor time constant during high temperatures
synchronous reference frame. The output of the neural network
or at deeply saturated conditions.
is either the estimated flux magnitude ψ̂ or a unit vector of
the slip angle sin θsl , which can further be used to calculate
B. ML-Based State Estimation for the Field-Oriented Control the unit vectors of the synchronous reference frame cos θe
of Induction Machines and sin θe with the measured rotor speed ωr . The test results
For the rotor field-oriented control, it is necessary to know have successfully demonstrated the high accuracy attainable by
the instantaneous magnitude |ψ̂r | and position θ̂ of the rotor the neural network flux estimator with the maximum absolute
flux. In the direct FOC scheme, both of them need to be error of 0.03 p.u. and with an RMS error of 0.1%, which
directly estimated based on the IM voltage model, the IM validates that data-driven neural network flux estimators may
current model, or the ML-based flux observer shown in Fig. 3. be a feasible alternative to other flux estimation methods based
Specifically, the IM voltage model in the stationary refer- on models derived by experts based on pre-knowledge.
ence frame can be written as At around the same time, [10] proposes a neural flux
observer scheme consisting of two neural networks, namely
diαs Lm dψαr the neural flux estimator and the neural stator estimator. While
vαs = Rs · iαs + σLs + · ,
dt Lr dt the neural flux estimator is trained in a similar fashion to
(1)
diβs Lm dψβr
vβs = Rs · iβs + σLs + · , 2 Frequently referred to as the Gopinath observer approach.
dt Lr dt

4
estimate the rotor flux magnitude, the proposed neural stator
estimator is able to continuously tune the rotor time constant
Tr = Lr /Rr for generating an accurate slip frequency com-
∗
mand ωsl in the indirect FOC of induction machines. Rather vs
than estimating the rotor flux magnitude using ML-based Reference Model ψr
methods, a neural network decoupling controller is designed (Voltage Model)
in [105] to generate the currents and slip commands (i∗ds , i∗qs ,
∗ is +
and ωsl ). Trained using the flux and torque commands (ψr∗ and
∗ -
Tem ), the outputs of this three-layer ANN are compared with
the outputs of the conventional decoupling controller, and the Adaptive Model
resulting errors are used to tune this neural network with either (Current Model) ψ̂r
back-propagation or the Levenberg-Marquardt algorithm. Sim-
ω̂r
ulation results also demonstrate the accuracy of the proposed Adaptation
neural network decoupling controller as an alternative to the Mechanism
conventional indirect FOC decoupling controller of induction
machines.

2) ML-Based Flux Observers for the Rotor-Flux-Oriented

Direct Vector Control: Contrary to the rotor-flux-oriented
Fig. 5. Illustration of the rotor flux-based MARS components replaced by
indirect direct vector control scheme where the unit vectors neural networks.
cos(θe ) and sin(θe ) are generated by estimating the slip
frequency in a feed-forward manner, the unit vectors in the
direct FOC scheme are directly estimated from the d and q- and f ) and output responses (is , ψs , and ωr ) are taken. These
axis components of the rotor flux linkage derived from the signals, which inherently include parameter variations and
voltage model in (1) or the current model in (2), and these saturation of the motor, are used to train an ANN to identify
models can also be completely or partially replaced by ML- the inverse dynamics of the motor until the sum-squared error
based methods, as presented in [11], [105]–[107]. of the a and b phase stator flux (ψas and ψbs ) is below the
An ML-based flux estimator of feedback signals needed desired level. Then the rotor flux can be calculated from the
for the direct vector control is first implemented in [11], estimated stator flux using
where a two-layer neural network with 20 neurons in the Llr
hidden layer is trained using the estimated stator flux (ψ̂αs ψ̂dr = (ψ̂ds − σLs ids ),
Lm
and ψ̂βs ) by integrating the back-EMF and the measured stator Lr
ψ̂qr = (ψ̂qs − σLs iqs ), (3)
currents (iαs and iβs ) transformed into the stationary reference L
frame, and the outputs are estimations of feedback signals qm
including the magnitude of the rotor flux |ψ̂r |, unit vectors |ψ̂r | = (ψ̂dr )2 + (ψ̂qr )2 ,
cos(θe ) and sin(θe ), and torque T̂em . Despite exhibiting cer-
tain advantages over the conventional flux estimator, such as where ψ̂dr and ψ̂qr are the estimated rotor flux components
faster execution speed, harmonic ripple immunity, and fault expressed in the rotor reference frame, and σ is the leakage
tolerance characteristics, the proposed neural flux estimator coefficient of the induction machine defined earlier. The unit
also brings an increased amount of fluctuation and noise in all vectors can thus be calculated as
of the estimated signals. This happens because the neural flux ψ̂dr
cos (θe ) = ,
observer proposed in [11] is designed as a pattern recognition |ψ̂r |
system without any adaptation mechanism. To overcome this (4)
ψ̂qr
issue, [106] expands the training set and exploits information sin (θe ) = .
on the variation or detuning of the motor parameters obtained |ψ̂r |
via simulation. Specifically, random noise within 10% of the However, it should be noted that direct measurement of
reference voltage is added to the stator voltage to enhance the the stator flux used to train the neural network requires the
variety of the training set in the neighborhood of the desired induction motor to be modified to install flux sensors, such as
operating conditions. Moreover, the motor parameters are also Hall-effect devices and search coils, which is not appropriate
varied within a suitably designed region in the parameter for general-purpose industrial motors. Additionally, by using
space. The implemented neural network flux observer has 4 the model-based motor equations in (3), it is assumed that
inputs, 3 output neurons, and a single hidden layer with 20 parameters Lr and Lm are weakly affected by saturation,
neurons. which might not be the case for many highly-utilized induc-
Besides developing an ANN-based rotor flux estimator for tion machines, e.g., designed for the automotive industry. To
the indirect FOC, [105] also presents a neural stator flux overcome this issue, [108] proposes a hybrid ML model with
estimator for the direct FOC to replace the conventional a structured ANN which allows estimating of both the stator
method that requires the integration of the back-EMF. With flux as well as the electromagnetic machine torque thanks to
the IM drive in operation, measurements of input signals (vs introducing a priori expert knowledge on the system dynamics.

5
Here, the stator flux is only estimated as an intermediate 1) Adaptive Current Model Replaced by ML-Based Flux
quantity while only a torque measurement (and not a stator Observers: Some of the earliest attempts in designing ML-
flux sensor) is required to perform the data-driven training. based rotor flux MRAS speed observers are presented in [5],
[13], [114], where a two-layer ANN is proposed to replace
the conventional adaptive current model described in (2). The
C. ML-Based Rotor Flux Model Reference Adaptive System
estimated rotor flux from the ANN is compared with its target
(MRAS) Speed Observer
value from the reference voltage model, and the total error
The conventional rotor flux-based model reference adaptive between the target and the estimated rotor flux is then back-
system (MRAS) estimator is introduced in [109], and the propagated to adjust the weights of the neural network, after
structure of which is shown in Fig. 5. This speed observer which the ANN’s output will coincide with the desired value.
mainly consists of two mathematical models – the reference Instead of using the classical adaptation mechanism for speed
model and the adaptive model, as well as an adaptation estimation as outlined in (6) and (7), the estimated speed is
mechanism to produce the estimated speed. This scheme is represented as one of the ANN weights updated online using
one of the most commonly used rotor speed estimators and a backpropagation algorithm.
many attempts have been made to improve its performance Further enhancements of this scheme are presented in [115]
according to the literature, and it is later proven from control and [116], where an adaptive linear neural network is em-
theories that both speed and rotor flux estimation are possible ployed to represent the adaptive current model. Additionally,
using only measurements of stator electrical quantities [110]. this ANN is tuned using the sampled stator currents and the
The reference model is typically represented by the IM rotor flux-linkage components coming from the model-based
voltage model in the stationary reference frame in (1), while reference voltage model, indicating that such an adaptive ANN
the adaptive model is typically represented by the IM current model is used in prediction mode rather than in simulation
model in the stationary reference frame in (2). The presence mode found in [5], [13], [114]. Both the recursive and the or-
of cross-coupling in the speed-dependent components in the dinary least square algorithms are used to train the ANN online
adaptive model (2) can lead to an instability issue [111], to obtain the rotor speed information. When compared with the
[112], therefore, it is common to use the rotor flux equations nonlinear back-propagation algorithm used in [5], [13], [114],
represented in the rotor reference frame as the proposed linear neural MRAS observer achieves better
Lm behavior in zero-speed operation at no load, as well as lower
ψ̂dr = ids , complexity and computational burden. A similar approach is
1 + Tr · s
(5) also proposed in [117] for the linear induction motor drive.
Lm
ψ̂qr = iqs ,
1 + Tr · s 2) Reference Voltage Model Replaced by ML-Based Flux
Observers: Despite the success and improvement of ANN-
where ids and iqs are the stator current components, ψ̂dr and
based flux observers replacing the conventional adaptive cur-
ψ̂qr are the rotor flux components all expressed in the rotor
rent model in the MRAS sensorless control algorithm, there
reference frame.
are still problems with the IM drive’s performance down to
The design of the adaptation mechanism is based mainly on
zero speed. For example, it is reported in [114] that the speed
Popov’s hyperstability theory, and as a result of applying this
estimation performance is only acceptable when “the operating
theory, the signal of the speed tuning error εω can be written
frequency is bigger or equal to 2 Hz, or else fluctuations will
as [113]
exist in the speed estimation that “may lead to the halting of
εω = ψ̂αr ψβr − ψ̂βr ψαr . (6) the system.” It is further revealed in [116] that the maximum
A PI controller is typically used to minimize this error, instantaneous speed estimation error at zero speed is above
which in turn generates the estimated speed at its output [113] 10 rad s−1 with its adaptive current model replaced by an
  ANN, despite the fact that such error is as high as 20 1/s
Ki using the approach proposed in [114].
ω̂r = Kp + εω . (7)
s To improve the sensorless drive performance at low and zero
Despite being a simpler and less computationally intensive speeds, [118] proposes a new MRAS scheme that employs
method when compared with many other sensorless control an ANN flux observer to entirely replace the conventional
methods, the main problems associated with it lie in its reference voltage model, rather than the adaptive current model
low-speed performance due to machine parameter sensitivity, as described in the earlier methods. This method tends to work
stator voltage/current acquisition, inverter nonlinearity, and better at low and zero speeds as when compared with a voltage
pure integration for the stator flux. Since many model-based model-based flux observer, an ANN does not employ pure
estimation techniques rely on the back-EMF voltage, which integration and is less sensitive to motor parameter variations.
is very small and even vanishes at zero stator frequency, As illustrated in Fig. 5, a multilayer feedforward ANN that
these techniques will fail at or around zero speed [113]. To estimates the rotor flux from present and past samples of the
overcome these issues, various ML-based rotor flux MRAS terminal voltages and currents is used to replace the reference
speed observers are proposed in the literature [5], [13], [89], voltage model. The experimental results show a significantly
[90], [114]–[119]. improved low and zero-speed performance at no load versus
the conventional MRAS approach. It is further revealed for

6
a zero speed and 20 % load case, the speed estimation error nonlinearities such as ReLU or sigmoid, they are shown to
at steady state is as low as 7 1/ min, which is much lower perform reasonably well on a couple of low-frequency, real-
than the method replacing the adaptive current model with world datasets [151]. In fact, sin/cos transformations are com-
ML-based flux observers. monly used when learning in time-series cyclical data, such
as the machine saliency discussed in this subsection. It is also
3) Adaptation Mechanism Replaced by ML-Based Speed
worthwhile to mention that besides induction machines, this
Estimators: The performance deficiency of the conventional
ML-based saliency tracking technique can also be extended
MRAS approach at low speeds due to pure integration and
to other machines, including permanent magnet machines and
machine parameter variations can also be mitigated by re-
synchronous reluctance machines.
placing the fixed-gain PI controller used in the adaptation
2) ML-Based Online Parameter Estimation of Induction
mechanism with ML-based control schemes [89], [90], [119].
Machines: Many supervised ML models have also been used
For example, a two-layer ANN is employed in [119] to
to perform online parameter estimation to enable more reliable,
replace such PI controllers, and the error between rotor flux
robust, and high-performance IM drives [144]–[148]. The
estimations from the conventional reference voltage model
performance of IM drives, especially those controlled by using
and from the adaptive current model is back-propagated to
indirect FOC, is inherently sensitive to the accuracy of the
the ANN to perform online training. The experimental results
rotor time constant Tr used to estimate the slip frequency ωsl .
demonstrate satisfactory speed estimation with less than 1 %
It is reported that the rotor resistance Rr may vary up to 100%
relative error when the induction machine is operating down
in certain applications over the entire range of operation due
to 10 1/ min.
to rotor heating [144], thus leading to compromised dynamic
drive response if it is not estimated in a real-time manner.
D. ML-Based Parameter and Model Identification of Induction To address the aforementioned issues, [144] and [145] have
Machine Drives proposed an online rotor resistance estimator using a simple
1) ML-Based Saliency Tracking for the Sensorless Control two-layer ANN trained by minimizing the error between the
of Induction Machines: ML models can be used to learn the rotor flux linkages based on an IM analytical voltage model
nonlinear dependencies of the machine saliency with respect and the output of this ANN. Since the analytical voltage
to its load and flux levels [140], which is crucial for reducing model also requires the knowledge of stator resistance Rs that
errors in the estimated rotor angle in IM drives with signal may also vary up to 50% during operation, another online
injection-based sensorless control. Different neural network estimator for Rs has been added using either a fuzzy nonlinear
types and learning methods are implemented and their perfor- mapping [144] or another ANN [145]. The proposed ANN-
mances are compared in [141]. The results demonstrate that based rotor resistance estimator was deployed onto a dSPACE
for the specific self-commissioning problem on an induction DS1104 controller board and was executed in 1 kHz, and
machine with closed rotor slots, the multi-layer perception net- satisfactory speed estimation can be obtained by using the
work shows the best performance followed by the functional proposed rotor resistance estimator. In [148], an online rotor
link neural network, whereas the time-delayed neural network resistance identification method is developed based on an
is only applicable using an extensive amount of training data. Elman network, which is typically a three-layer network with
Similarly, a physical model-based neural network, also the addition of a set of context units connected to the middle
referred to as the structured neural network, is employed to (hidden) layer fixed with a weight of one. These fixed back-
compensate for such saturation-induced saliencies [142] and connections could save a copy of the previous values of the
to perform automatic self-commissioning [143]. Originally hidden unit, making itself capable of adapting to time-varying
proposed in [149], structured neural networks have their characteristics and reflecting the dynamic characteristics of
interconnections between neurons determined by the physical a system. The experimental results show that the proposed
model, and their neuron basis functions are selected based method is able to enhance the dynamic response of speed
on physical representations. Therefore, a structured neural regulation of an IM drive.
network uses sinusoidal and cosinusoidal functions as its Besides the rotor resistance Rr , it is also essential to perform
activation functions with physical meaning, versus a “classical online parameter estimation of the mutual inductance Lm at
random (unstructured) feedforward neural network” that uses any operating condition to achieve optimal drive performance.
generic activation functions (such as a sigmoid function). This parameter typically decreases with the saturation of the
This structured neural network is also claimed to have a magnetic path, reflecting an inverse relationship that could be
significantly reduced training time with a simpler structure highly nonlinear. Two ANNs taking the form of a feedforward
than traditional neural networks. The experimental results in MLP and a recurrent network similar to an Elman network
[142] demonstrate that the estimated rotor position error using are proposed in [146] to estimate the mutual inductance.
such a structured neural network is roughly in line with those Although both networks were trained using the same dataset,
reported in [140] and [150]. It is further reported in [143] that the simulation results revealed that the recurrent network
this technique has the advantages of reducing commissioning maintains a filtering action that is advantageous during the
time and automating the process versus traditional methods oscillation of input data, while the feedforward network shows
such as look-up tables. a smaller error between the value developed in the network and
Despite the fact that sinusoidal activation functions are not the value from the motor model. Therefore, the feedforward
commonly used and do not fall into the generally applicable ANN is selected for experimental validation and it is shown

7
 TABLE I
ML A PPLICATIONS IN I NDUCTION M ACHINE D RIVES .

Applications References
ML-Based Control
Mimicking hysteresis current controllers to generate desired switching patterns [2], [3]
Replacing classical PI/PID current and speed controllers [7], [8], [101]–[104]
Generating the optimal flux command [120]–[122]
Achieving robust controller response against load disturbances [123], [124]
Implementing inverse optimal control [125]
ML-Based State Estimation
Functioning as flux observers for the rotor-flux-oriented indirect vector control [9], [10], [105]
Functioning as flux observers for the rotor-flux-oriented direct vector control [11], [105]–[108]
Replacing the MRAS adaptive current model with ML-based speed observers [114]–[117]
Replacing the MRAS reference voltage model with ML-based speed observers [118]
Replacing the MRAS adaptation mechanism model with ML-based speed estimators [89], [90], [119]
Formulating a current error-based MRAS speed observer [126], [127]
Developing full-order and reduced-order speed observers [128]–[130]
Correcting the estimated rotor speed from sensorless nonlinear control [131], [132]
ML-Based Signal Processing
Constituting a cascaded low-pass filter to obtain more accurate stator flux vectors [133], [134]
Training a neural notch filter to estimate the rotor flux at low speeds [135]
Introducing delayless finite impulse response and infinite impulse response filters [136]
ML-Based PWM Synthesis
Synthesizing space vector PWM [134], [137]–[139]
ML-Based Parameter and Model Identification
Learning the nonlinear machine saliency with respect to its load and flux levels [140], [141]
Compensating for saturation-induced saliencies in signal injection-based sensorless control [142], [143]
Performing online identification and parameter estimation [144]–[148]

that the accuracy of the speed estimation has been significantly challenging real-world conditions Therefore, similar to those
improved, which further enhances the overall controller perfor- applied to IM drives, ML-based controllers are also proposed
mance. Compared with the conventional feedforward ANNs to in the literature to improve the dynamic response of PM
be trained online, the adaptive linear neural (ADALINE) net- machine drives [153]–[157].
work has a simpler structure with only a single-layer network,
and its weights are can be interpreted physically. In [147], the For example, ANNs are also implemented as speed con-
IM model is approximated by two first-order subsystems with trollers in PMSM drives based on motor dynamics and non-
appropriate assumptions at the low and the high frequency, linear load characteristics [153]–[155]. In [153] and [154],
which can be readily used to design two ANALINE networks an ANN speed controller is developed to generate the q-axis
to identify an IM’s electrical parameters at standstill. After current reference i∗q (n), and the input of which are selected
online training, the stator resistance Rs and the stator cyclic as the speed of the motor at the present and previous two
inductance Lr are identified via the low-frequency ADALINE sample intervals [ωr (n), ωr (n − 1), ωr (n − 2)] in addition to
network, while the rotor time-constant Rs and the leakage the previous sample of the q-axis current reference i∗q (n − 1).
factor σ are identified via the high-frequency network. Apart The ANN speed controller can be integrated into the vector
from the above papers introduced in detail, readers are also control scheme of the PMSM drive with the initial weights and
referred to a comprehensive review paper on performing online biases obtained through the offline training using simulation
identification and parameter estimation in IM drives for more data. Then these weights and biases are updated online when
details [152]. an error between the actual output and the target of the
ANN exceeds a preset value. The robustness of the proposed
ANN scheme is validated using experiments against parameter
E. Summary variations [153] and load disturbances [154] in real time.
A summary of the aforementioned literature on ML models In addition, an ANN-based speed controller consisting of a
applied to induction machine drives is presented in TABLE I. radial basis function network is proposed in [155] and the
network is trained online to adapt to system uncertainties.
III. M ACHINE L EARNING -BASED P ERMANENT M AGNET The error between the reference speed and the measured
S YNCHRONOUS M ACHINE D RIVES speed is fed into the proposed ANN-based controller and
its weights and biases are trained online. The experimental
A. ML-Based Controllers for PMSM Drives results with load disturbances demonstrate that the proposed
A satisfactory current or speed controller should enable a ANN speed controller is able to regulate the motor speed
PM machine drive to follow any reference signal taking into in a more stable manner and with fewer transients when
account the effects of load impact, temperature, saturation, compared with the conventional PI controller. Furthermore, a
and parameter variations. However, as presented in the earlier brain emotional learning-based intelligent controller is further
analysis in Section II-A, conventional PI and PID controllers proposed in [156] and [157] to control the motor speed with
can lack the structural ability to achieve these objectives under very fast response and robustness with respect to disturbances

8
and manufacturing imperfections. Sliding-Mode Observer Filtering & PLL
Other relevant literature on this topic includes the hard-
-
ware/software controller designs using fuzzy ANNs for brush- +
less DC motor drives [158]–[163], achieving robust controller
response [164], [165], as well as formulating sliding mode
[166] and adaptive control schemes [167]–[169] for PM motor
drives.
Fig. 6. Simplified block diagram on ML serving as different components of
the SMO-based position observer based on the extended EMF model of PM
B. ML-Based State Estimation for PMSM Drives machines.

1) ML-Based State Estimation for the Sensorless Control

of PMSM Drives: A number of classical state estimation where
methods have been developed to achieve the sensorless control  
−Rs /Ld ω̂e (Lq − Ld ) /Ld
of PMSM drives, such as state observers, Kalman filters, dis- A=
−ω̂e (Lq − Ld ) /Ld −Rs /Ld
turbance observers, MRAS observers, sliding-mode observers,
      (11)
high-frequency signal injection [170]–[172], etc. However,
1/Ld 0 î u
these techniques usually suffer from DC drift due to motor B= , îs = αs us = αs ,
0 1/Ld îβs uβs
parameter variations and the influence of inverter nonlinearities
[173]. To overcome these issues, a wide variety of ML-based and ω̂e is the estimated electrical speed. The sliding-mode con-
methods are implemented to improve the existing sensorless trol function z contains the useful rotor position information
control schemes. and is defined as z = g · F (îs − is ), where g is the gain of the
Similar to the MRAS method for induction machines in control function and F (îs − is ) can be a signum, saturation,
Section II-C, the MRAS for PM machines also needs an or sigmoid function [182]. With F (îs − is ) being selected as
adaptation mechanism to provide accurate speed and position a saturation function and the gain of the control function g
estimations. However, the conventional adaptation mechanism being greater than the maximum value of the extended EMF,
is mostly linear, making it challenging to account for the namely g > |e|max , the observer can be asymptotically stable
effects of torque constant and stator resistance variations and the state can converge to S = is − is = 0 in a finite
on the rotor speed and position estimations. Therefore, a time. Therefore, the relation between the estimated EMF and
two-layer ANN is implemented in [174] as the nonlinear the control function becomes [176]
adaptation mechanism as shown in Fig. 5, and the experimental
results demonstrate that the proposed method is able to track ês = z. (12)
these varying parameters at different speeds with consistent Conventionally, the position estimate can be calculated
performance. directly from EMF estimates through an arc-tangent function
ML methods have also been extensively applied to improve as
the popular sliding-mode observer (SMO) designed using θ̂e = − tan−1 (êαs /êβs ) . (13)
the extended EMF model of PM machines [175]–[180]. The
voltage equation of the PM machine in the stationary reference However, the presence of noise signals may adversely affect
frame can be expressed as [181] the accurate estimation of rotor position. This is especially
       the case when using the arc-tangent function during the zero-
vαs Rs + pLd ωe (Ld − Lq ) iαs e
= + αs , crossing of EMF signals. To improve the position estimation
vβs −ωe (Ld − Lq ) Rs + pLq iβs eβs for industrial applications, a software phase-lock-loop (PLL) is
(8)
typically used to obtain rotor position from the estimated EMF
where vαs , vβs are the stator voltage components, iαs , iβs are
information. Moreover, normalization of the EMF is often
the stator current components, and eαs , eβs are the extended
necessary for the position observer due to the magnitude of the
EMF components all expressed in the stationary reference
EMF varying at different velocities. In this way, the equivalent
frame. Ld and Ld are the inductance of the d- and q-axis,
position error signal of the EMF model can be obtained as
Rs is the stator resistance, ωe is the rotor electrical speed, and
1 h    i
p is the differential operator, respectively.
εf = q −eαs cos θ̂e − eβs sin θ̂e . (14)
The extended EMF is defined as e2αs + e2βs
   
eα − sin(θe )
= [(Ld − Lq ) (ωe id − piq ) + ωe ψf ] , (9)
eβ cos(θe ) Therefore, the position observer based on the software PLL
can be expressed as
where θe is the rotor position and ψf is the permanent magnet
flux linkage. θ̂e = (1/s) (ki /s + kp ) εf , (15)
An SMO can then be designed based on the extended EMF
model of the PM machine in (8) to extract the rotor spatial where kp and ki are the proportional and integral gains of the
information contained in (9) as PI controller commonly used for the software PLL.
Fig. 6 illustrates implementations of ML models replacing
î˙ s = Aîs + B (us − z) , (10) different subsystems of the sensorless position observer based

9
on the extended EMF model of PM machines. In the pertinent irreversible demagnetization of permanent magnets as well as
literature, [175] proposed a five-layer wavelet fuzzy neural severe deterioration of other motor components, thus it is of
network (WFNN) to replace the conventional PI controller high concern for the machine’s control strategy and will result
in the PLL. To obtain good control performance in transient in oversized motor and inverter designs leading to lesser device
states and deal with the uncertainties of PM machines, both the utilization and higher material cost [184].
rotor angle estimation error and its derivative are selected as The traditional approaches for temperature estimation are
inputs to the network. In order to train the WFNN effectively typically accomplished by using sensor-based methods or by
and guarantee the convergence of the rotor angle estimation estimating electrical parameters such as the stator winding
error, the varied learning rates are derived based on the resistance, stator inductance, permanent magnet flux linkage,
analysis of a discrete-type Lyapunov function. When compared etc., using state-space observers or high-frequency injection.
with the PID-based PLL, the proposed WFNN-based PLL However, precise temperature estimation, especially for pre-
is able to reduce the average rotor position estimation error dicting the latent and highly dynamic magnet temperature, still
from 4.06° to 2.22° based on the experimental results. Due remains a challenging task, while the conventional methods
to the influence of inverter non-idealities and flux spatial still prove unfeasible in a commercial context [184]. In this
harmonics, the (6k ± 1) order harmonics are typically present regard, data-driven approaches could become very competitive
in EMF estimates, resulting in the (6k ± 1) harmonic ripples once proven that they could deliver highly accurate tempera-
in the estimated rotor position and speed and compromise ture estimations at low to moderate model sizes that would
their accuracy. To mitigate this issue, [176] designed a multi- allow them to run in real time in embedded systems.
input, single-output, and single-layer ADALINE network to
track and compensate for these (6k ± 1) order harmonics In the spirit of pursuing this goal, a comprehensive bench-
in the estimated EMF signals before they are fed into the mark study has been conducted in [184] that empirically
PLL. By continuously updating the filter weights online, the evaluates the temperature estimation accuracy of permanent
experimental results demonstrated that this ADALINE-based magnets using many classical ML models, including ordi-
filter is able to effectively suppress the (6k ± 1) harmonic nary least squares (OLS), support vector regression, k-nearest
ripples in the estimated rotor position and reduce its maximum neighbors, randomized trees, and multilayer perception (MLP)
position error from 8.3° to 2.2°. Finally, [183] proposes an feedforward neural networks. All of these ML models have
entirely data-driven sensorless PMSM torque control scheme been trained using the same test bench data collected from
which does not require any offline training. The proposed a three-phase, 52 kW automotive traction PMSM. This work
sensorless control algorithm can commission itself in a fully also reveals the full potential of simpler ML models, especially
automated fashion, i.e., pre-knowledge regarding a specific linear regression and simple feed-forward neural networks
drive system is not required. with optimized hyperparameters, in terms of their regres-
To design an EMF-based observer independent of any sion accuracy, model size, and data demand in comparison
machine parameters, [177]–[179] developed an ANN observer to parameter-heavy deep neural networks, which are imple-
that is trained to map between the dataset of the inputs mented in [185] in the form of recurrent neural networks
(iα , iβ , vα , vβ ) and those of the outputs (sin(θe ), cos(θe )), and temporal convolutional networks (TCN). For example,
followed by a PI-based PLL that tracks the rotor speed infor- the mean squared errors of OLS and MLP are 3.10 K2 and
mation based on the processed position error, and subsequently 3.20 K2 , respectively, while the TCN can further reduce this
the rotor position by performing integration. The ANN-based error by more than 50% to 1.52 K2 . However, this accuracy
observer has been tested on a 32-bit micro-controller and the comes at a cost of using as many as 67 k model parameters
inference time is around 5 µs. It is also revealed in [179] and its inference duration is taking 115 times longer than that
that the performance of this ANN-based position estimator of the OLS. The simpler ML models of OLS and MLP, on the
is very poor at zero and low-speed regions, and there are no other hand, only have 109 and 1.8 k model parameters with
comparative results presented using the conventional PI-based inference duration of 1.0 and 14.8, which are normalized with
observer. The conventional SMO, however, is also known to respect to the OLS model.
have compromised performance at the standstill and low-speed The potential of ML models can be further expanded by
conditions due to the amplitude of the back-EMF being almost adding more interpretability at their design stage, thus allowing
zero, [180] thus integrates an ANN-based angle compensation humans to capture relevant knowledge from such models
scheme into an iterative SMO that successfully mitigates this concerning relationships either contained in data or learned
issue. Specifically, the experimental results demonstrate that at by the model. [186] thus introduces a novel thermal neural
a low speed of 100 revolutions per minute, the maximum rotor network (TNN) that unifies both the consolidated knowledge in
position estimation error can be reduced to 4° from around 70° the form of heat-transfer-based lumped-parameter models, and
using an iterative SMO without the proposed neural network, data-driven nonlinear function approximation with supervised
and this large error might also indicate an issue in the way machine learning. Experiments on the same electric motor data
that the iterative SMO is designed or implemented. set show that this TNN is able to achieve accurate temperature
2) ML-Based Temperature Estimation of PMSM Drives: estimation with a mean squared error of 3.18 K2 at only 64
Besides estimating the states of motors that are directly related model parameters. A detailed review of temperature estimation
to drive controllers, temperature estimation is also a focal point methods for PMSMs, including the application of different ML
for PMSM drives. This is because overheating can trigger models, can be found in [187].

10
C. ML-Based Parameter and Model Identification of PMSM namely the stator resistance Rs , the stator inductance of the
Drives d− and q− axis Ld , Lq , and also the permanent magnet flux
linkage ψf . All of these online estimations have been proven
Similar to IM drives, accurate online parameter estimation to improve the dynamic and steady-state characteristics of the
and model identification are also crucial to achieving robust drive system.
and high-performance PMSM drives across their entire range In addition to the discussions above, readers are also re-
of operations. One of the earliest attempts at using ML-based ferred to the review paper on ML-based online identification
methods for this purpose is reported in [188], which proposes and parameter estimation of PM machines in [197] for more
a standard, 3-layer feedforward ANN to estimate the torque details on this topic.
constant and stator resistance of a PM motor, the values of
which will be used for torque ripple minimization of a dead-
beat predictive current controller. The simulation results show D. Summary
that the drive system is insensitive to these parameter changes
after implementing this ANN-based parameter estimator and A summary of the aforementioned literature on ML models
the torque ripple is reduced from 5% to 3%. applied to induction machine drives is presented in TABLE
II. The next-generation reinforcement learning-based motor
The simple structure and low computational demand of
control schemes, with most of the existing literature carried
the ADALINE network have also been leveraged in PMSM
out on PM machines, can be found in Section V.
drives for online model identification. For example, [189]
implements a direct instantaneous torque and flux controller
that requires accurate knowledge of the instantaneous elec- IV. M ACHINE L EARNING T ECHNIQUES A PPLIED TO
tromagnetic torque, stator flux vector, and machine electri- D RIVE I NVERTERS AND S ENSORS
cal parameters in order to accomplish a high-performance
instantaneous torque control scheme. All of these quantities As the inverter and sensors are also important parts of any
are estimated online using an ADALINE-based PM motor modern drive system, this section discusses state-of-the-art ML
model that is trained through back-propagation by minimizing techniques applied to those drive components, particularly in
the mean squared error between the measured q-axis current terms of modeling and compensating the inverter non-ideal
and its estimated value from the ADALINE network. When characteristics [191], [198]–[200] and condition monitoring of
compared with the conventional torque control with decoupled sensors used to provide critical feedback signals in a motor
PI current controllers, the experimental results reveal that the drive system [201]–[206].
torque ripple has been reduced from 8.5% to only 0.5% at
10 1/ min when using the ADALINE-based motor model, and
this ML-enabled drive system is able to offer fast and smooth A. ML-Based Modeling and Compensation of the Drive In-
torque response with enough robustness against disturbances verter Non-Ideal Characteristics
and parameter variations. Similarly, an online parameter es- In most motor drive applications, the stator phase voltages
timator based on a variable step-size ADALINE network is have modulated signal forms due to inverter switching and are
proposed in [190] to identify the PMSM parameters such as therefore difficult to measure. While it is technically feasible
the stator synchronous inductance Ls , the stator resistance Rs , to measure them directly such as using delta-sigma modula-
and the permanent magnet flux linkage ψf . The identification tors, additional cost and integration effort have prevented the
results of motor parameters are then substituted into the widespread implementation of phase voltage measurements in
prediction model of a deadbeat predictive current controller, mass-produced drives [187]. Therefore, an accurate inverter
which eliminates the current static error caused by parameter model is required to estimate the phase voltages ūabc from
mismatch and effectively improves the parameter robustness the reference voltage information ū∗abc in the motor control
of the controller. algorithm, as shown in Fig. 7(a).
Furthermore, [191] develops a data-driven recursive least However, due to various non-ideal characteristics of the
squares estimation method for online motor parameter identifi- drive inverter such as the interlocking time, non-ideal switch-
cation to improve the prediction accuracy of the finite-control- ing behaviors due to parasitics, signal delays, and the forward
set model-predictive-current (FCS-MPC) control of PM drives. voltage drop across semiconductors and cables, an analytical
The PMSM model parameters can be recursively corrected white-box modeling approach requiring simulation step times
with each new measurement and, therefore, the resulting FCS- in the nanosecond range is hardly feasible in a control context
MPC algorithm enabled by this data-driven method is able [198]. Therefore, a black-box inverter model utilizing ML
to outperform a baseline white-box model derived from first- and data-driven approaches is considered favorable for this
order physical principles [192]. To overcome the global forget- task. To train such an inverter model or compensation scheme
ting of (ultra-)local models [193], [194] extends the adaptive incorporating the inverter’s non-ideal characteristics, a large
local model approach with a long-term memory to allow in- amount of data samples needs to be collected a priori that
stant model reconfiguration to already visited operating points. cover the complete operating envelope of a motor drive system.
In [195], [196], a novel adaptive decoupling controller is also An exemplary dataset of 234,500 samples has been collected
introduced based on radial basis function neural network to in [198] on a two-level IGBT inverter and published on Kaggle
estimate the uncertain and time-varying motor parameters, [207].

11
 TABLE II
ML A PPLICATIONS IN P ERMANENT M AGNET M ACHINE D RIVES .
CTIONS ON INDUSTRIAL ELECTRONICS, VOL. 68, NO. 9, SEPTEMBER 2021
Applications References
ML-Based Control
Replacing classical PI/PID current and speed controllers [153]–[157]

Gray-Box and Black-Box Implementing the hardware/software designs for brushless DC motor drives
Achieving robust controller response
Implementing sliding mode control on a PM linear servo motor drive system
[158]–[163]
[164], [165]
[166]

Three-Phase Inverter Designing adaptive controller schemes

ML-Based State Estimation
Replacing the MRAS adaptation mechanism model with ML-based speed estimators
[167]–[169]

[174]

or Electrical Drives Improving different subsystems of the popular back-EMF-based observer with PLL
ML-Based Parameter and Model Identification
Performing online or offline identification and parameter estimation
[175]–[180]

[188]–[194]
Executing adaptive decoupling control considering uncertain and time-varying parameters [195], [196]
Member, IEEE, and JoachimEstimating
Böcker, the Senior
temperatureMember,
of permanentIEEE
magnets & multiple stator components [184]–[186]

Based on this open dataset, a comprehensive data-driven

ompared
e control black-box inverter model is established using a neural network
pecially if [199] to approximate the following function
vers (e.g.,
a distur-
fnn (dabc , dabc,pre , iabc,0 , iabc,1 , udc,0 , udc,1 ) = ūabc , (16)
al control
proaches
ufficiently which provides the set mean voltage vector ūabc with respect
gh or low to the relevant set duty vector of the current and the previous
on in the PWM period (dabc , dabc,pre ), the measured phase current
x inverter
is article. vectors at the beginning of the current and the next PWM
WO-LEVEL
apping ofTHREE-PHASE INVERTER MODELS 8651 period (iabc,0 , iabc,1 ), and the measured DC-link voltage
he actual Fig. 1. Schematic integration of(a)an inverter model and inverter com- values at the beginning of the current and next PWM period
arison, a pensation scheme into field-oriented control. (udc,0 , udc,1 ).
dentified
and a
Similarly, a suitable function for the inverter compensation
ehensive scheme illustrated in Fig. 7(a) can be formulated by rearrang-
s of both are generally not measured for cost reasons, the observer’s ing the output variable and the input features of the inverter
he phase performance highly depends on the modeled set voltage of the model
ot-mean- inverter. Especially at low-speed levels when only small voltage
k voltage fnnc (ū∗abc , dabc,pre , iabc,−1 , iabc,−2
black-box amplitudes are applied to the motor, the utilization of a voltage (17)
nV.[25]. information that deviates only little from the true value can udc,−1 , udc,−2 ) = d∗abc
already result in dramatically high observation errors [1], [2].
g, neural
Thus, for increasing the control performance, accurate inverter with the corresponding signals defined in Fig. 7(b).
ning, sys-
models and compensation schemes, which aim to cancel out all For the neural network representing the ML-based inverter
inverter non-ideal effects, are needed. Their integration into the model, a basic feedforward network layout is chosen, while
standard field-oriented control (FOC) framework is visualized hyperparameters such as the number of hidden layers, the
in Fig. 1. number of neurons, and the type of activation functions are
accuracy Normally, the basic procedure is to determine a suitable determined by Bayesian optimization. A detailed compara-
one hand, inverter model and to derive from this the compensation scheme tive study with a gray-box inverter model combining first-
therefore, [3]. Up to now, various inverter models have been presented in order principles from physics with data-driven-based parame-
fficiency) the literature: basic models that consider only the interlocking ter identification reveals that the ML-based black-box model
e voltage time [4]–[6], simple but physically motivated approaches that can precisely estimate the phase voltages per switching cycle
onn within
a drive are based on mean current values during the switching peri- with a root-mean-square error of less than 0.65 V at a 560 V
ters. The
integrated ods [7], [8], and more sophisticated ones that intend to model of DC-link voltage level, outperforming the gray-box model
mined by
voltages the inverter precisely with component deviations between the that only achieves an error of less than 1.1 V [199]. The
The BO three
Fig. 6. phases [9], with
Visualization parasitic
of the utilized effectsfor
(b)signals [10],
the [11],
trainingwith current
process of scope of the aforementioned gray-box inverter model has been
software the
Fig.BB
7. approaches
transients duringand
(a) Schematic their time
a switching
integration of andependence:
intervalmodel
inverter NN
[8], and
[9], NNcompensa-
[12],
M inverter
and repre-
Cor with further extended in [200] to also estimate the power losses
2020 and sent the NNsintofor electric
the inverter model (19)and
and
tion [28] zero-current
tion scheme clamping effects
machine (i.e.,
drives, (b)the
a phase inverter compensation
current reaches
visualization zero in the motor and inverter, and the parameter of which is
of the utilized
August 25, scheme (22),
signals for therespectively.
training process of the black-box inverter model with time
supported
nput and within the interlocking time [13], [14]). The
dependence: NNM and NNC represent the neural networks for the inverter
disadvantages also obtained via a data-driven approach based on particle
89029890. of those swarm optimization. It is envisioned that based on this precise
map the model (16) existing white-box
and the inverter models
compensation are (17),
scheme thatrespectively
their parameter-
[199].
mal
ical value
Drives ization is difficult due to insufficient information from data loss modeling, an optimal motor operation strategy can be
any (e-mail:
C. Inverter
sheets Compensation
and that parasitic effectsScheme can via
be onlyNNs covered with a developed in response to system changes in real time during
1] within
upb.de). high degreeNNs
Basically, of uncertainty. Thus, for
describe nonlinear identifying
models they operation.
the model
and, hence,
maximum
e available
parameters,
cannot often However,
be inverted. manual experiments
by rearranging [7],the[8]output
or automatized
variable
ng BO, a procedures
and the inputatfeatures
standstillin [3],
(16),[9],
the[15]
aimneed
of thetocompensation
be performed.ap- But12
follows, proach can analogously be formulated as identifying a function
permitted, but republication/redistribution requires IEEE permission.
 TABLE III
C ONDITION M ONITORING OF M OTOR D RIVE S ENSORS U SING ML M ODELS .

ML model Sensor type Fault type Inference time References

Extreme learning machine Current, speed, and DC-voltage sensors Stuck, erratic, and offset faults 10 ms [201]
Random vector functional link network Current sensor∗ Stuck, erratic, and offset faults 22 ms [202]
Random vector functional link network Speed sensor Stuck, erratic, and gain faults 228 ms [203]
Decision tree, support vector machines Current sensor Stuck and offset faults N/A [204]
ANN Current sensor Stuck, erratic, and gain faults 0.8 ms – 2.03 s [205]
ANN Current sensor Stuck and gain faults N/A [206]
∗ Also detects IGBT faults.

B. ML-Based Condition Monitoring of Motor Drive Sensors model-based methods are fast and independent of operating
conditions, their performance is also highly dependent on
Sensors are indispensable parts of any modern electric ma-
the accuracy of the model and its parameters. For signal-
chine drive to provide accurate real-time feedback signals for
based methods, the diagnostic process is based on the real-
enabling high-performance closed-loop controls. Specifically,
time evaluation of fault signatures obtained by manual feature
current sensors and rotor speed/position sensors are typically
extraction. For example, the average normalized current value,
required in most drive applications, while DC-link voltage
the magnitude of the harmonic frequency components, and
sensors are needed for implementing advanced features such
the asymmetry between the phase currents can all be used to
as the sensorless flux observer or online parameter estimation.
identify current sensor faults. Despite their independence from
However, the sensors in the drive system are prone to various
system parameters and models, these signal-based methods are
faults due to aging, vibration, humidity, and surrounding
highly sensitive to the load conditions of the drive system
interference [201]. The typical fault modes of the sensor can
and often require some expert domain knowledge to manually
be broadly classified as stuck faults, erratic faults, gain/offset
select the useful features.
faults, drift faults, and spike faults [208]:
Based on recent advancements, ML-based methods also
1) Stuck faults: The sensor’s output gets stuck at a fixed have the potential to become a promising alternative to the
value, and this can also be viewed as a complete failure. condition monitoring of sensors for motor drive applications.
2) Erratic faults: Variance of the sensor output significantly A summary of currently available publications on this topic
increases above the normal value. using ML techniques is presented in TABLE III. While only a
3) Offset faults: The output of the sensor possesses a con- few publications exist and most of them are also fairly recent
stant offset value some time after calibration. compared to model-based and signal-based approaches, it is
4) Drift/gain faults: The output of the sensor keeps increas- reported that ML-based solutions demonstrate some superior
ing or decreasing linearly from the normal state. condition monitoring performance with higher generalization
5) Spike faults: Spikes are observed in the sensor output at capability and increased robustness [201]–[203], [206]. How-
fixed intervals. ever, one potential concern still remains related to the model
Possible consequences of the above fault modes on the size/parameters and the corresponding inference time. This
different sensors in the drive system can be briefly discussed range can vary from 10 ms to a few seconds and some of
as follows [201]. A current sensor fault can result in an which may not be fast enough to meet the requirement of
imbalanced current flowing into the motor, causing overheat- certain drive applications.
ing and fluctuation/instability in speed and torque control. A
speed sensor fault affects the desired orthogonal alignment of V. R EINFORCEMENT L EARNING -E NABLED N EXT
the stator field and the torque component (q-axis) of current G ENERATION E LECTRIC M ACHINE D RIVES
in a drive, thereby leading to wobbles and fluctuations in Despite the widespread applications of RL in AlphaGo,
motor speed and phase currents. The voltage sensor fault can robots, and self-driving cars, RL has only recently been
negatively affect the performance of the flux observer and introduced to the control of electric machine drives [68]–[78].
the estimation of motor parameters. In summary, all of these Similar to the vision of self-driving cars where a car can drive
consequences resulting from the erroneous feedback due to itself and take its passengers to their desired destinations,
sensor failures could lead to degraded control performance or RL-enabled electric machine drives are expected to meet
even drive system shutdown. various performance requirements and efficiency specifications
Traditional sensor fault monitoring methods can be divided by automatically learning their optimal control policies via
into model-based and signal-based methods [209]. Model- direct interactions with the actual motors. This entire workflow
based methods aim to evaluate and monitor the difference can be completely automated and does only require minimum
between the measured output of the actual system and the human design effort as well as a priori model knowledge.
output generated by the model, which is typically obtained To make this RL-enabled motor control scheme a compet-
using state observers or MRAS-based approaches for contin- itive alternative to classical methods, many exemplary works
uous estimation of rotor speed and phase currents. While the have been presented over the course of the past few years

13
 RNING

ground plan- More recently, another important step is accomplished to-

control policy wards introducing RL to the control of physical motor drives,
s (a policy up- which involves the complete workflow of transferring an
decision-time RL controller from offline simulation to online training and
d allows more inference on a real motor drive system, as illustrated in Fig.
control policy 8(b) [74]. In order to outsource computational heavy RL com-
nference. putations, edge computing based on an internet of the things
framework is utilized. Consequently, only the control policy
inference must be calculated in real time on the embedded
that RL ap- controller while the actual RL training algorithms are calcu-
mance in sim- lated in an asynchronous fashion using dedicated computing
a basic proof resources. It is further envisioned that such an implementation
ontrol
ING context
will also be possible for low-cost applications in the future
n open-source
(a)
AI Gym stan- FIGURE 1. Simplified schematic of the overall control and drive system
using typical system-on-chip (SoC) embedded hardware with
structure; note that all gray shaded parts are control-related while from FPGA, as will be detailed in the next section of implementing
a simulation- the RL agent’s perspective, both the coordinate transforms and PWM are
-based control part of the environment, i.e., they are pre/post-processing steps outside ML-based motor drives in embedded systems. Furthermore,
ded to ek = the RL agent’s core software. [210] extends the scheme with an online safeguarding method
motor models.
nd o k+2
ent on a real-, an to prevent unsafe drive operations due to random exploration
vation
ack after
of limited actions. By doing so, online torque control learning can be
ric motor.
he-art control reward function, which penalizes policies exceeding the safety accomplished in less than 10 minutes on a real-world system.
vestigated the bounds. Besides the pioneering work mentioned above, readers are
-world motor Furthermore, electric motor control systems contain mul- also referred to other state-of-the-art literature employing RL
ayed in the tiple inherent forms of delays, e.g., calculation time of the
to PM machine [75]–[77] and switched reluctance machine
ation to real- controller hardware or the modulation scheme of the power
This action drives [78] for more details.
faced, as electronic converter [20]. These can be modeled as a one-step
asbeproposed
motor control, delay in the application of the agent’s actions, as described
to estimate
nts, measure- in Sec. IV-C. Such delays slow down the learning process
or example, VI. I MPLEMENTING M ACHINE L EARNING -BASED
n Although
the electri-an of RL agents significantly. To tackle a τ d -step delay before E LECTRIC M ACHINE D RIVES IN E MBEDDED S YSTEMS
utilized
ns to reduce in or- actions take effect, [21] appends the last τ d applied actions to
A. A Brief History of Embedded Systems for Electric Machine
physical
The actions sys- the observation of the RL agent. Alternatively, [22] uses re- Drives
the transfer current neural network agents and a special reward allocation
system dueis
urately
ard network match- to properly assign reward to past actions. Due to the lack of a suitable ANN application-specific
emember
discussed in In summary, the overwhelming majority of investigations integrated circuit (ASIC) or FPGA in the 1990s, the exper-
the
oblem in drive in the field of RL are based on simulations without any in- imental validation of the first ML-based control algorithms
ck into the teraction to real-world physical systems [12]. Addressing and
allows the for electric drives has been performed on microcomputers or
other game- FIGURE solving issues when transferring
6. Setup of the development
RL-based
(b)process control approaches
including the online RL microprocessors, focusing on available parallel architectures.
]. to8.real-world
remote rapid controlapplications, specifically for the field of electric
real-time re-Fig. (a) Simplified prototyping
schematic of toolchain.
the overall reinforcement-learning-based For example, the first ANN-based current controller to identify
al turnaroundcontrol driveandsystems,
drive systemis therefore an important
structure partitioning objectand
the agent ofenvironment,
research inand
NTS and control the induction machine dynamics presented in [7]
me constraint,(b) order setup of to the
be able to transfer
development data-driven
process includingcontrol
the online techniques
RL remote intorapid
uses an 25 MHz INMOS T800 transputer with a 32-bit integer
to comply
hardware be-control prototyping
industrial toolchain
processes [74].long run.
in the
ectric specific hardware architecture including the motor, controller processor that runs in parallel with a 64-bit floating-point unit
rence motor
and the and workstation is presented. Finally, important implementa-
B. CONTRIBUTION on a single chip [211]. Due to hardware limitations, the final
d on different tion details for the tests are described.
ssary. (21) exploring the boundary and tackling
In this work, the transfer from simplified offline simulation-the unsolved problems attainable sampling rate is 500 Hz with a two-layer ANN of
in motor con- [68]–[74].
based The
training first
to proof-of-concept
online training and of the
inference RL-based
on real current
motor 8 inputs, 12 hidden nodes, and two outputs. It is reported
A.drive
controlWORKFLOW
in a PMSM
systems is FROM drive
presented.SIMULATION
is Apresented
Python-based TO in THE
[68], TEST
rapid which BENCH
control success-
pro- that the stator currents will show signs of growing instability
roy the mo-
e limits of the
upply (e.g.,
RL algorithms
The
fully development
totyping
validates toolchain
the of
2 is
basic RL motor
developed
design controllers
that
architecture allows can be
online
shown split
intraining
Fig. into
8(a) with the increase of its electrical frequency until reaching a
nt complies
nstance, three
[17]and a steps
onunderlines
remote asplatform
shown in(edge
the potential Fig. computing)
6. First,
of theusing
the data-driven gym-electric-motor
measurements
controller de- point as low as 1.27 Hz, where the ANN controller behaves
s used
yer toolbox
[10]. sign.
correcting obtained [10]
fromcan
To accelerate be development
an embedded
the used withcontrollertheand standardized
(cf.training
Fig. 1).of interface
Therefore,
RL agents wildly. Therefore, it is suggested that the 500 Hz prototype
9]
term forces
rlim is from
the OpenAI
training Gym
process [11].
is Therewith,
executed many
asynchronously
thefor electric motor control, an open-source gym-electric-motor different
in the general-
back- ANN current controller must be increased by an order of
y shaping thePython purpose
ground. RL
This agents
toolbox toolchain from
is publishedallows several
to
inrapidlyPython
[69], test
[70], libraries
and validatecan
which vari-be
contains magnitude, and higher speeds of computation will be required
adapted
ous RL
models and tested ineasily
ofalgorithms
different DC forthree-phase
the context
and this use case.
of electric drive
motorAlso,
control different
with-for from the hardware.
variants
out the necessity
investigations to implement the training process within the
(22) easily accessible(e.g., on training
simulation. Thisparameters
package can andbenetwork
readily used ar- Additionally, two model reference adaptive speed neural
chitectures) can be executed in a
puter science, theto compare the trained RL agents with other state-of-the-art
simple and quick manner.
controllers proposed in [212], [213] are implemented in x86
We use both terms Afterwards, selected
2 The full rapid control RL algorithms
prototyping toolchainandwith parameter
an extendedspecifica-
technical
control approaches.
documentation
For
is available
the
as an
same purpose,
attachmentremote
a data
to this publication.
set consisting microcomputers with only a 500 Hz sampling rate, though they
eters to bal- tions are tested with the presented training setup on a
of about 40 million data points is recorded at a test bench are still shown to compare favorably against the benchmark
the reward real-time controlled SIL model utilizing an embedded rapid
for a 57-kW PM machine drive and is published on 2,Kaggle
VOLUME 2021 PI controllers during transients [213]. Furthermore, [214] and
control prototyping hardware system. The batched learning
[71], [72]. A deep Q-learning (DQN) direct torque controller is [215] provide an exemplar study that runs an ANN-based
s represent- under the real-time control and the proper transfer from a pure
further implemented for PM machines by aligning the limited current controller and the rest of the indirect FOC control on
given refer- simulation framework to an embedded hardware framework
number of distinct switching states
ntrol loops). is tested with this setup. Furthermore, the RL agent’s weights
of voltage source inverters a Texas Instruments TMS320C30 DSP. Despite implementing
and DQN’s finite control set framework.
d long-term are pre-trained in the SIL simulation. Finally, the chosen al- certain optimization strategies such as performing the hyper
an-squared- gorithm is trained and tested on the test bench. The training
ing control on the workstation as well as the controller can stay the same 14
error can be when exchanging the SIL model with the real motor. Solely
tangent sigmoid function by a look-up table, the final attain- the inference of an MLP with 3 hidden layers and 64 neurons
able sampling frequency is still only 1 kHz due to hardware in each layer can be completed in 7.36 µs with 32 parallel
limitations, though researchers always tend to use the best neuron control units. It is also reported from the proposed
available parallel hardware. implementation that it is more efficient for the inference of
During this time, a variety of algorithmic approaches are deeper MLPs (more hidden layers) compared to MLPs with a
also proposed in [216]–[219] to accelerate the continual on- high number of neurons per layer and fewer hidden layers.
line training and to enable a sampling frequency of at least An overview of recent achievements in the area of FPGA
10 kHz for modern electric machine drives. These acceler- and GPU-based implementations for reinforcement learning
ation methods include efficient parallelization methods such is provided in [227].
as output separation and tandem parallelization [216]; the Besides widely-used embedded systems based on DSPs and
random weight change algorithm to replace the conventional FPGAs, some rapid control prototyping platforms are also
backpropagation for online training [217], [219]; as well as leveraged to deploy ML-based motor control applications.
various techniques to reduce the computational demand [218]. For example, [228] proposes an ML-based induction machine
With the evolution of hardware capabilities in the new drive composed of a parallel combination of the classical PI
century, ML-based controllers for machine drives have been structure and the radial basis function neural network on a
advanced to execute at or above the desired switching fre- dSPACE DS1103 card. In addition, the inference of ANN and
quency. As presented in [220]–[222], all of the computations the rest of the vector control algorithm are also implemented
related to the same two-layer ANN proposed in [218] are on a dSPACE DS1103 in [104] with a sampling frequency up
now able to run at 10 kHz to identify the system dynamics to 10 kHz. The hardware experiments further reveal that when
within 1 ms using the pre-trained weights. The ML-based compared with the PI controllers, the ANN-based controllers
controller is deployed on a 333 MHz Analog Devices ADSP- can achieve much better current tracking performance with a
21369 DSP that is capable of executing at 2 giga floating-point low PWM switching frequency of 4 kHz, which further yields
instructions per second (GFLOPS). An interface card is also possibilities to improve the motor drive efficiency by lowering
used to host two FPGAs in charge of handling the high-speed its switching loss.
parallel data coming from the data acquisition system [222].
In [166], a field-oriented control PM linear machine drive is
B. Selecting Appropriate Embedded Systems for ML-Based
implemented on a 24 MHz Xilinx XC2V1000 FPGA with a
Electric Machine Drives
switching frequency of 15 kHz. In addition, ML models have
also been on FPGAs integrated with the National Instruments Although various ML-based electric machine drives have
CompactRIO controller for a two-mass electric drive system. been successfully implemented in embedded systems with
Specifically, a multilayer perception network is implemented DSPs [84], [91], [161], [214], [229], FPGAs [166], [223]–
for the speed estimation [223] and an ADALINE model is [225], [230]–[232], and embedded GPUs ( [233]–[237]) during
implemented as a speed controller [224]. the past 30 years, most of them have rather shallow network
Similarly, while the validation of ML-based flux observers structures and slow PWM cycles. Fortunately, the develop-
is only carried out in simulation in the 1990s [9]–[11], [105], ment of hardware platforms for parallel computing, including
[106], [138], the evolution of hardware platforms, especially GPUs, FPGAs, and TPUs, has significantly promoted the fast
FPGAs, has further advanced the implementation and valida- evolution and deployment of ML algorithms for industrial
tion of ML-based observers on the hardware. For example, a applications in recent years. A clear example is the currently
flux observer with two cascaded ANNs has been implemented very active domain of perception algorithms for advanced
in [107] using a single XC3S400 FPGA from Xilinx, and driver-assistance systems and autonomous driving. Based on
validation of the proposed FPGA controller is performed on the parallel properties inherent in such deep neural networks
a hardware-in-the-loop (HIL) test platform using a real-time applied to electric machine drives, an FPGA-based or GPU-
digital simulator with a 50 µs time step. Efficient inference of based implementation also seems promising and is highly
ML modes is also achieved in [225] when deploying a stator recommended in [38].
flux-oriented induction machine drive on a Stratix 2 FPGA by Due to their intrinsic architecture, however, GPUs are only
Altera. Specifically, the computational time for synthesizing efficient for processing data with large batch sizes that fit into
the SVPWM using a three-layer ANN and for estimating the the scope of CNNs. On the other hand, the control of electric
stator flux using a three-layer RNN are only 1.7 µs and 1.0 µs, machines will always utilize a handful of real-time measured
respectively. signals that have vastly different data representations than raw
However, it should be noted that most of the aforementioned pixel data processed primarily by CNNs. As such, GPUs may
ML models still have shallow structures with only one or two not be the most appropriate platforms for electric machine
hidden layers of a neural network with dozens of neurons at drives that require ultra-low latency and high interfacing
most, and the inference of significantly larger neural networks flexibility – though both of which are the strengths of FPGAs.
can be achieved using today’s mainstream FPGAs. A recent Therefore, we’ll focus on FPGAs for the remainder of this
study that seeks to investigate the boundary conditions to section with detailed discussions provided as follows.
use multilayer perception (MLP) networks in motor control For ML-based high-performance electric machine drives, an
applications has been performed on the Xilinx Zynq UltraScale ultra-low latency in the order of microseconds will be needed
ZU9EG FPGA with 2,520 DSP slices [226], which reveals that because the control frequency of which is typically in the range

15
 TABLE IV FPGAs, it seems there’s still big headroom for FPGAs to
P ERFORMANCE COMPARISON BETWEEN CPU, GPU, AND FPGA FOR THE run inference on deeper and larger neural networks for
INFERENCE OF NEURAL NETWORKS [238].
motor controls. For example, the current implementation
CPU GPU FPGA in [226] uses 32 DSP-slices to get to a point where the
latency is below 10 µs, while the low-end Xilinx Zynq-
Throughput Lowest Highest High
Latency Highest Medium Lowest 7020 offers 220 DSP-slices [240] and Xilinx UltraScale
Power Medium Highest Lowest ZU2EG offers 240 DSP-slices [241].
Energy Efficiency Worst Medium Best 2) Excellent interfacing flexibility: FPGAs can be repro-
Device Size Small Large Small
Development Easiest Easy Hard grammed for different functionalities and data types
Library Support Sufficient Sufficient Limited [242]. They also excel at handling data input from mul-
Flexibility Limited Limited Flexible tiple sensors, such as current sensors, voltage sensors,
thermocouples, encoders/resolvers, and accelerometers.
These features make FPGAs very flexible when optimiz-
of 10 kHz to 40 kHz, hence the maximum available calculation ing hardware acceleration of ML inference for electric
time for each control loop is tc = 25 µs to tc = 100 µs. Exclud- machine drives.
ing the time needed for ADC sampling, signal scaling/filtering, 3) High throughput: Based on the tightly-coupled SoC ar-
software-based protection logic, etc., the available time for the chitecture, FPGAs can deliver high throughput by op-
inference of deep neural networks has to be always lower than timizing hardware acceleration of ML inference in the
a full control cycle. Furthermore, machine drives will also programmable logic (PL) part and other non-critical
need to interface with many different categories of sensors to functions in the processing system (PS). Additionally,
properly perform the control, estimation, and monitoring of they also have the capability of hardware software co-
electric machines for different industrial applications. design to achieve optimized balancing between the two.
In addition to the low latency and high interfacing flexibility These desirable features could bring matched throughput
discussed earlier, there are also many other advantages of with end-to-end applications that are able to deliver
using an FPGA for the inference of deep neural networks, significantly better performance than fixed-architecture
as presented in TABLE IV. We’ll also elaborate on how AI accelerators such as GPUs. That’s because with a
these advantages are particularly relevant to motor control GPU, the other non-critical functions of the application
applications as follows: must still run in software without the performance or
1) Low latency: Latency is important in the inference of efficiency of custom hardware acceleration.
neural networks as it is directly tied to their real-time 4) Affordable cost: Large GPUs can be excessively costly
performance. FPGAs offer clear advantages over GPUs to be considered suitable for many electric drive appli-
and CPUs with lower latencies, which is a prerequisite cations, including home appliances, pumps, fans, or even
for applications that run inference in real-time, such as electric vehicles, while FPGAs are often more affordable.
the control of electric machines (including online RL By integrating additional capabilities onto the same chip
training). This advantage can be attributed to the fact that thanks to its SoC architecture, designers can also save
FPGAs can be configured to directly access peripheral on cost and board space. In addition, FPGAs have long
hardware components, such as sensors or input data product life cycles, measured in years or decades. This
sources. Directly combining this with implementations characteristic makes them ideal for use in industrial,
for the required preprocessing in the FPGA fabric pro- defense, medical, and automotive markets as it further
vides a very high bandwidth and much lower latency. On reduces maintenance costs. Despite the costs of FPGAs
the other hand, the communication between GPUs and being still expensive when compared with the standard
hardware components is less efficient, since a standard micro-controllers that host classical FOC/DTC motor
bus (USB or PCIe) is typically required to access the control algorithms, the reconfigurable SoC can offer an
hardware, and a host system (or an embedded CPU) integrated and much simpler design of the software pro-
needs to be employed [238]. Furthermore, based on their gram and the hardware FPGA image. More importantly,
architecture, requiring a high number of threads running there’s a great potential for using ML-based methods in
in parallel, GPUs can provide high bandwidth only at terms of quick exploration and domain adaptation on
the cost of high latency since they are only efficient for motor control over existing methods than run on these
large batch sizes. As a qualitative example, [226] shows ultra-low-cost micro-controllers.
that the latency of a reinforcement learning-based motor 5) Low power consumption: With FPGAs, designers can
control application can be reduced to as low as 7.36 µs fine-tune the hardware according to the application to
on FPGAs, which is sufficient for a control frequency help meet energy efficiency requirements. FPGAs can
of 100 kHz. Specifically, the deployed neural network also provide a variety of functions to improve the energy
has 9,224 variables and the inference is performed using efficiency of the chip. It’s possible to use a portion of an
32 DSP-slices, which are offered by the programmable FPGA for a function instead of the entire chip, allowing
logic part of the Xilinx FPGA to efficiently implement the FPGA to host multiple functions in parallel and the
multiplications and multiply-accumulate operations. Al- ability of dynamic reconfiguration [242].
though the number of DSP-slices is a limited resource on

16
 Ultra Low Power Low Power Low-cost SoC Combining High-Performance SoC Combining
Architecture
Microcontrollers Microcontrollers Reconfigurable Microprocessor Reconfigurable Microprocessor,
Type
w/ ML Accelerators w/ ML Accelerators SoC and GPU SoC FPGA and AI Engines

Typ. Theoretical
0.01 to 1 0.1 to 5 0.1 to 2 1 to 100 1 to 10 10 to 100
Peak TOPS

Typ. Power
<0.1 0.1 to 5 1 to 5 5 to 100 5 to 50 10 to 100
Range in Watt

Representative Maxim Int. NXP Xilinx Nvidia Xilinx Xilinx

Device Example MAX 78000 i.MX8 M Plus Zynq 7000 Jetson NX Zynq Ultrascale+ Versal

RT Cores Hi-Perf. Hi-Perf. Hi-Perf. Hi-Perf. Hi-Perf.

(Lockstep) Cores Cores Cores Cores Cores

AI
ML Acc. ML Acc. FPGA GPU FPGA GPU FPGA
Engines

Fig. 9. Overview of relevant embedded platform types in the market, illustrating a simplified block diagram of their topology, providing indicative ranges
for the typical theoretical peak TOPS performance and power consumption for each type and one representative device example, figure adapted from [239].

SoC
- of current reconfigurable SoCs and can even be modified
-

System
u*
-
ARMCore 0 y -- dynamically in milliseconds during operation to meet changing
Yref -
I
� - �
� workload requirements [246]. The latest Xilinx edge Versal
Outer ML VE1752 is now shipping out to customers [247] and it could
�
I

loop
Model become a favorable embedded platform for next-generation
Inference y
ADCs
X

1-=
�
�
motor drive applications.
1
Shared memory

I I
ARMCore 1
C. Implementing Machine Learning-Based Motor Control in
FPGA init.
FPGA FPGAs
Data logging
Fig. 10 depicts a simplified example of the implementation
Communication of an ML-based motor control algorithm on a dual-core
i
reconfigurable SoC. First, the measurements are read from
the ADCs and processed by digital filters implemented in
Fig. 10. FPGA-based SoC structure for the inference of ML models for motor the FPGA. Subsequently, the inference of neural networks is
control applications, figure adapted from
. [243] executed in the FPGA that also estimates the current state
x(k). The reference command (torque, speed, or position)
yref (k) is provided by an outer control loop that runs on the
Based on the aforementioned comparisons, it can be con- ARM Core 0. The interface between Core 0 and FPGA is
cluded that FPGAs, especially those based on the SoC archi- realized by the integrated advanced extensible interface (AXI).
tecture, are among the most promising digital technologies for The other depicted ARM Core 1 is generally not part of the
implementing ML-based smart controllers in electric drives. control loop, but it is responsible for many “housekeeping”
Specifically, the reconfigurable SoC consists of memory, mi- tasks, such as data logging, communication with other systems
croprocessors, analog interfaces, an on-chip network, and a and users, and the initialization of the FPGA, which includes
programmable logic block. Additionally, heterogeneous multi- all the libraries, all the tenants, the real-time operating sys-
processing SoC (MPSoC) architectures offer better perfor- tem (RTOS), drivers, and application programming interfaces
mance in terms of power and performance when compared (API), etc.
with monolithic cores [244]. Examples of such a new class of However, it is also worthwhile to mention that FPGAs can
reconfigurable SoCs are the Xilinx All-Programmable Zynq, be difficult to program as they require significant hardware
the Altera reconfigurable SoC, and the Actel/Microsemi M1 design expertise or long learning curves for optimal use,
[245]. Fig. 9 gives an overview of the different architectures and the task of converting sequential, high-level software
that are available on the market, providing an indication descriptions into fully optimized, parallel hardware architec-
of typical performance and power ranges. In 2018, Xilinx tures is tremendously complex [248]. This limitation is only
also launched a new programmable chip architecture called becoming more profound when deploying ML algorithms with
the adaptive compute acceleration platform (ACAP), a re- a deep structure and a large number of parameters. Fortunately,
programmable multi-core compute architecture with new dedi- instead of starting from scratch, there are many different tools
cated AI engines integrated into the device. With this heteroge- and customized environments to streamline this process. To
neous approach, the architecture goes beyond the capabilities provide some examples, we’ll present some potential ways to

17
 Applications be used in a manner analogous to software libraries.
Specifically, PYNQ provided a base hardware system
with an aspect of generality that includes almost all
modules in the PYNQ board for flexible reuse, such
Software as interfacing blocks for DMA, audio, video, I2C, and
components from logic tools. Neural network acceler-
ators can then be implemented through such overlays,
as presented in [251], which deployed a recurrent neural
Hardware Design Vivado
network language model for speech recognition.
However, it should be noted that new accelerators have
to be developed from scratch within the PYNQ framework,
and similar to FPGAs in general, it is mostly limited to
© MERL
6/7/2021 CONFIDENTIAL 9

Fig. 11. PYNQ – an open-source project from Xilinx that features an easy the inference of a neural network, and the online learning
software interface and framework for rapid prototyping and development
[249]. through back-propagation is usually difficult to implement on
such low-cost FPGAs in general due to limited resources.
Alternatively, some on-device learning approaches that do not
deploy a trained ML-based controller for electric drives in the rely on back-propagation for training have been proposed for
FPGA. FPGAs [252]–[254].
2) Matlab HDL Coder and Xilinx System Generator (XSG):
1) PYNQ – Python Productivity for Zynq: PYNQ is an
HDL Coder provides a workflow advisor that automates the
open-source project from Xilinx that makes it easier to use Xil-
programming of Xilinx, Microsemi, and Intel FPGAs [255].
inx platforms by using the Python language and libraries [249].
Specifically, it can generate portable, synthesizable Verilog
Compatible with Zynq, Zynq UltraScale+, Zynq RFSoC, and
and VHDL code from over 300 HDL-ready Simulink blocks,
Alveo accelerator boards, the PYNQ platform improves the
MATLAB functions, and Stateflow charts. With HDL Coder,
productivity of designers already working with Zynq, and it
programming FPGAs for ML-based motor control applications
reduces the barrier to entry for users with limited experience
can be achieved at a high level of abstraction, and the gener-
in hardware design. Fig. 11 illustrates the general concept of
ated HDL code can be imported and compiled into customized
the PYNQ framework consisting of three layers:
IP cores using the Intel Quartus or the Xilinx Vivado Design
• Upper Layer (Applications): The upper layer of the Suite.
PYNQ stack enables user interaction using one or more Besides the HDL Coder, Xilinx also developed its own
Jupyter Notebooks, which are hosted on Zynq’s Arm Xilinx System Generator (XSG) that adds Xilinx-specific
processors, also known as the processing system. Custom blocks to Simulink for system-level simulation and hardware
functionalities specific to each application can be created deployment. We can also integrate System Generator blocks
by writing Python code and using many open-source with native Simulink blocks for HDL code generation on the
Python libraries. In addition to developing software-based desired neural network structure. In [226], for example, the
functionality running on the PS, Python code within VHDL code for two multi-layer perceptions (MLP) neural
the notebook can also offload processing to hardware networks is also generated by the HDL Coder.
modules operating on the PL [250]. Interaction with By adopting such a model-based workflow utilizing the
hardware is achieved using the Python APIs and drivers HDL Coder, the proper functioning of the system can be first
that are provided as part of the PYNQ framework. The examined by simulation and co-simulation in Matlab, then
programmer’s experience of using hardware blocks is the block design is integrated into the FPGA architecture in
therefore very similar to calling functions from a software the form of an IP core. This workflow is very convenient for
library — a software developer can call a hardware high-level integration of various IP blocks created using the
block without any need to understand the internals of Matlab/Simulink graphical interface, especially for those who
the hardware design. are not familiar with hardware description languages such as
• Middle Layers (Software): In the middle layer, the PYNQ VHDL and Verilog. Also, the debugging and verification of
framework includes a Linux-based OS, bootloaders to HDL designs become easy and flexible with the Simulink
initiate system start-up, a web server to host Jupyter toolbox, though the performance and resource utilization of
notebooks, and a set of drivers for interacting with such toolboxes may not yield the optimal design compared to
elements of the Zynq hardware system. Thus, the design experienced FPGA designers.
effort of developing common software elements of an 3) Deep Learning Processor Unit (DPU): Besides the high-
embedded system is significantly reduced, and new users level synthesis (HLS) tool that can compile deep learning
are expected to get started quickly with Zynq. C/C++ code for programmable logic in the hardware [256],
• Lower Layer (Hardware): The bottom layer of the stack Xilinx also developed The Deep Learning Processor Unit
represents a hardware system design, which would nor- (DPU) intellectual property (IP) core that can be integrated
mally be created in Vivado requiring significant hardware into the programmable logic of selected Zynq-7000 SoC, Zynq
design expertise. In PYNQ, however, hardware system UltraScale+ MPSoC, and Versal AI edge devices with direct
designs are often referred to as overlays and they can connections to the processing system. Specifically, this DPU

18
 Chapter 2: Overview

Figure 3: Example System with Integrated DPU

D. Others
DisplayPort
USB3.0
Arm Cortex-
A53
Arm
Cortex
The actual hardware design of FPGAs can be performed by
-R5
SATA3.1
PCIe Gen2 combining any of the methods mentioned above. In addition.
GigE
USB2.0
some advanced high-level synthesis (HLS) tools, such as the
UART Controller DDR
SPI
Quad SPI
AXI Interconnect Auto-HLS [259], can be used to directly generate synthesiz-
NAND
SD able C code of the ML models and to conduct latency/resource
AXI Interconnect
estimation and FPGA accelerator generation.
AXI AXI
The framework of FINN [260] can also be adapted to build
DMA
Interconnect Interconnect

Color_
fast and flexible FPGA accelerators by reducing the weights
demosaic gamma
conversion
and activations of ML models for motor drive applications to
MIPI MIPI DPU
low bit width or even binary values. This method is especially
Camera
CSI2
CSI2 well-suited for CNNs that contain significant redundancy, and
a similar motor drive performance is expected against the
original ML model without adapting to the FINN framework.
Fig. 12. Example system with an integrated deep learning processor unit In addition to embedded control systems, commercial rapid
(DPU) [257].
Vitis AI Development Kit control prototyping (RCP) systems have also been used in
The Vitis™ AI development environment is used for AI inference on Xilinx® hardware platforms. deploying ML-based motor control algorithms. Such systems
It consists of optimized IP cores, tools, libraries, models, and example designs.
include the dSPACE MicroLabBox and DS1006MC [74],
As shown in the following figure, the Vitis AI development kit consists of AI Compiler, AI
is aQuantizer,
programmable engine
AI Optimizer, AI Profiler, dedicated
AI Library, to Library
and Xilinx Runtime convolutional
(XRT). neural which implement a deep deterministic policy gradient algo-
networks. This unit includes the register configuration module, rithm that learns the current control policy for a PM motor.
the data controller module, and the convolution computing Moreover, open-source software and hardware RCP systems,
module. The DPU has a specialized instruction set, which such as UltraZohm [261] or AMDC [262], can contribute
allows the DPU to work efficiently on many convolutional to distributing open ML-based drive control and monitoring
neural networks, including VGG, ResNet, GoogLeNet, YOLO, solution.
SSD, MobileNet, FPN, etc. The figure below shows an ex-
ample system block diagram with the Xilinx UltraScale+ VII. F UTURE C HALLENGES AND T RENDS
PG338 (v3.3) February
MPSoC using3, 2021
a camera input. The DPU is integrated www.xilinx.com
into10the
Send Feedback
Zynq DPU This paper provides a comprehensive state-of-the-art review
system through an AXI interconnect to perform deep learning
of ML-based solutions addressing the control and monitoring
inference tasks such as image classification, object detection,
of electric drives. Despite the continued progress of relevant
and semantic segmentation [257].
publications in this field, there are still some unresolved issues
This Xilinx DPU IP module is provided at no additional that need to be addressed as future work:
cost with the Xilinx Vivado Design Suite. However, it should • Development effort: Although FPGAs can offer better
be noted that as a CNN IP core, DPU is highly tailored for energy efficiency, connectivity, and flexibility, one major
computer vision and image recognition-related applications, challenge of using FPGAs is the engineering effort in
where users are expected to prepare the instructions and input development. Unlike GPU development which requires
image data in the specific memory address that DPU can only software engineering skills, the development of
access. Although CNNs are seldom used to tackle control FPGAs requires hardware configuration skills as well.
tasks of high complexities - such as electric machine drives, The complexity of implementing ML models on FP-
the convolutional layers can often be deployed as a part of GAs makes their manual design processes very time-
the reinforcement learning algorithms. For example, in order consuming, even for a seasoned FPGA engineer. In
to learn good policies with just pixel inputs, the authors addition, although many researchers have focused on
of the deep deterministic policy gradient (DDPG) algorithm ML inference on FPGAs, very few research papers have
used 3 convolutional layers to provide an easily separable explored their training on FPGAs, or how to optimize
representation of state space [258]. the architecture design on FPGAs for training. This is
However, one should also note that [258] learns from particularly needed for deploying RL algorithms to motor
raw pixels and, therefore, processes image data where CNNs control applications as the core of RL is to be able
shine. As discussed earlier, electric machine drives utilize to have the agent interact with the environment in real
a completely different data representation than what CNNs time and learn a policy (control law) in a trial-an-error
are primarily used to process. Since CNNs are really costly way. Therefore, an automated design workflow from the
in computation, the prospect of their applications tailored RL’s neural network architecture to the hardware design
for electric drives on low-cost embedded platforms such as is necessary to enable efficient and effective training of
FPGAs for low-dimensional control tasks is somewhat unclear. RL control on FPGAs (i.e., to not only utilize the FPGA
Nevertheless, if CNNs are selected to accomplish certain for policy inference but also for online policy learning).
motor control tasks, we can still benefit from this DPU IP If an effective automated design workflow is developed,
core by taking advantage of its built-in convolutional layers researchers and engineers can quickly develop various
and integrating them with other layers of the neural network ML models for motor control applications without the
designed in custom IP cores. need to possess deep knowledge about hardware design.

19
 • Application effort: ML is data-hungry and normally one [6] S. A. Mir, D. S. Zinger, and M. E. Elbuluk, “Fuzzy controller for
needs to train an ML model individually for each drive inverter fed induction machines,” IEEE Trans. Ind. Appl., vol. 30, no. 1,
pp. 78–84, Jan./Feb. 1994.
system at an expensive test bench, hence the speedy [7] M. T. Wishart and R. G. Harley, “Identification and control of induction
transfer of an ML method between different applications machines using artificial neural networks,” IEEE Trans. Ind. Appl.,
is an issue for the industrial mass production usage. This vol. 31, no. 3, pp. 612–619, May/June 1995.
[8] Y.-S. Kung, C.-M. Liaw, and M. Ouyang, “Adaptive speed control
issue can be addressed from both the software and the for induction motor drives using neural networks,” IEEE Trans. Ind.
hardware emulation perspectives. In terms of software, Electron., vol. 42, no. 1, pp. 25–32, 1995.
some ML algorithms are specifically designed to enable [9] A. K. Toh, E. P. Nowicki, and F. Ashrafzadeh, “A flux estimator for
field oriented control of an induction motor using an artificial neural
transfer learning with strong domain adaptation capa- network,” in Proc. of 1994 IEEE Industry Applications Society Annual
bilities. Additionally, the hardware platform of different Meeting, vol. 1, 1994, pp. 585–592.
electric drive systems can be emulated in the hardware- [10] J. Theocharis and V. Petridis, “Neural network observer for induction
motor control,” IEEE Control Syst. Mag., vol. 14, no. 2, pp. 26–37,
in-the-loop environment, making it a lot easier to collect April 1994.
a sufficient amount of emulated data to train ML models [11] M. G. Simoes and B. K. Bose, “Neural network based estimation of
for any industry application. feedback signals for a vector controlled induction motor drive,” IEEE
Trans. Ind. Appl., vol. 31, no. 3, pp. 620–629, May/June 1995.
• Safety: Since ML is always subject to some kind of
[12] F.-Z. Peng and T. Fukao, “Robust speed identification for speed-
stochastic learning, a method’s output should be also sensorless vector control of induction motors,” IEEE Trans. Ind. Appl.,
considered a stochastic one. Therefore, the probabil- vol. 30, no. 5, pp. 1234–1240, 1994.
[13] L. Ben-Brahim, “Motor speed identification via neural networks,” IEEE
ity of failing is intrinsic to an ML model, which can Ind. Appl. Mag., vol. 1, no. 1, pp. 28–32, 1995.
cause trouble if an ML technique produces outliers for [14] G. C. Sousa, B. K. Bose, and J. G. Cleland, “Fuzzy logic based on-
estimation or control action. As a result, there could line efficiency optimization control of an indirect vector-controlled
induction motor drive,” IEEE Trans. Ind. Electron., vol. 42, no. 2, pp.
be negative impacts on the behaviors of mechatronic 192–198, April 1995.
systems, thereby compromising their prospects in safety- [15] P. Mehrotra, J. E. Quaicoe, and R. Venkatesan, “Development of
critical applications. an artificial neural network based induction motor speed estimator,”
in PESC Record. 27th Annual IEEE Power Electronics Specialists
• Interpretability: ML models are very complex and diffi-
Conference, vol. 1, 1996, pp. 682–688.
cult to understand or explain, as it is reported in [226] [16] S. Weerasooriya and M. A. El-Sharkawi, “Identification and control
that a recent ML model proposed for electric drive of a DC motor using back-propagation neural networks,” IEEE Trans.
Energy Convers., vol. 6, no. 4, pp. 663–669, Dec. 1991.
applications can have close to ten thousand parameters, [17] M. A. Rahman and M. A. Hoque, “Online self-tuning ANN-based
not to mention those commercially deployed ML models speed control of a PM DC motor,” IEEE/ASME Trans. Mechatronics,
for natural language processing or image recognition vol. 2, no. 3, pp. 169–178, Sept. 1997.
[18] H. Tsai, A. Keyhani, J. Demcko, and D. Selin, “Development of
tasks that could have millions or billions of parameters. a neural network based saturation model for synchronous generator
Although interpretability does not ensure safety by itself, analysis,” IEEE Trans. Energy Convers., vol. 10, no. 4, pp. 617–624,
it is important for monitoring functional safety and for Dec. 1995.
[19] D. S. Reay, M. Mirkazemi-Moud, T. C. Green, and B. W. Williams,
understanding where the models are failing. Therefore, “Switched reluctance motor control via fuzzy adaptive systems,” IEEE
more in-depth investigations regarding the interpretability Control Systems Magazine, vol. 15, no. 3, pp. 8–15, 1995.
and explainability of ML models are necessary for their [20] P. V. Goode and M.-y. Chow, “Using a neural/fuzzy system to extract
heuristic knowledge of incipient faults in induction motors. part i-
commercial deployment in drive applications. methodology,” IEEE Trans. Ind. Electron., vol. 42, no. 2, pp. 131–138,
Upon resolving many of the practical issues mentioned April 1995.
above, it is anticipated that the ML-based data-driven control [21] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “AI techniques in
induction machines diagnosis including the speed ripple effect,” IEEE
and monitoring schemes will be able to deliver unparal- Trans. Ind. Appl., vol. 34, no. 1, pp. 98–108, Jan./Feb. 1998.
leled performance in terms of quick exploration and domain [22] ——, “Recent developments of induction motor drives fault diagnosis
adaptation. Therefore, it has great potential to become the using AI techniques,” IEEE Trans. Ind. Electron., vol. 47, no. 5, pp.
994–1004, Oct. 2000.
next-generation electric machine drive technology over the [23] R. M. Tallam, T. G. Habetler, and R. G. Harley, “Self-commissioning
existing model-driven methods currently implemented in low- training algorithms for neural networks with applications to electric
cost microcontrollers. machine fault diagnostics,” IEEE Trans. Power Electron., vol. 17, no. 6,
pp. 1089–1095, Nov. 2002.
R EFERENCES [24] M. A. Awadallah and M. M. Morcos, “Application of AI tools in fault
diagnosis of electrical machines and drives-an overview,” IEEE Trans.
[1] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- Energy Convers., vol. 18, no. 2, pp. 245–251, June 2003.
sentations by back-propagating errors,” nature, vol. 323, no. 6088, pp. [25] X. Huang, T. G. Habetler, and R. G. Harley, “Detection of rotor
533–536, 1986. eccentricity faults in a closed-loop drive-connected induction motor
[2] F. Harashima, Y. Demizu, S. Kondo, and H. Hashimoto, “Application using an artificial neural network,” IEEE Trans. Power Electron.,
of neutral networks to power converter control,” in Conference Record vol. 22, no. 4, pp. 1552–1559, July 2007.
of the IEEE Industry Applications Society Annual Meeting,, 1989, pp. [26] M. B. K. Bouzid, G. Champenois, N. M. Bellaaj, L. Signac, and
1086–1091. K. Jelassi, “An effective neural approach for the automatic location of
[3] M. R. Buhl and R. D. Lorenz, “Design and implementation of neural stator interturn faults in induction motor,” IEEE Trans. Ind. Electron.,
networks for digital current regulation of inverter drives,” in Confer- vol. 55, no. 12, pp. 4277–4289, Dec. 2008.
ence Record of the 1991 IEEE Industry Applications Society Annual [27] S. Mohagheghi, R. G. Harley, T. G. Habetler, and D. Divan, “Con-
Meeting, 1991, pp. 415–421. dition monitoring of power electronic circuits using artificial neural
[4] B.-R. Lin and R. G. Hoft, “Power electronics inverter control with networks,” IEEE Trans. Power Electron., vol. 24, no. 10, pp. 2363–
neural networks,” in Proc. Eighth Annual Applied Power Electronics 2367, Oct. 2009.
Conference and Exposition,, 1993, pp. 128–134. [28] M. Cirrincione, M. Pucci, and G. Vitale, Power converters and AC
[5] L. Ben-Brahim and R. Kurosawa, “Identification of induction motor electrical drives with linear neural networks. CRC Press, 2017.
speed using neural networks,” in Conference Record of the Power
Conversion Conference-Yokohama 1993, 1993, pp. 689–694.

20
[29] B. K. Bose, “Neural network applications in power electronics and [53] S. Barmada, N. Fontana, L. Sani, D. Thomopulos, and M. Tucci, “Deep
motor drives—an introduction and perspective,” IEEE Trans. Ind. learning and reduced models for fast optimization in electromagnetics,”
Electron., vol. 54, no. 1, pp. 14–33, Feb. 2007. IEEE Trans. Magn., vol. 56, no. 3, pp. 1–4, March 2020.
[30] B. K. Bose, “Global energy scenario and impact of power electronics [54] A. Khan, M. H. Mohammadi, V. Ghorbanian, and D. Lowther, “Ef-
in 21st century,” IEEE Trans. Ind. Electron., vol. 60, no. 7, pp. 2638– ficiency map prediction of motor drives using deep learning,” IEEE
2651, 2013. Trans. Magn., vol. 56, no. 3, pp. 1–4, March 2020.
[31] B. K. Bose, Power electronics and motor drives: Advances and trends. [55] T. Guillod and J. W. Kolar, “From brute force grid search to artificial
Academic press, 2020. intelligence: Which algorithms for magnetics optimization?: Workshop
[32] ——, “Artificial intelligence techniques in smart grid and renewable at virtual psma industry session on design of magnetics for different
energy systems—some example applications,” Proc. IEEE, vol. 105, circuit topologies,” in 35th IEEE Applied Power Electronics Conference
no. 11, pp. 2262–2273, Nov. 2017. (APEC 2020), 2020.
[33] ——, “Artificial intelligence techniques: How can it solve problems [56] H. Li, S. R. Lee, M. Luo, C. R. Sullivan, Y. Chen, and M. Chen,
in power electronics?: An advancing frontier,” IEEE Power Electron. “Magnet: A machine learning framework for magnetic core loss
Mag., vol. 7, no. 4, pp. 19–27, Dec. 2020. modeling,” in 2020 IEEE 21st Workshop on Control and Modeling
[34] S. Zhao, F. Blaabjerg, and H. Wang, “An overview of artificial for Power Electronics (COMPEL), 2020, pp. 1–8.
intelligence applications for power electronics,” IEEE Trans. Power [57] J. Hao, S. Suo, Y. Yang, Y. Wang, W. Wang, and X. Chen, “Optimiza-
Electron., April 2020. tion of torque ripples in an interior permanent magnet synchronous
[35] L. Gao, “The decade of deep learning,” [Online]. Available: motor based on the orthogonal experimental method and miga and rbf
https://bmk.sh/2019/12/31/The-Decade-of-Deep-Learning/, Accessed: neural networks,” IEEE Access, vol. 8, pp. 27 202–27 209, 2020.
Jul. 2022. [58] V. Parekh, D. Flore, and S. Schöps, “Deep learning-based prediction
[36] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification of key performance indicators for electrical machines,” IEEE Access,
with deep convolutional neural networks,” Advances in neural infor- vol. 9, pp. 21 786–21 797, 2021.
mation processing systems, vol. 25, 2012. [59] T. Sato and M. Fujita, “A data-driven automatic design method for
[37] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. electric machines based on reinforcement learning and evolutionary
Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” optimization,” IEEE Access, 2021.
Advances in neural information processing systems, vol. 30, 2017. [60] S. Barmada, N. Fontana, A. Formisano, D. Thomopulos, and M. Tucci,
[38] E. Monmasson, M. Hilairet, G. Spagnuolo, and M. Cirstea, “System- “A deep learning surrogate model for topology optimization,” IEEE
on-Chip FPGA devices for complex electrical energy systems control,” Trans. Magn., June 2021.
IEEE Ind. Electron. Mag., March 2021. [61] H. Sasaki, Y. Hidaka, and H. Igarashi, “Explainable deep neural
[39] K. Liu and Z.-Q. Zhu, “Position-offset-based parameter estimation network for design of electric motors,” IEEE Trans. Magn., June 2021.
using the ADALINE NN for condition monitoring of permanent- [62] Y. Li, G. Lei, G. Bramerdorfer, S. Peng, X. Sun, and J. Zhu, “Machine
magnet synchronous machines,” IEEE Trans. Ind. Electron., vol. 62, learning for design optimization of electromagnetic devices: Recent
no. 4, pp. 2372–2383, April 2015. developments and future directions,” Applied Sciences, vol. 11, no. 4,
[40] R. R. Kumar, G. Cirrincione, M. Cirrincione, A. Tortella, and M. An- p. 1627, 2021.
driollo, “A topological neural-based scheme for classification of faults [63] J. Saha, D. Hazarika, N. B. Y. Gorla, and S. K. Panda, “Machine
in induction machines,” IEEE Trans. Ind. Appl., vol. 57, no. 1, pp. learning aided optimization framework for design of medium-voltage
272–283, Jan./Feb. 2021. grid-connected solid-state-transformers,” IEEE Trans. Emerg. Sel. Top-
[41] B. Bengherbia, R. Kara, A. Toubal, M. O. Zmirli, S. Chadli, and ics Power Electron., 2021.
P. Wira, “FPGA implementation of a wireless sensor node with a built- [64] Y.-m. You, “Multi-objective optimal design of permanent magnet
in ADALINE neural network coprocessor for vibration analysis and synchronous motor for electric vehicle based on deep learning,” Applied
fault diagnosis in machine condition monitoring,” Measurement, vol. Sciences, vol. 10, no. 2, p. 482, 2020.
163, p. 107960, 2020. [65] N. Gabdullin, S. Madanzadeh, and A. Vilkin, “Towards end-to-end deep
[42] S. Zhang, S. Zhang, B. Wang, and T. G. Habetler, “Deep learning learning performance analysis of electric motors,” in Actuators, vol. 10,
algorithms for bearing fault diagnostics—a comprehensive review,” no. 2, 2021, p. 28.
IEEE Access, vol. 8, pp. 29 857–29 881, 2020. [66] A. Mayr, M. Weigelt, M. Masuch, M. Meiners, F. Hüttel, and J. Franke,
[43] A. G. Nath, S. S. Udmale, and S. K. Singh, “Role of artificial “Application scenarios of artificial intelligence in electric drives pro-
intelligence in rotor fault diagnosis: A comprehensive review,” Artificial duction,” Procedia Manufacturing, vol. 24, pp. 40–47, 2018.
Intelligence Review, pp. 1–60, 2020. [67] A. Mayr, D. Kißkalt, A. Lomakin, K. Graichen, and J. Franke, “To-
[44] J. Lee and J.-I. Ha, “Temperature estimation of PMSM using wards an intelligent linear winding process through sensor integration
a difference-estimating feedforward neural network,” IEEE Access, and machine learning techniques,” Procedia CIRP, vol. 96, pp. 80–85,
vol. 8, pp. 130 855–130 865, 2020. 2021.
[45] Y. Cai, Y. Cen, G. Cen, X. Yao, C. Zhao, and Y. Zhang, “Temperature [68] M. Schenke, W. Kirchgässner, and O. Wallscheid, “Controller design
prediction of PMSMs using pseudo-siamese nested LSTM,” World for electrical drives by deep reinforcement learning: A proof of
Electric Vehicle Journal, vol. 12, no. 2, p. 57, 2021. concept,” IEEE Trans. Ind. Informat., vol. 16, no. 7, pp. 4650–4658,
[46] S. Zhang, S. Li, R. G. Harley, and T. G. Habetler, “An efficient multi- 2019.
objective bayesian optimization approach for the automated analytical [69] A. Traue, G. Book, W. Kirchgässner, and O. Wallscheid, “Toward
design of switched reluctance machines,” in Proc. IEEE Energy Conv. a reinforcement learning environment toolbox for intelligent electric
Congr. Expo. (ECCE), 2018, pp. 4290–4295. motor control,” IEEE Trans. Neural Netw. Learn. Syst., 2020.
[47] ——, “Visualization and data mining of multi-objective electric ma- [70] P. Balakrishna, G. Book, W. Kirchgässner, M. Schenke, A. Traue, and
chine optimizations with self-organizing maps: A case study on O. Wallscheid, “gym-electric-motor (gem): A python toolbox for the
switched reluctance machines,” in Proc. IEEE Energy Conv. Congr. simulation of electric drive systems,” Journal of Open Source Software,
Expo. (ECCE), 2018, pp. 4296–4302. vol. 6, no. 58, p. 2498, 2021.
[48] A. Khan, V. Ghorbanian, and D. Lowther, “Deep learning for magnetic [71] S. Hanke, O. Wallscheid, and J. Böcker, “Data set description: Identi-
field estimation,” IEEE Trans. Magn., vol. 55, no. 6, pp. 1–4, June 2019. fying the physics behind an electric motor–data-driven learning of the
[49] S. Doi, H. Sasaki, and H. Igarashi, “Multi-objective topology optimiza- electrical behavior (part i),” arXiv preprint arXiv:2003.07273, 2020.
tion of rotating machines using deep learning,” IEEE Trans. Magn., [72] ——, “Data set description: Identifying the physics behind an electric
vol. 55, no. 6, pp. 1–5, June 2019. motor–data-driven learning of the electrical behavior (part ii),” arXiv
[50] H. Sasaki and H. Igarashi, “Topology optimization accelerated by deep preprint arXiv:2003.06268, 2020.
learning,” IEEE Trans. Magn., vol. 55, no. 6, pp. 1–5, June 2019. [73] M. Schenke and O. Wallscheid, “A deep q-learning direct torque
[51] S. Zhang, S. Zhang, S. Li, L. Du, and T. G. Habetler, “Visualization of controller for permanent magnet synchronous motors,” IEEE Open J.
multi-objective switched reluctance machine optimization at multiple Ind. Electron. Soc.,, 2021.
operating conditions with t-sne,” in Proc. IEEE Energy Conv. Congr. [74] G. Book, A. Traue, P. Balakrishna, A. Brosch, M. Schenke, S. Hanke,
Expo. (ECCE), 2019, pp. 3793–3798. W. Kirchgässner, and O. Wallscheid, “Transferring online reinforce-
[52] T. Guillod, P. Papamanolis, and J. W. Kolar, “Artificial neural network ment learning for electric motor control from simulation to real-world
(ANN) based fast and accurate inductor modeling and design,” IEEE experiments,” IEEE Open J. Power Electron., vol. 2, pp. 187–201,
Open J. Power Electron., vol. 1, pp. 284–299, 2020. 2021.

21
[75] T. Schindler, L. Foss, and A. Dietz, “Comparison of reinforcement motor performance enhancement,” Nature communications, vol. 11,
learning algorithms for speed ripple reduction of permanent magnet no. 1, pp. 1–11, 2020.
synchronous motor,” in IKMT 2019-Innovative small Drives and Micro- [96] M. Hannan, J. A. Ali, A. Mohamed, and A. Hussain, “Optimization
Motor Systems; 12. ETG/GMM-Symposium, 2019, pp. 1–6. techniques to enhance the performance of induction motor drives: A
[76] S. Bhattacharjee, S. Halder, A. Balamurali, M. Towhidi, L. V. Iyer, review,” Renewable and Sustainable Energy Reviews, vol. 81, pp. 1611–
and N. C. Kar, “An advanced policy gradient based vector control of 1626, 2018.
PMSM for EV application,” in 2020 10th International Electric Drives [97] R. Colom, S. Karama, R. E. Jung, and R. J. Haier, “Human intelligence
Production Conference (EDPC), 2020, pp. 1–5. and brain networks,” Dialogues in Clinical Neuroscience, 2022.
[77] F. F. El-Sousy, M. M. Amin, G. A. A. Aziz, and A. Al-Durra, “Adaptive [98] K. J. Åström and T. Hägglund, “The future of PID control,” Control
neural-network optimal tracking control for permanent-magnet syn- engineering practice, vol. 9, no. 11, pp. 1163–1175, 2001.
chronous motor drive system via adaptive dynamic programming,” in [99] R.-J. Wai, R.-Y. Duan, J.-D. Lee, and H.-H. Chang, “Wavelet neural
2020 IEEE Industry Applications Society Annual Meeting, 2020, pp. network control for induction motor drive using sliding-mode design
1–8. technique,” IEEE Trans. Ind. Electron., vol. 50, no. 4, pp. 733–748,
[78] H. Alharkan, S. Saadatmand, M. Ferdowsi, and P. Shamsi, “Optimal Aug. 2003.
tracking current control of switched reluctance motor drives using [100] T. Huber, W. Peters, and J. Böcker, “Voltage controller for flux
reinforcement q-learning scheduling,” IEEE Access, vol. 9, pp. 9926– weakening operation of interior permanent magnet synchronous motor
9936, 2021. in automotive traction applications,” in IEEE International Electric
[79] K. P. Seng, P. J. Lee, and L. M. Ang, “Embedded intelligence on FPGA: Machines & Drives Conference (IEMDC), 2015.
Survey, applications and challenges,” Electronics, vol. 10, no. 8, p. 895, [101] T.-J. Ren and T.-C. Chen, “Robust speed-controlled induction motor
2021. drive based on recurrent neural network,” Electric power systems
[80] W. Kirchgässner, M. Schenke, O. Wallscheid, and D. Weber, research, vol. 76, no. 12, pp. 1064–1074, Feb. 2006.
“Reinforcement learning course material,” 2020, Paderborn University. [102] A. Rubaai and M. D. Kankam, “Adaptive tracking controller for
[Online]. Available: https://github.com/upb-lea/reinforcement learn- induction motor drives using online training of neural networks,” IEEE
ing course materials Trans. Ind. Appl., vol. 36, no. 5, pp. 1285–1294, Sept./Oct. 2000.
[81] P. Vas, Artificial-intelligence-based electrical machines and drives: [103] A. Rubaai, R. Kotaru, and M. D. Kankam, “Online training of parallel
Application of fuzzy, neural, fuzzy-neural, and genetic-algorithm-based neural network estimators for control of induction motors,” IEEE Trans.
techniques. Oxford university press, 1999, vol. 45. Ind. Appl., vol. 37, no. 5, pp. 1512–1521, Sept./Oct. 2001.
[82] P. Z. Grabowski, M. P. Kazmierkowski, B. K. Bose, and F. Blaabjerg, [104] X. Fu and S. Li, “A novel neural network vector control technique for
“A simple direct-torque neuro-fuzzy control of PWM-inverter-fed in- induction motor drive,” IEEE Trans. Energy Convers., vol. 30, no. 4,
duction motor drive,” IEEE Trans. Ind. Electron., vol. 47, no. 4, pp. pp. 1428–1437, Dec. 2015.
863–870, Aug. 2000. [105] A. Ba-Razzouk, A. Cheriti, G. Olivier, and P. Sicard, “Field-oriented
[83] S. M. Gadoue, D. Giaouris, and J. W. Finch, “Genetic algorithm control of induction motors using neural-network decouplers,” IEEE
optimized PI and fuzzy sliding mode speed control for DTC drives,” Trans. Power Electron., vol. 12, no. 4, pp. 752–763, July 1997.
in World Congress on Engineering, 2007, pp. 475–480. [106] P. Marino, M. Milano, and F. Vasca, “Linear quadratic state feedback
[84] M. Suetake, I. N. da Silva, and A. Goedtel, “Embedded DSP-based and robust neural network estimator for field-oriented-controlled induc-
compact fuzzy system and its application for induction-motor v/f tion motors,” IEEE Trans. Ind. Electron., vol. 46, no. 1, pp. 150–161,
speed control,” IEEE Trans. Ind. Electron., vol. 58, no. 3, pp. 750– Feb. 1999.
760, March 2011. [107] D. Zhang and H. Li, “A stochastic-based FPGA controller for an
[85] N. V. Naik, A. Panda, and S. P. Singh, “A three-level fuzzy-2 DTC induction motor drive with integrated neural network algorithms,” IEEE
of induction motor drive using SVPWM,” IEEE Trans. Ind. Electron., Trans. Ind. Electron., vol. 55, no. 2, pp. 551–561, Feb. 2008.
vol. 63, no. 3, pp. 1467–1479, 2015. [108] M. Stender, O. Wallscheid, and J. Böcker, “Accurate torque estimation
[86] S. Singh, A. K. Panda et al., “An interval type-2 fuzzy-based DTC of for induction motors by utilizing a hybrid machine learning approach,”
IMD using hybrid duty ratio control,” IEEE Trans. Power Electron., in IEEE International Power Electronics and Motion Control Confer-
vol. 35, no. 8, pp. 8443–8451, 2020. ence (PEMC), 2021.
[87] F.-J. Lin, R.-J. Wai, C.-H. Lin, and D.-C. Liu, “Decoupled stator-flux- [109] C. Schauder, “Adaptive speed identification for vector control of
oriented induction motor drive with fuzzy neural network uncertainty induction motors without rotational transducers,” in Conference Record
observer,” IEEE Trans. Ind. Electron., vol. 47, no. 2, pp. 356–367, of the IEEE Industry Applications Society Annual Meeting,, 1989, pp.
April 2000. 493–499.
[88] S. M. Gadoue, D. Giaouris, and J. Finch, “Artificial intelligence-based [110] P. Vaclavek, P. Blaha, and I. Herman, “AC drive observability analysis,”
speed control of DTC induction motor drives—A comparative study,” IEEE Trans. Ind. Electron., vol. 60, no. 8, pp. 3047–3059, Aug. 2013.
Electric Power Systems Research, vol. 79, no. 1, pp. 210–219, 2009. [111] P. L. Jansen and R. D. Lorenz, “A physically insightful approach to
[89] S. M. Gadoue, D. Giaouris, and J. W. Finch, “MRAS sensorless vector the design and accuracy assessment of flux observers for field oriented
control of an induction motor using new sliding-mode and fuzzy-logic induction machine drives,” IEEE Trans. Ind. Appl., vol. 30, no. 1, pp.
adaptation mechanisms,” IEEE Trans. Energy Convers., vol. 25, no. 2, 101–110, 1994.
pp. 394–402, June 2010. [112] Y. B. Zbede, S. M. Gadoue, and D. J. Atkinson, “Model predictive
[90] T. Ramesh, A. K. Panda, and S. S. Kumar, “Type-2 fuzzy logic control MRAS estimator for sensorless induction motor drives,” IEEE Trans.
based MRAS speed estimator for speed sensorless direct torque and Ind. Electron., vol. 63, no. 6, pp. 3511–3521, June 2016.
flux control of an induction motor drive,” ISA Trans., vol. 57, pp. 262– [113] P. Vas, Sensorless vector and direct torque control. Oxford Univ.
275, 2015. Press, 1998.
[91] M. Demirtas, “DSP-based sliding mode speed control of induction [114] L. Ben-Brahim, S. Tadakuma, and A. Akdag, “Speed control of in-
motor using neuro-genetic structure,” Expert Systems with Applications, duction motor without rotational transducers,” IEEE Trans. Ind. Appl.,
vol. 36, no. 3, pp. 5533–5540, 2009. vol. 35, no. 4, pp. 844–850, Jan./Feb. 1999.
[92] F.-J. Lin, P.-K. Huang, and W.-D. Chou, “Recurrent-fuzzy-neural- [115] M. Cirrincione, M. Pucci, G. Cirrincione, and G.-A. Capolino, “A
network-controlled linear induction motor servo drive using genetic new TLS-based MRAS speed estimation with adaptive integration for
algorithms,” IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 1449–1461, high-performance induction machine drives,” IEEE Trans. Ind. Appl.,
June 2007. vol. 40, no. 4, pp. 1116–1137, July/Aug. 2004.
[93] F.-J. Lin, L.-T. Teng, J.-W. Lin, and S.-Y. Chen, “Recurrent functional- [116] M. Cirrincione and M. Pucci, “An MRAS-based sensorless high-
link-based fuzzy-neural-network-controlled induction-generator system performance induction motor drive with a predictive adaptive model,”
using improved particle swarm optimization,” IEEE Trans. Ind. Elec- IEEE Trans. Ind. Electron., vol. 52, no. 2, pp. 532–551, April 2005.
tron., vol. 56, no. 5, pp. 1557–1577, May 2009. [117] M. Cirrincione, A. Accetta, M. Pucci, and G. Vitale, “MRAS speed
[94] M. A. Hannan, J. A. Ali, A. Mohamed, U. A. U. Amirulddin, N. M. L. observer for high-performance linear induction motor drives based on
Tan, and M. N. Uddin, “Quantum-behaved lightning search algorithm linear neural networks,” IEEE Trans. Power Electron., vol. 28, no. 1,
to improve indirect field-oriented fuzzy-PI control for im drive,” IEEE pp. 123–134, Jan. 2013.
Trans. Ind. Appl., vol. 54, no. 4, pp. 3793–3805, 2018. [118] S. M. Gadoue, D. Giaouris, and J. W. Finch, “Sensorless control
[95] M. Hannan, J. A. Ali, M. H. Lipu, A. Mohamed, P. J. Ker, T. I. of induction motor drives at very low and zero speeds using neural
Mahlia, M. Mansor, A. Hussain, K. M. Muttaqi, and Z. Dong, “Role of network flux observers,” IEEE Trans. Ind. Electron., vol. 56, no. 8, pp.
optimization algorithms based fuzzy controller in achieving induction 3029–3039, Aug. 2009.

22
[119] S.-H. Kim, T.-S. Park, J.-Y. Yoo, and G.-T. Park, “Speed-sensorless [139] S. K. Mondal, J. O. Pinto, and B. K. Bose, “A neural-network-based
vector control of an induction motor using neural network speed space-vector PWM controller for a three-level voltage-fed inverter
estimation,” IEEE Trans. Ind. Electron., vol. 48, no. 3, pp. 609–614, induction motor drive,” IEEE Trans. Ind. Appl., vol. 38, no. 3, pp.
June 2001. 660–669, May/June 2002.
[120] E. Abdin, G. Ghoneem, H. Diab, and S. Deraz, “Efficiency optimization [140] T. M. Wolbank, J. L. Machl, and T. Jager, “Combination of signal
of a vector controlled induction motor drive using an artificial neural injection and neural networks for sensorless control of inverter fed
network,” in IECON’03. 29th Annual Conference of the IEEE Industrial induction machines,” in 2004 IEEE 35th Annual Power Electronics
Electronics Society (IEEE Cat. No. 03CH37468), vol. 3, 2003, pp. Specialists Conference (IEEE Cat. No. 04CH37551), vol. 3, 2004, pp.
2543–2548. 2300–2305.
[121] B. Pryymak, J. M. Moreno-Eguilaz, and J. Peracaula, “Neural network [141] T. M. Wolbank, M. A. Vogelsberger, R. Stumberger, S. Mohagheghi,
flux optimization using a model of losses in induction motor drives,” T. G. Habetler, and R. G. Harley, “Comparison of neural network
Mathematics and computers in simulation, vol. 71, no. 4-6, pp. 290– types and learning methods for self commissioning of speed sensor-
298, April 2006. less controlled induction machines,” in 2007 IEEE Power Electronics
[122] O. S. Ebrahim, M. A. Badr, A. S. Elgendy, and P. K. Jain, “ANN- Specialists Conference, 2007, pp. 1955–1960.
based optimal energy control of induction motor drive in pumping [142] P. Garcia, F. Briz, D. Raca, and R. D. Lorenz, “Saliency-tracking-based
applications,” IEEE Trans. Energy Convers., vol. 25, no. 3, pp. 652– sensorless control of AC machines using structured neural networks,”
660, Sept. 2010. IEEE Trans. Ind. Appl., vol. 43, no. 1, pp. 77–86, Jan./Feb. 2007.
[123] C.-Y. Huang, T.-C. Chen, and C.-L. Huang, “Robust control of induc- [143] P. Garcia, D. Reigosa, F. Briz, D. Raca, and R. D. Lorenz, “Automatic
tion motor with a neural-network load torque estimator and a neural- self-commissioning for secondary-saliencies decoupling in sensorless-
network identification,” IEEE Trans. Ind. Electron., vol. 46, no. 5, pp. controlled AC machines using structured neural networks,” in 2007
990–998, Oct. 1999. IEEE International Symposium on Industrial Electronics, 2007, pp.
[124] T.-T. Sheu and T.-C. Chen, “Self-tuning control of induction motor 2284–2289.
drive using neural network identifier,” IEEE Trans. Energy Convers., [144] B. Karanayil, M. F. Rahman, and C. Grantham, “Stator and rotor
vol. 14, no. 4, pp. 881–886, Dec. 1999. resistance observers for induction motor drive using fuzzy logic and
[125] E. Quintero-Manriquez, E. N. Sanchez, R. G. Harley, S. Li, and R. A. artificial neural networks,” IEEE Trans. Energy Convers., vol. 20, no. 4,
Felix, “Neural inverse optimal control implementation for induction pp. 771–780, Dec. 2005.
motors via rapid control prototyping,” IEEE Trans. Power Electron., [145] ——, “Online stator and rotor resistance estimation scheme using arti-
vol. 34, no. 6, pp. 5981–5992, 2018. ficial neural networks for vector controlled speed sensorless induction
[126] S. M. Gadoue, D. Giaouris, and J. Finch, “A neural network based stator motor drive,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 167–176,
current MRAS observer for speed sensorless induction motor drives,” Feb. 2007.
in 2008 IEEE International Symposium on Industrial Electronics, 2008, [146] M. Wlas, Z. Krzeminski, and H. A. Toliyat, “Neural-network-based
pp. 650–655. parameter estimations of induction motors,” IEEE Trans. Ind. Electron.,
[127] T. Orlowska-Kowalska, M. Dybkowski, and K. Szabat, “Adaptive vol. 55, no. 4, pp. 1783–1794, April 2008.
sliding-mode neuro-fuzzy control of the two-mass induction motor [147] A. Bechouche, H. Sediki, D. O. Abdeslam, and S. Haddad, “A novel
drive without mechanical sensors,” IEEE Trans. Ind. Electron., vol. 57, method for identifying parameters of induction motors at standstill
no. 2, pp. 553–564, Feb. 2010. using ADALINE,” IEEE Trans. Energy Convers., vol. 27, no. 1, pp.
[128] M. Cirrincione, M. Pucci, G. Cirrincione, and G.-A. Capolino, “An 105–116, March 2012.
adaptive speed observer based on a new total least-squares neuron for [148] B. Fan, Z. Yang, W. Xu, and X. Wang, “Rotor resistance online
induction machine drives,” IEEE Trans. Ind. Appl., vol. 42, no. 1, pp. identification of vector controlled induction motor based on neural
89–104, Jan./Feb. 2006. network,” Mathematical Problems in Engineering, vol. 2014, 2014.
[129] ——, “Sensorless control of induction motors by reduced order ob- [149] D. R. Seidl, “Motion and motor control using structured neural net-
server with mca exin+ based adaptive speed estimation,” IEEE Trans. works,” Ph.D. dissertation, Univ. Wisconsin-Madison, 1996.
Ind. Electron., vol. 54, no. 1, pp. 150–166, Feb. 2007. [150] F. Briz, M. W. Degner, P. Garcı́a, and J. M. Guerrero, “Rotor position
[130] A. Accetta, M. Cirrincione, M. Pucci, and G. Vitale, “Neural sensorless estimation of AC machines using the zero-sequence carrier-signal
control of linear induction motors by a full-order luenberger observer voltage,” IEEE Trans. Ind. Appl., vol. 41, no. 6, pp. 1637–1646, 2005.
considering the end effects,” IEEE Trans. Ind. Appl., vol. 50, no. 3, pp. [151] G. Parascandolo, H. Huttunen, and T. Virtanen, “Taming the waves:
1891–1904, May/June 2014. Sine as activation function in deep neural networks,” 2016.
[131] H. Abu-Rub, J. Guzinski, Z. Krzeminski, and H. A. Toliyat, “Speed [152] J. M. Gutierrez-Villalobos, J. Rodrı́guez-Reséndiz, E. A. Rivas-Araiza,
observer system for advanced sensorless control of induction motor,” and V. Mucino, “A review of parameter estimators and controllers for
IEEE Trans. Energy Convers., vol. 18, no. 2, pp. 219–224, 2003. induction motors based on artificial neural networks,” Neurocomputing,
[132] M. Wlas, Z. Krzeminski, J. Guzinski, H. Abu-Rub, and H. A. Toliyat, vol. 118, pp. 87–100, 2013.
“Artificial-neural-network-based sensorless nonlinear control of induc- [153] M. Rahman and M. Hoque, “On-line adaptive artificial neural network
tion motors,” IEEE Trans. Energy Convers., vol. 20, no. 3, pp. 520–528, based vector control of permanent magnet synchronous motors,” IEEE
Sept. 2005. Trans. Energy Convers., vol. 13, no. 4, pp. 311–318, Dec. 1998.
[133] L. E. Da Silva, B. K. Bose, and J. O. Pinto, “Recurrent-neural-network- [154] Y. Yi, D. M. Vilathgamuwa, and M. A. Rahman, “Implementation
based implementation of a programmable cascaded low-pass filter used of an artificial-neural-network-based real-time adaptive controller for
in stator flux synthesis of vector-controlled induction motor drive,” an interior permanent-magnet motor drive,” IEEE Trans. Ind. Appl.,
IEEE Trans. Ind. Electron., vol. 46, no. 3, pp. 662–665, June 1999. vol. 39, no. 1, pp. 96–104, Jan./Feb. 2003.
[134] J. Pinto, B. K. Bose, and L. E. B. da Silva, “A stator-flux-oriented [155] L. Guo and L. Parsa, “Model reference adaptive control of five-phase
vector-controlled induction motor drive with space-vector PWM and IPM motors based on neural network,” IEEE Trans. Ind. Electron.,
flux-vector synthesis by neural networks,” IEEE Trans. Ind. Appl., vol. 59, no. 3, pp. 1500–1508, March 2012.
vol. 37, no. 5, pp. 1308–1318, Sept./Oct. 2001. [156] C. Lucas, D. Shahmirzadi, and N. Sheikholeslami, “Introducing BEL-
[135] M. Cirrincione, M. Pucci, G. Cirrincione, and G.-A. Capolino, “A BIC: Brain emotional learning based intelligent controller,” Intelligent
new adaptive integration methodology for estimating flux in induction Automation & Soft Computing, vol. 10, no. 1, pp. 11–21, 2004.
machine drives,” IEEE Trans. Power Electron., vol. 19, no. 1, pp. 25– [157] E. Daryabeigi, G. A. Markadeh, and C. Lucas, “Interior permanent
34, Jan. 2004. magnet synchronous motor (IPMSM), with a developed brain emo-
[136] J. Zhao and B. K. Bose, “Neural-network-based waveform processing tional learning based intelligent controller (BELBIC),” in 2009 IEEE
and delayless filtering in power electronics and AC drives,” IEEE Trans. International Electric Machines and Drives Conference, 2009, pp.
Ind. Electron., vol. 51, no. 5, pp. 981–991, Oct. 2004. 1633–1640.
[137] A. Bakhshai, J. Espinoza, G. Joos, and H. Jin, “A combined artificial [158] A. Rubaai, R. Kotaru, and M. D. Kankam, “A continually online-trained
neural network and DSP approach to the implementation of space neural network controller for brushless DC motor drives,” IEEE Trans.
vector modulation techniques,” in IAS’96. Conference Record of the Ind. Appl., vol. 36, no. 2, pp. 475–483, March/April 2000.
1996 IEEE Industry Applications Conference Thirty-First IAS Annual [159] A. Rubaai, D. Ricketts, and M. D. Kankam, “Development and
Meeting, vol. 2, 1996, pp. 934–940. implementation of an adaptive fuzzy-neural-network controller for
[138] J. O. Pinto, B. K. Bose, L. B. Da Silva, and M. P. Kazmierkowski, brushless drives,” IEEE Trans. Ind. Appl., vol. 38, no. 2, pp. 441–447,
“A neural-network-based space-vector PWM controller for voltage-fed March/April 2002.
inverter induction motor drive,” IEEE Trans. Ind. Appl., vol. 36, no. 6,
pp. 1628–1636, Nov./Dec. 2000.

23
[160] A. Rubaai, M. J. Castro-Sitiriche, and A. R. Ofoli, “Design and [181] Z. Chen, M. Tomita, S. Doki, and S. Okuma, “An extended electromo-
implementation of parallel fuzzy PID controller for high-performance tive force model for sensorless control of interior permanent-magnet
brushless motor drives: An integrated environment for rapid control synchronous motors,” IEEE transactions on Industrial Electronics,
prototyping,” IEEE Trans. Ind. Appl., vol. 44, no. 4, pp. 1090–1098, vol. 50, no. 2, pp. 288–295, 2003.
July/Aug. 2008. [182] G. Wang, Z. Li, G. Zhang, Y. Yu, and D. Xu, “Quadrature pll-
[161] ——, “DSP-based laboratory implementation of hybrid fuzzy-PID con- based high-order sliding-mode observer for ipmsm sensorless control
troller using genetic optimization for high-performance motor drives,” with online mtpa control strategy,” IEEE Transactions on Energy
IEEE Trans. Ind. Appl., vol. 44, no. 6, pp. 1977–1986, Nov./Dec. 2008. Conversion, vol. 28, no. 1, pp. 214–224, 2012.
[162] A. Rubaai and P. Young, “EKF-based PI-/PD-like fuzzy-neural-network [183] A. Brosch, F. Tinazzi, O. Wallscheid, M. Zigliotto, and J. Böcker,
controller for brushless drives,” IEEE Trans. Ind. Appl., vol. 47, no. 6, “Finite Set Sensorless Control With Minimum a Priori Knowledge and
pp. 2391–2401, Nov./Dec. 2011. Tuning Effort for Interior Permanent Magnet Synchronous Motors,”
[163] ——, “Hardware/software implementation of fuzzy-neural-network techrxiv, 2023.
self-learning control methods for brushless DC motor drives,” IEEE [184] W. Kirchgässner, O. Wallscheid, and J. Böcker, “Data-driven permanent
Trans. Ind. Appl., vol. 52, no. 1, pp. 414–424, Jan./Feb. 2016. magnet temperature estimation in synchronous motors with supervised
[164] R.-J. Wai, “Total sliding-mode controller for PM synchronous servo machine learning: A benchmark,” IEEE Trans. Energy Convers., 2021.
motor drive using recurrent fuzzy neural network,” IEEE Trans. Ind. [185] ——, “Estimating electric motor temperatures with deep residual
Electron., vol. 48, no. 5, pp. 926–944, Oct. 2001. machine learning,” IEEE Trans. Power Electron., vol. 36, no. 7, pp.
[165] F.-J. Lin, L.-T. Teng, and H. Chu, “A robust recurrent wavelet neural 7480–7488, 2020.
network controller with improved particle swarm optimization for [186] ——, “Thermal neural networks: lumped-parameter thermal modeling
linear synchronous motor drive,” IEEE Trans. Power Electron., vol. 23, with state-space machine learning,” Engineering Applications of Arti-
no. 6, pp. 3067–3078, Nov. 2008. ficial Intelligence, vol. 117, p. 105537, 2023.
[166] F.-J. Lin, J.-C. Hwang, P.-H. Chou, and Y.-C. Hung, “FPGA-based [187] O. Wallscheid, “Thermal monitoring of electric motors: State-of-the-art
intelligent-complementary sliding-mode control for PMLSM servo- review and future challenges,” IEEE Open J. Ind. Appl, 2021.
drive system,” IEEE Trans. Power Electron., vol. 25, no. 10, pp. 2573– [188] T. Liu, I. Husain, and M. Elbuluk, “Torque ripple minimization with on-
2587, Oct. 2010. line parameter estimation using neural networks in permanent magnet
[167] F.-J. Lin and P.-H. Chou, “Adaptive control of two-axis motion control synchronous motors,” in Conference Record of 1998 IEEE Industry
system using interval type-2 fuzzy neural network,” IEEE Trans. Ind. Applications Conference. Thirty-Third IAS Annual Meeting (Cat. No.
Electron., vol. 56, no. 1, pp. 178–193, Jan. 2009. 98CH36242), vol. 1, 1998, pp. 35–40.
[168] K. A. Abuhasel, F. F. El-Sousy, M. F. El-Naggar, and A. Abu- [189] Y. A.-R. I. M. Mohamed, “A novel direct instantaneous torque and flux
Siada, “Adaptive RCMAC neural network dynamic surface control control with an ADALINE-based motor model for a high performance
for permanent-magnet synchronous motors driven two-axis XY table,” DD-PMSM,” IEEE Trans. Power Electron., vol. 22, no. 5, pp. 2042–
IEEE Access, vol. 7, pp. 38 068–38 084, 2019. 2049, Sept. 2007.
[169] F. F. El-Sousy, M. F. El-Naggar, M. Amin, A. Abu-Siada, and [190] Z. Wang, M. Yang, L. Gao, Z. Wang, G. Zhang, H. Wang, and X. Gu,
K. A. Abuhasel, “Robust adaptive neural-network backstepping control “Deadbeat predictive current control of permanent magnet synchronous
design for high-speed permanent-magnet synchronous motor drives: motor based on variable step-size ADALINE neural network parameter
Theory and experiments,” IEEE Access, vol. 7, pp. 99 327–99 348, identification,” IET Electric Power Applications, vol. 14, no. 11, pp.
2019. 2007–2015, 2020.
[170] M. Linke, R. Kennel, and J. Holtz, “Sensorless position control of [191] A. Brosch, S. Hanke, O. Wallscheid, and J. Böcker, “Data-driven re-
permanent magnet synchronous machines without limitation at zero cursive least squares estimation for model predictive current control of
speed,” in IEEE 2002 28th Annual Conference of the Industrial permanent magnet synchronous motors,” IEEE Trans. Power Electron.,
Electronics Society. IECON 02, vol. 1, 2002, pp. 674–679. vol. 36, no. 2, pp. 2179–2190, 2020.
[171] ——, “Sensorless speed and position control of synchronous machines [192] A. Brosch, O. Wallscheid, and J. Böcker, “Torque and inductances
using alternating carrier injection,” in IEEE International Electric estimation for finite model predictive control of highly utilized per-
Machines and Drives Conference, 2003. IEMDC’03., vol. 2, 2003, pp. manent magnet synchronous motors,” IEEE Transactions on Industrial
1211–1217. Informatics, vol. 17, no. 12, pp. 8080–8091, 2021.
[172] J. Holtz, “Acquisition of position error and magnet polarity for sen- [193] F. Tinazzi, P. G. Carlet, S. Bolognani, and M. Zigliotto, “Motor
sorless control of PM synchronous machines,” IEEE Trans. Ind. Appl., parameter-free predictive current control of synchronous motors by
vol. 44, no. 4, pp. 1172–1180, July/Aug. 2008. recursive least-square self-commissioning model,” IEEE Transactions
[173] D. Xu, B. Wang, G. Zhang, G. Wang, and Y. Yu, “A review of on Industrial Electronics, vol. 67, no. 11, pp. 9093–9100, 2020.
sensorless control methods for AC motor drives,” CES Trans. Electri. [194] A. Brosch, O. Wallscheid, and J. Böcker, “Long-term memory recursive
Mach. Syst., vol. 2, no. 1, pp. 104–115, 2018. least squares online identification of highly utilized permanent magnet
[174] M. E. Elbuluk, L. Tong, and I. Husain, “Neural-network-based model synchronous motors for finite-control-set model predictive control,”
reference adaptive systems for high-performance motor drives and IEEE Transactions on Power Electronics, vol. 38, no. 2, pp. 1451–
motion controls,” IEEE Trans. Ind. Appl., vol. 38, no. 3, pp. 879–886, 1467, 2023.
May/June 2002. [195] H. Jie, G. Zheng, J. Zou, X. Xin, and L. Guo, “Adaptive decoupling
[175] F.-J. Lin, Y.-C. Hung, J.-M. Chen, and C.-M. Yeh, “Sensorless IPMSM control using radial basis function neural network for permanent
drive system using saliency back-EMF-based intelligent torque ob- magnet synchronous motor considering uncertain and time-varying
server with MTPA control,” IEEE Trans. Ind. Informat., vol. 10, no. 2, parameters,” IEEE Access, vol. 8, pp. 112 323–112 332, 2020.
pp. 1226–1241, May 2014. [196] ——, “Speed regulation based on adaptive control and RBFNN for
[176] G. Zhang, G. Wang, D. Xu, and N. Zhao, “ADALINE-network-based PMSM considering parametric uncertainty and load fluctuation,” IEEE
PLL for position sensorless interior permanent magnet synchronous Access, vol. 8, pp. 190 147–190 159, 2020.
motor drives,” IEEE Trans. Power Electron., vol. 31, no. 2, pp. 1450– [197] M. S. Rafaq and J.-W. Jung, “A comprehensive review of state-of-the-
1460, Feb. 2016. art parameter estimation techniques for permanent magnet synchronous
[177] Z. Makni and W. Zine, “Rotor position estimator based on machine motors in wide speed range,” IEEE Trans. Ind. Informat., vol. 16, no. 7,
learning,” in IECON 2016-42nd Annual Conference of the IEEE pp. 4747–4758, July 2020.
Industrial Electronics Society, 2016, pp. 6687–6692. [198] M. Stender, O. Wallscheid, and J. Böcker, “Data set description: Three-
[178] W. Zine, Z. Makni, E. Monmasson, K. Chen, L. Idkhajine, and phase IGBT two-level inverter for electrical drives,” Jul. 2020, DOI:
B. Condamin, “Hybrid sensorless control strategy for EV applications 10.13140/RG.2.2.23335.37280.
based on high frequency signal injection and machine learning,” in [199] M. Stender, O. Wallscheid, and J. Boecker, “Comparison of gray-
2017 IEEE Vehicle Power and Propulsion Conference (VPPC), 2017, box and black-box two-level three-phase inverter models for electrical
pp. 1–5. drives,” IEEE Trans. Ind. Electron., vol. 68, no. 9, pp. 8646–8656,
[179] W. Zine, Z. Makni, E. Monmasson, L. Idkhajine, and B. Condamin, 2020.
“Interests and limits of machine learning-based neural networks for [200] M. Stender, O. Wallscheid, and J. Böcker, “Gray-box loss model
rotor position estimation in EV traction drives,” IEEE Trans. Ind. for induction motor drives,” in 2021 IEEE 19th International Power
Informat., vol. 14, no. 5, pp. 1942–1951, May 2018. Electronics and Motion Control Conference (PEMC), 2021, pp. 447–
[180] M.-S. Wang and T.-M. Tsai, “Sliding mode and neural network control 453.
of sensorless PMSM controlled system for power consumption and
performance improvement,” Energies, vol. 10, no. 11, p. 1780, 2017.

24
[201] B. Gou, Y. Xu, Y. Xia, G. Wilson, and S. Liu, “An intelligent time- Conference on Devices, Circuits and Systems, 2004., vol. 1, 2004, pp.
adaptive data-driven method for sensor fault diagnosis in induction 225–229.
motor drive system,” IEEE Trans. Ind. Electron., vol. 66, no. 12, pp. [222] J. Restrepo, J. Viola, R. Harley, and T. Habetler, “Induction machine
9817–9827, 2018. current loop neuro controller employing a lyapunov based training
[202] B. Gou, Y. Xu, Y. Xia, Q. Deng, and X. Ge, “An online data-driven algorithm,” in 2007 IEEE Power Engineering Society General Meeting,
method for simultaneous diagnosis of IGBT and current sensor fault 2007, pp. 1–8.
of three-phase PWM inverter in induction motor drives,” IEEE Trans. [223] T. Orlowska-Kowalska and M. Kaminski, “FPGA implementation of
Power Electron., vol. 35, no. 12, pp. 13 281–13 294, 2020. the multilayer neural network for the speed estimation of the two-mass
[203] Y. Xia, Y. Xu, B. Gou, and Q. Deng, “A learning-based method for drive system,” IEEE Trans. Ind. Informat., vol. 7, no. 3, pp. 436–445,
speed sensor fault diagnosis of induction motor drive systems,” IEEE Aug. 2011.
Trans. Instrum. Meas., 2021. [224] M. Kaminski and T. Orlowska-Kowalska, “FPGA implementation of
[204] R. Argawal, D. Kalel, M. Harshit, A. D. Domnic, and R. R. Singh, ADALINE-based speed controller in a two-mass system,” IEEE Trans.
“Sensor fault detection using machine learning technique for automo- Ind. Informat., vol. 9, no. 3, pp. 1301–1311, Aug. 2013.
bile drive applications,” in 2021 National Power Electronics Conference [225] A. M. Soares, L. C. Leite, J. O. Pinto, L. E. da Silva, B. K. Bose,
(NPEC), 2021, pp. 1–6. and M. E. Romero, “Field programmable gate array (FPGA) based
[205] M. Dybkowski and K. Klimkowski, “Artificial neural network applica- neural network implementation of stator flux oriented vector control
tion for current sensors fault detection in the vector controlled induction of induction motor drive,” in 2006 IEEE International Conference on
motor drive,” Sensors, vol. 19, no. 3, p. 571, 2019. Industrial Technology, 2006, pp. 31–34.
[206] K. Jankowska and M. Dybkowski, “Design and analysis of current [226] T. Schindler and A. Dietz, “Real-time inference of neural networks
sensor fault detection mechanisms for PMSM drives based on neural on FPGAs for motor control applications,” in 2020 10th International
networks,” Designs, vol. 6, no. 1, p. 18, 2022. Electric Drives Production Conference (EDPC), 2020, pp. 1–6.
[207] M. Stender, O. Wallscheid, and J. Böcker, “Data set – three-phase IGBT [227] M. Rothmann and M. Porrmann, “A survey of domain-specific archi-
two-level inverter for electrical drives (data),” [Online]. Available: tectures for reinforcement learning,” IEEE Access, vol. 10, pp. 13 753–
https://www.kaggle.com/stender/inverter-data-set, Accessed: Jul. 2022. 13 767, 2022.
[208] S. U. Jan, Y.-D. Lee, J. Shin, and I. Koo, “Sensor fault classification [228] M. Kaminski, “Nature-inspired algorithm implemented for stable radial
based on support vector machine and statistical time-domain features,” basis function neural controller of electric drive with induction motor,”
IEEE Access, vol. 5, pp. 8682–8690, 2017. Energies, vol. 13, no. 24, p. 6541, 2020.
[209] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and [229] P. Vas, A. Stronach, and M. Neuroth, “DSP-based speed-sensorless
fault-tolerant techniques—-Part I: Fault diagnosis with model-based vector controlled induction motor drives using AI-based speed estima-
and signal-based approaches,” IEEE Trans. Ind. Electron., vol. 62, tor and two current sensors,” in 1998 Seventh International Conference
no. 6, pp. 3757–3767, 2015. on Power Electronics and Variable Speed Drives (IEE Conf. Publ. No.
[210] M. Schenke, B. Haucke-Korber, and O. Wallscheid, “Finite-Set Direct 456), 1998, pp. 442–446.
Torque Control via Edge Computing-Assisted Safe Reinforcement [230] Q. N. Le and J.-W. Jeon, “Neural-network-based low-speed-damping
Learning for a Permanent Magnet Synchronous Motor,” techrxiv, 2023. controller for stepper motor with an FPGA,” IEEE Trans. Ind. Electron.,
[211] B. Burton, R. G. Harley, G. Diana, and J. L. Rodgerson, “Implemen- vol. 57, no. 9, pp. 3167–3180, Sept. 2010.
tation of a neural network to adaptively identify and control vsi-fed [231] N. K. Quang, N. T. Hieu, and Q. Ha, “FPGA-based sensorless PMSM
induction motor stator currents,” IEEE Trans. Ind. Appl., vol. 34, no. 3, speed control using reduced-order extended Kalman filters,” IEEE
pp. 580–588, May/June 1998. Trans. Ind. Electron., vol. 61, no. 12, pp. 6574–6582, Dec. 2014.
[212] K.-K. Shyu, H.-J. Shieh, and S.-S. Fu, “Model reference adaptive speed [232] E. Monmasson, L. Idkhajine, M. N. Cirstea, I. Bahri, A. Tisan, and
control for induction motor drive using neural networks,” IEEE Trans. M. W. Naouar, “FPGAs in industrial control applications,” IEEE Trans.
Ind. Electron., vol. 45, no. 1, pp. 180–182, Feb. 1998. Ind. Informat., vol. 7, no. 2, pp. 224–243, May 2011.
[213] T.-C. Chen and T.-T. Sheu, “Model reference neural network controller [233] I. S. Mohamed, S. Rovetta, T. D. Do, T. Dragicević, and A. A. Z.
for induction motor speed control,” IEEE Trans. Energy Convers., Diab, “A neural-network-based model predictive control of three-phase
vol. 17, no. 2, pp. 157–163, June 2002. inverter with an output LC filter,” IEEE Access, vol. 7, pp. 124 737–
[214] M. Mohamadian, E. Nowicki, A. Chu, F. Ashrafzadeh, and J. Salmon, 124 749, 2019.
“DSP implementation of an artificial neural network for induction [234] H. T. Truong, B. P. Ta, Q. A. Le, D. M. Nguyen, C. T. Le, H. X.
motor control,” in CCECE’97. Canadian Conference on Electrical and Nguyen, H. T. Do, H. T. Nguyen, and K. P. Tran, “Light-weight
Computer Engineering. Engineering Innovation: Voyage of Discovery. federated learning-based anomaly detection for time-series data in
Conference Proc., vol. 2, 1997, pp. 435–437. industrial control systems,” Computers in Industry, vol. 140, p. 103692,
[215] M. Mohamadian, E. Nowicki, F. Ashrafzadeh, A. Chu, R. Sachdeva, 2022.
and E. Evanik, “A novel neural network controller and its efficient DSP [235] N. Lehment, F. Kaelber, and F. Jespers, “The interplay between silicon
implementation for vector-controlled induction motor drives,” IEEE capability and system architecture for cognitive power systems,” in
Trans. Ind. Appl., vol. 39, no. 6, pp. 1622–1629, Nov./Dec. 2003. PCIM Europe 2022; International Exhibition and Conference for
[216] B. Burton and R. Harley, “Linear speed-up parallel implementation of Power Electronics, Intelligent Motion, Renewable Energy and Energy
continually online trained neural networks for identification and control Management. VDE, 2022, pp. 1–10.
of fast processes [induction motor control],” in IAS’96. Conference [236] S. Gawde, S. Patil, S. Kumar, P. Kamat, K. Kotecha, and A. Abraham,
Record of the 1996 IEEE Industry Applications Conference Thirty-First “Multi-fault diagnosis of industrial rotating machines using data-driven
IAS Annual Meeting, vol. 3, 1996, pp. 1718–1724. approach: A review of two decades of research,” arXiv preprint
[217] B. Burton, F. Kamran, R. G. Harley, T. G. Habetler, M. A. Brooke, arXiv:2206.14153, 2022.
and R. Poddar, “Identification and control of induction motor stator [237] P. Blaha, M. Kozovsky, Z. Havranek, M. Dosedel, F. Wotawa, D. Kauf-
currents using fast on-line random training of a neural network,” IEEE mann, A. Amukhtar, I. Nica, F. Klück, and H. Felbinger, “Real-time
Trans. Ind. Appl., vol. 33, no. 3, pp. 697–704, May/June 1997. predictive maintenance–artificial neural network based diagnosis,” in
[218] B. Burton and R. G. Harley, “Reducing the computational demands Artificial Intelligence for Digitising Industry–Applications. River
of continually online-trained artificial neural networks for system Publishers, 2022, pp. 83–101.
identification and control of fast processes,” IEEE Trans. Ind. Appl., [238] Y. Tao, R. Ma, M.-L. Shyu, and S.-C. Chen, “Challenges in energy-
vol. 34, no. 3, pp. 589–596, May/June 1998. efficient deep neural network training with FPGA,” in Proceedings of
[219] B. Burton, R. G. Harley, and T. G. Habetler, “High bandwidth direct the IEEE/CVF Conference on Computer Vision and Pattern Recognition
adaptive neurocontrol of induction motor current and speed using Workshops, 2020, pp. 400–401.
continual online random weight change training,” in 30th Annual [239] M. Dendaluce Jahnke, F. Cosco, R. Novickis, J. Perez Rastelli, and
IEEE Power Electronics Specialists Conference. Record.(Cat. No. V. Gomez-Garay, “Efficient neural network implementations on parallel
99CH36321), vol. 1, 1999, pp. 488–494. embedded platforms applied to real-time torque-vectoring optimization
[220] J. Restrepo, B. Burton, R. Harley, and T. Habetler, “Practical imple- using predictions for multi-motor electric vehicles,” Electronics, vol. 8,
mentation of a neuro controller using a DSP based system,” in Proc. of no. 2, p. 250, 2019.
the Fifth IEEE International Caracas Conference on Devices, Circuits [240] Xilinx, “Zynq-7000 SoC Data Sheet: Overview (DS190),” [On-
and Systems, 2004., vol. 1, 2004, pp. 293–297. line]. Available: https://docs.xilinx.com/v/u/en-US/ds190-Zynq-7000-
[221] ——, “Ann based current control of a vsi fed AC machine using line Overview, Accessed: Jul. 2022.
coordinates,” in Proceedings of the Fifth IEEE International Caracas

25
[241] ——, “Zynq UltraScale+ MPSoC Data Sheet: Overview (DS891),” [253] R. Ito, M. Tsukada, and H. Matsutani, “An on-device federated learning
[Online]. Available: https://docs.xilinx.com/v/u/en-US/ds891-zynq- approach for cooperative model update between edge devices,” IEEE
ultrascale-plus-overview, Accessed: Jul. 2022. Access, vol. 9, pp. 92 986–92 998, 2021.
[242] Intel, “FPGA vs. GPU for deep learning,” [Online]. [254] H. Watanabe, M. Tsukada, and H. Matsutani, “An FPGA-based on-
Available: https://www.intel.com/content/www/us/en/artificial- device reinforcement learning approach using online sequential learn-
intelligence/programmable/fpga-gpu.html, Accessed: Jul. 2022. ing,” in 2021 IEEE International Parallel and Distributed Processing
[243] P. Karamanakos, E. Liegmann, T. Geyer, and R. Kennel, “Model Symposium Workshops (IPDPSW), 2021, pp. 96–103.
predictive control of power electronic systems: Methods, results, and [255] MathWorks, “HDL coder – Generate VHDL and Verilog
challenges,” IEEE Open J. Ind. Appl, vol. 1, pp. 95–114, 2020. code for FPGA and ASIC designs,” [Online]. Available:
[244] I. Bahri, L. Idkhajine, E. Monmasson, and M. E. A. Benkhelifa, https://www.mathworks.com/products/hdl-coder.html, Accessed:
“Hardware/software codesign guidelines for system on chip FPGA- Jul. 2022.
based sensorless AC drive applications,” IEEE Trans. Ind. Informat., [256] R. O. Hassan and H. Mostafa, “Implementation of deep neural networks
vol. 9, no. 4, pp. 2165–2176, Nov. 2013. on FPGA-CPU platform using Xilinx SDSOC,” Analog Integrated
[245] S. M. S. Trimberger, “Three ages of FPGAs: A retrospective on the Circuits and Signal Processing, pp. 1–10, 2020.
first thirty years of FPGA technology,” Proc. IEEE, vol. 103, no. 3, [257] Xilinx, “Zynq DPU v3.3 product guide,” [Online]. Available:
2015. https://www.xilinx.com/content/dam/xilinx/support/documents/ip doc-
[246] Embedded Computing Design, “Xilinx adaptive compute acceleration umentation/dpu/v3 3/pg338-dpu.pdf, Accessed: Jul. 2022.
platform (ACAP) moves beyond FPGA,” [Online]. Available: [258] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa,
https://embeddedcomputing.com/technology/processing/chips-and- D. Silver, and D. Wierstra, “Continuous control with deep reinforce-
socs/xilinx-moves-beyond-fpga-with-adaptive-compute-acceleration- ment learning.” in ICLR (Poster), 2016.
platform-acap, Accessed: Sept. 2022. [259] C. Hao, X. Zhang, Y. Li, S. Huang, J. Xiong, K. Rupnow, W.-M. Hwu,
[247] Xilinx, “Breakthrough performance/watt for sensor to AI to real- and D. Chen, “FPGA/DNN co-design: An efficient design methodology
time control - Versal AI edge series now shipping!” [Online]. for IoT intelligence on the edge,” in 2019 56th ACM/IEEE Design
Available: https://www.xilinx.com/about/blogs/adaptable-advantage- Automation Conference (DAC), 2019, pp. 1–6.
blog/2022/versal-ai-edge-series-now-shipping.html, Accessed: Oct. [260] Y. Umuroglu, N. J. Fraser, G. Gambardella, M. Blott, P. Leong,
2022. M. Jahre, and K. Vissers, “Finn: A framework for fast, scalable
[248] D. Gschwend, “Zynqnet: An FPGA-accelerated embedded convolu- binarized neural network inference,” in Proceedings of the 2017
tional neural network,” Master’s thesis, ETH Zürich, 2016. ACM/SIGDA international symposium on field-programmable gate
[249] Xilinx, “PYNQ: Python productivity,” [Online]. Available: arrays, 2017, pp. 65–74.
http://www.pynq.io/, Accessed: Jul. 2022. [261] E. Liegmann, T. Schindler, P. Karamanakos, A. Dietz, and R. Kennel,
[250] L. Crockett, D. Northcote, C. Ramsay, F. Robinson, and R. Stewart, Ex- “Ultrazohm—an open-source rapid control prototyping platform for
ploring Zynq MPSoC: With PYNQ and machine learning applications. power electronic systems,” in International Aegean Conference on
Strathclyde Academic Media, 2019. Electrical Machines and Power Electronics (ACEMP) & International
[251] Y. Hao and S. Quigley, “The implementation of a deep recurrent Conference on Optimization of Electrical and Electronic Equipment
neural network language model on a Xilinx FPGA,” arXiv preprint (OPTIM), 2021.
arXiv:1710.10296, 2017. [262] Severson-Group, “Severson-group/amdc-hardware: Circuit board
[252] M. Tsukada, M. Kondo, and H. Matsutani, “A neural network-based designs for controlling advanced motor systems.” [Online]. Available:
on-device learning anomaly detector for edge devices,” IEEE Trans. https://github.com/Severson-Group/AMDC-Hardware
Comput., vol. 69, no. 7, pp. 1027–1044, 2020.

26