This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
1

An Overview of Artificial Intelligence Applications

for Power Electronics
Shuai Zhao, Member, IEEE, Frede Blaabjerg, Fellow, IEEE, and Huai Wang, Senior Member, IEEE

Abstract—This paper gives an overview of the Artificial Intel- solid foundation for the AI in power electronics. AI is able
ligence (AI) applications for power electronic systems. The three to exploit data to improve product competitiveness by global
distinctive life-cycle phases, design, control, and maintenance design optimization, intelligent control, system health status
are correlated with one or more tasks to be addressed by
AI, including optimization, classification, regression, and data estimation, etc. As a result, the research in power electronics
structure exploration. The applications of four categories of AI can be conducted from a data-driven perspective, which is
are discussed, which are expert system, fuzzy logic, metaheuristic beneficial especially to complex and challenging cases.
method, and machine learning. More than 500 publications Due to the specific challenges and characteristics of power
have been reviewed to identify the common understandings, electronic systems, e.g., high tuning speed in control, high
practical implementation challenges, and research opportunities
in the application of AI for power electronics. This paper is sensitivity in condition monitoring for aging detection, etc.,
accompanied by an Excel file listing the relevant publications for the implementation of AI in power electronics has its own
statistical analytics. features that are different from other engineering areas, e.g.,
Index Terms—Artificial intelligence, design, intelligent con- image classification. Therefore, there is a pressing need for
troller, predictive maintenance, power electronic systems, prog- an overview of AI in power electronics to expedite synergy
nostics and health management research and interdisciplinary applications. Based on literature
review, in this paper the applications of AI in power electronics
are categorized into three aspects, i.e., design, control, and
I. I NTRODUCTION maintenance.
OWADAYS artificial intelligence (AI) is expanding
N rapidly and is one of the most salient research areas
during the last several decades [1], [2]. The aim of AI
is to facilitate systems with intelligence that is capable of
human-like learning and reasoning. It possesses tremendous
advantages and has been successfully applied in numerous
industrial areas including image classification, speech recog-
nition, autonomous cars, computer vision, etc. With immense
potentials, power electronics benefit from the development of
AI. There are various applications, including design optimiza-
tion of power module heatsink [3], intelligent controller for
multi-color light-emitting diode (LED) [4], maximum power
point tracking (MPPT) control for wind energy conversion
systems [5], [6], anomaly detection for inverter [7], remaining
useful life (RUL) prediction for supercapacitors [8], etc. By Fig. 1. Annual number of publications of AI in power electronics since 1990.
implementing AI, power electronic systems are embedded
with capabilities of self-awareness and self-adaptability, and Fig. 1 shows the annual number of publications related to
therefore the system autonomy can be improved. AI for power electronics since 1990. The statistical data are
Meanwhile, the rapid development of data science, includ- based on searching the IEEE Xplore from the journals IEEE
ing sensor technology, internet-of-things (IoT), edge com- Transactions on Power Electronics, IEEE Journal of Emerging
puting, digital twin [9], and big data analytics [10], [11], and Selected Topics in Power Electronics, IEEE Transactions
provides a wide variety of data for power electronic systems on Industrial Electronics, IEEE Transactions on Industrial
throughout different phases of its life-cycle. The increasing Informatics, and IEEE Transactions on Industry Applications.
volume of data enables immense opportunities and lays a The data of 2020 are up to May 2020. As a result, a total
of 444 relevant journal papers are identified, which can be
This work was supported by the Innovation Fund Denmark through the found in the supplemental Excel file. It can be seen that
project of Advanced Power Electronic Technology and Tools (APETT) and
the Villum Foundation through the project of Light-AI for Congitive Power the implementations of AI have been drastically increased
Electronics. and experienced a spectacular dynamism over the last few
Shuai Zhao, Frede Blaabjerg, and Huai Wang are with the Department years. The number of publications for control is continuously
of Energy Technology, Aalborg University, Aalborg 9220, Denmark. (Email:
[email protected]; [email protected]; [email protected]) (Corresponding author: Huai increasing and it is the most active research area. Since 2007,
Wang). there is an increase regarding the design and maintenance

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
2

applications, and such trends are more evident in the last two Moreover, it presents the quantitative information of the
years. method usage percentages and application trend.
It is found that several existing reviews in the literature 3) The advantages and limitations of AI algorithms are
are related to this topic. In [12], the metaheuristic methods comprehensively investigated. Exemplary applications
for stochastic optimization for power quality and waveform, are provided for AI in each life-cycle stage, where the
circuit design, and control tuning are reviewed. It focuses on challenges and future research directions are discussed.
the optimization tasks only. The details of neural network The rest of this paper is organized as follows. Section II
(NN) in industrial applications are presented in [13] with the presents the functions, methods, and milestones of AI in power
design of network structure, training methods, and application electronics. The applications of AI in design, control, and
considerations. It covers a broad scope of engineering appli- maintenance are discussed in Section III, Section IV, and
cations beyond power electronics. In [14], a comprehensive Section V, respectively. The outlook on the AI applications
review is given on the applications of NN in power electronics. for power electronics is put forward in Section VI. Finally,
Several specific examples of control and system identification conclusions are given in Section VII.
are detailed. Nevertheless, other AI techniques, such as fuzzy
logic, metaheuristic methods, etc., have not been discussed. II. F UNCTIONS AND M ETHODS OF A RTIFICIAL
Although these techniques are further discussed in [15], it I NTELLIGENCE FOR P OWER E LECTRONIC S YSTEMS
emphasizes on illustrative examples while an in-depth anal-
Fig. 2 gives a summary of the methods, functions, and
ysis of AI algorithms is not provided. In [16], an intensive
applications of AI for power electronics. It can be seen that AI
discussion of metaheuristic methods for MPPT in photovoltaic
has been extensively applied to the three distinctive life-cycle
(PV) systems is presented. In [17], the AI techniques applied
phases of power electronic systems, including design, control,
to PV systems are reviewed, which focuses on the specific PV
and maintenance.
applications only.
As a functional layer between AI and power electronic
Maintenance [18] in power electronics is a topic that applications, the essential functions of AI are categorized
includes reliability, condition monitoring, remaining useful as optimization, classification, regression, and data structure
life prediction, etc. Several review papers in the last decade exploration:
can be found in [19]–[22]. In [19], a state-of-the-art analysis
• Optimization: It refers to find an optimal solution max-
of the condition monitoring and fault detection in power
imizing or minimizing objective functions from a set of
electronics is presented. However, it only includes a very
available alternatives given constraints, equalities, or in-
limited AI-based fault detection methods. In [20], a review
equalities that the solutions have to satisfy. For example,
of condition monitoring techniques for capacitors in power
in the design task, optimization serves as a tool to explore
electronic converters is presented. It includes only the AI-
an optimal set of parameters that maximize or minimize
based parameter identification methods. In [21], the methods
design goals with design constraints.
in Prognostics and Health Management (PHM) of information
• Classification: It deals with assigning input information
and electronics-rich systems are summarized. This paper only
or data with a label indicating one of the k discrete
discusses the category of AI algorithms in the PHM area, while
classes. Specifically, anomaly detection and fault diag-
there is no algorithm detail or comparative analysis. In [22],
nosis in maintenance is a typical classification task to
machine learning methods applied in reliability management
determine fault labels with condition monitoring infor-
of energy systems are summarized. It focuses on the machine
mation.
learning method and the maintenance task only. A tutorial
• Regression: By identifying the relationship between input
[23] regarding “Artificial Intelligence Applications to Power
variables and target variables, the goal of regression is
Electronics” is presented on the 2019 IEEE Energy Conversion
to predict the value of one or more continuous target
Congress and Exposition. It serves as an introductory level
variables given input variables. For example, an intelli-
presentation. Nevertheless, the desirable details of the AI
gent controller can be facilitated with a regression model
algorithms and their comparisons are not available.
between the input electrical signals and the output control
As a result, it lacks a comprehensive review of the AI variables.
algorithms and applications for power electronics. From a life- • Data structure exploration: It consists of data clustering
cycle perspective, this paper aims to fill this gap and compre- that discovers groups of similar data within a dataset,
hensively review the published research in power electronics density estimation that determines the distribution of
using AI techniques, which needs a systematic consolidation. data within the input space, and data compression that
The contributions of this paper include projects high-dimensional data down to low-dimensional
1) The AI algorithms in power electronics are systemati- data for feature reduction. For example, in maintenance,
cally investigated from a life-cycle perspective, where the degradation state clustering is within the data structure
the relationships of the relevant AI algorithms, their exploration category.
essential functions, and the relevant applications are According to the surveyed 444 relevant journal papers, Fig. 3
identified. shows a Sankey diagram of application usage statistics of AI
2) A timeline map is provided to illustrate the milestones methods in the life-cycle of power electronic systems. Specifi-
of AI algorithms and power electronic applications. cally, the percentages of AI application in the design, control,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
3

Fig. 2. The application of artificial intelligence in the life-cycle of power electronic systems. Section II-A implies that the relevant discussions are presented
in part A of Section II.

A. Expert System
Expert system is the earliest method in AI that is effectively
implemented in industrial applications [17]. The expert system
[24]–[27] is essentially a database that integrates the expert
knowledge in a Boolean logic catalog, based on which the
IF-THEN rules in human brain reasoning are simulated. It
is an intelligent system simulating the inference process that
answers the why-and-how inquires based on the database. The
database is from either field expert experience or simulation
data, facts, and statements. It can be continuously updated.
The technical details of expert system are given in [17], and
several exemplary applications can be found in [15], [28].
It is worth mentioning that the applications of expert system
Fig. 3. Sankey diagram of AI methods and applications in each phase of the
are as low as 0.9% according to the usage statistics in Fig. 3.
life-cycle of power electronic systems. The statistical usages and percentages It is because the expert system is generally based on system
are based on the data in Fig. 1. principles and rules, which relates strongly to the system
of interest and lacks universality. It applies to well-defined
domains only with solid expert rules. Besides, due to the
rapid development of computational platforms, the functions
and maintenance are 9.8%, 77.8%, and 12.4%, respectively. of expert system can be replaced with other advanced AI
Regarding the functions, the percentages of optimization, methods (e.g., fuzzy logic and machine learning) with superior
classification, regression, and data structure exploration are capabilities in inference and approximation.
33.3%, 6.6%, 58.4%, and 1.7%. It shows that most of the
tasks of AI in power electronics are essentially regression and B. Fuzzy Logic
optimization. The AI methods can be generally categorized
Similar to expert system, fuzzy logic is also a rule-based
as expert system, fuzzy logic, metaheuristic methods, and
method while it extends the Boolean logic into a multi-valued
machine learning. Their application percentages are 0.9%,
case. Fuzzy logic is an ideal tool to tackle system uncertainties
21.3%, 32.0%, and 45.8%, respectively. It suggests that the
and noisy measurements [29]–[31]. Instead of using the precise
largest portion of AI in power electronics is with the machine
input crisp value directly, fuzzification is firstly performed with
learning. These methods will be detailed subsequently. Note
the fuzzy sets consisting of several membership functions to
that a comprehensive but still not exhaustive investigation
a range of 0 to 1. The fuzzy input signals are then aggregated
is conducted. Only the relevant AI methods that are widely
with fuzzy rules in the inference step. Defuzzification is
applied to power electronics are considered.
subsequently performed on the inference result by considering
the degree of fulfillment and output a crisp value. As a

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
4

result, the crisp value is manipulated in a fuzzy space that methods can be categorized as trajectory-based methods (tabu
completes nonlinear mapping between the input and output search method [43], simulated annealing method [44], etc.) and
with elaborately designed principles. population-based methods (genetic algorithm, particle swarm
In most applications, a fuzzy logic method mainly consists optimization [45], ant colony optimization, differential evo-
of four parts [30]: fuzzification, rule inference, knowledge lution [46], immune algorithm [47], etc.). For the trajectory-
base, and defuzzification. Firstly, fuzzification is performed on based methods, each exploration stage includes only one can-
the input of linguistic variables with membership functions, didate solution and it evolves into another solution according
including triangular, trapezoidal, Gaussian, bell-shaped, sin- to a certain rule. The performance of this method is mainly
gleton, and other customized shapes. Secondly, the inference based on the quality and efficiency of the rule. As a result, the
module integrates the signals together according to IF-THEN convergence speed of the trajectory-based methods is generally
fuzzy rules in the knowledge base derived from expert expe- slow and the final solution is prone to local rather than global
rience. Thirdly, defuzzification is performed on the signal for solution for non-convex optimization tasks. For the population-
output. One example of the fuzzy rule is based methods, multiple candidate solutions are randomly gen-
Antecedent: IF X is Medium AND Y is Zero, erated. At each iterative exploration, these candidate solutions
Consequent: Then Z is Positive. are diversified (e.g., crossover in the genetic algorithm) or
For both the antecedent and consequent, the degree of ful- incorporated and replaced with new candidate solutions to
fillment is determined by the membership functions. The improve the quality of the population at the present generation.
type of fuzzy inference scheme is categorized as Mamdani- As a result, the suitability of the population is iteratively
type [30], [32]–[35] and Takagi-Sugeno-Kang-type (TSK- improved to approach the optimal solution. Compared to the
type) [31], [36]–[38]. For the Mamdani-type fuzzy inference trajectory-based methods, they are superior in the convergence
scheme, the membership function of the antecedent and the speed, the global searching capability, and especially useful for
consequent are shape-based functions, e.g., triangular. For the large-scale optimization tasks. Nevertheless, the computational
TSK-type fuzzy inference scheme, the membership function burden of the population-based methods is more intensive.
of the antecedent part is identical to the Mamdani-type while This challenge needs to be considered for online application
that of the consequent is singleton at several constant values. cases where efficiency and speed are of most significance.
Typically, more fuzzy sets are needed for the Mamdani-type Table I shows a summary of the metaheuristic methods in the
scheme compared to the TSK-type scheme for the same task. area of power electronics with their advantages and limitations.
Compared to the fuzzy terms in the Mamdani-type, the mem- These metaheuristic methods are qualitatively compared in
bership function in the TSK-type scheme can be functional terms of several critical features including implementation
type as either linear or constant, which is more powerful and simplicity, global convergence, convergence speed, and par-
accurate in nonlinear approximation. More theoretical details allel capability.
of fuzzy logic are discussed in [15], [39]. Due to enormous advantages, most of the optimization tasks
Note that expert experience plays a critical role in the design in power electronics are solved with the population-based
of the membership function and the fuzzy rule, and such a methods. It can be seen from Table I that there are various
method is applicable to experts only in most cases. From this population-based methods with the improved variants for
perspective, the prior information and expert experience can optimization tasks in power electronics. They are developed
be coped with fuzzy logic and then incorporated with other and improved with different biological inspirations. In addition
AI techniques as a hybrid method. to the above widely applied metaheuristic methods, several
other emerging approaches have been applied in a limited
C. Metaheuristic Methods scale, e.g., biogeography-based optimization [72], crow search
algorithm [73], grey wolf optimization [74], firefly optimiza-
Once the optimization task of specific applications is for-
tion algorithm [16], bee algorithm [75], colonial competitive
mulated, the optimal solution can be obtained by either a
algorithm [76], teaching-learning-based optimization [77], etc.
deterministic programming method (e.g., linear or quadratic
It is worth mentioning that the selection of the best method is
programming) or a non-deterministic programming method,
not a simple task, which is application-dependent [12]. Genetic
i.e., metaheuristic method. The deterministic programming
algorithm and particle swarm optimization are the two most
methods need to calculate the gradient and Hessian matrices
popular metaheuristic methods applied to power electronics, as
[40], which is challenging for most of the optimization tasks
shown in Fig. 4. They are the fundamentals and representatives
in power electronics due to the complexity. Metaheuristic
for evolutionary algorithms and swarm intelligence algorithms,
methods serve as a general end-to-end tool that needs less
respectively, based on which various variants are developed.
expert experience and is efficient and scalable for various
Practitioners can choose the method considering its superiority
optimization tasks.
according to Table I.
The metaheuristic methods [12] are generally developed
with inspirations of biological evolution, e.g., genetic algo- Note that there is no guarantee for a global optimum for
rithm [41] by process of natural selection, ant colony opti- metaheuristic methods, but the solution is generally satisfac-
mization algorithm (ACO) [42] by simulating ants in finding tory and acceptable for most practical applications. For more
an efficient path for foods. The exploration of optimal solution theoretical details of the metaheuristic methods, readers can
is motivated by the trial-and-error process. The metaheuristic refer to [16], [78].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
5

Table I: The applications of metaheuristic methods in power electronics. Superior: +++, intermediate: ++, inferior: +

Advantages and Limitations

Type Algorithms Implementation Global Convergence Parallel Exemplary Applications
Simplicity Convergence Speed Capability
Design [45], [48], [49], Control
Particle swarm
+ +++ + Yes [50]–[56], Maintenance [57],
optimization (PSO)
[58].
Population- Design [3], [41], [59]–[63],
Genetic algorithm
based + +++ + Yes Control [64]–[68], Maintenance
(GA)
method [69]
Ant colony
++ ++ ++ Yes Design [70], Control [42]
optimization (ACO)
Differential
++ +++ ++ Yes Control [46], [71]
evolutionary (DE)
Immune algorithm
++ ++ ++ Yes Control [47]
(IA)
Trajectory- Tabu search method +++ + +++ No Design [43]
based Simulated
method +++ ++ +++ No Control [44]
annealing method

data points, i.e., the ones in the testing dataset, is termed as

the generalization. Since the training dataset comprises only
a limited amount of possible input-and-output pairs in most
cases, its generalization on new inputs is one of the most
critical performance factors of supervised learning methods.
Generally, supervised learning methods can be catego-
rized into connectionism-based methods (i.e., neural network
method), probabilistic graphical methods, and memory-based
methods (i.e., kernel method). For neural network methods,
knowledge learned from the training dataset is facilitated
Fig. 4. Usage statistics of population-based metaheuristic methods in opti- and transferred as the connection weights and structures of
mization of power electronics. The statistical results are obtained based on the network. Numerous research has been devoted to im-
the data in Fig. 1.
proving the performance of neural network methods. These
improvements are from two aspects for applications in power
electronics. The first aspect deals with enabling the uncertainty
D. Machine Learning
capability in handling the noisy signal of the neural network
Machine learning is designed to automatically discover to improve the method robustness. This feature is facilitated
principles and regularities with experience from either col- by integrating the fuzzy logic into the neural network as the
lected data or interactions by trial-and-error. For applications fuzzy neural network or its variants (e.g., adaptive neuro-
in power electronics, it is categorized as supervised learning, fuzzy inference system (ANFIS) [101]). The second aspect is
unsupervised learning, and reinforcement learning. for dynamic-performance improvement of the neural network
1) Supervised Learning: With the training dataset consist- to tackle time-series dataset cases, e.g., intelligent controller,
ing of input-and-output pairs, the supervised learning aims remaining useful life prediction. Compared to the conventional
to establish the mapping and functional relationships between neural network where the network weights are independent,
the inputs and outputs implicitly. This feature is especially the transient performance is facilitated by sharing weights
useful for cases in power electronics where system models between different layers and network cells. The weight sharing
are challenging to formulate. Generally, the tasks of the can be implemented either in a shallow scale with a convolu-
supervised learning include classification and regression. For tional structure (e.g., 1-D convolutional neural network, Time-
classification, its output of the input-and-output pairs in the delayed neural network (TDNN) [114]), or in full and deep
training dataset deals with a finite number of discrete cat- scale by using a recurrent unit as recurrent neural network
egories to be labeled. For example, the fault diagnosis for a [105]. Generally, the modeling capability of recurrent unit
multilevel inverter [94] is a typical classification task where the implementation is superior to the one with a convolutional
discrete fault label needs to be identified given the input fault structure. More theoretical details of the neural network meth-
information. For a regression task, the output of the input-and- ods are discussed in Chapter 5 of [1] and [13], [14].
output pairs consists of one or more continuous variables. An The probabilistic graphical methods obtain knowledge from
example of regression is the remaining useful life prediction the data by using a diagrammatic representation of input-
of IGBTs [114] where the output, i.e., the residual useful and-output pairs. The diagrammatic representation implies
lifetime, is a continuous variable. Once the model is trained, the conditional dependence relationship between the decision
it is ready to evaluate new data points that differ from the variables. The underlying relationship in the model is formu-
training dataset. The model capability in dealing with new lated in the Bayesian framework [1] and can be inferred in

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
6

Table II: Supervised learning methods and the applications to power electronics.

Type Method Variants Advantages and Limitations Exemplary Applications

Design [79], [80], Control [4],
Feed-forward neural
Conventional N/A (Not applicable) [51], [81]–[85], Maintenance
network (FFNN)
NN [86]–[97]
Compared to FFNN:
Radial basis function
– Simple network structure Control [50], Maintenance [98]
network (RBFN)
– Higher speed of training
Compared to conventional NN:
Fuzzy neural network – Capability of handling uncertainty
NN with Control [46], [71], [99], [100]
(FNN) – Incorporation of expert experience
fuzzy logic
Neural network (NN) method

– Higher speed of training

Adaptive neuro-fuzzy Compared to FNN:
Control [101], Maintenance
inference system – Automatic fuzzy-rule generation with less expert
[102], [103]
(ANFIS) experience
Compared to conventional NN:
Recurrent neural
– Better transient and dynamic capability Control [52], [53], [71], [99],
network (RNN) or
– Better sensitivity [104]–[107], Maintenance [108]
Elman NN (ENN)
– Slow speed in training
Nonlinear Compared to RNN:
autoregressive – Higher speed of training
Design [109], Maintenance
NN with network with – Better generalization capability
[110]
recurrent unit exogenous inputs – Better capability in dealing with long-term
(NARX) dependence
Compared to RNN:
Echo state network
– Only hidden-to-output weights need to determine Maintenance [111], [112]
(ESN)
– less severe issue of gradient explode and vanish
Compared to RNN:
Long short-term
– Most effective structure for practical application Maintenance [113]
memory (LSTM)
– No issue of gradient exploding and vanishing
Time-delayed neural Compared to conventional NN:
NN with network (TDNN) – Partial weight sharing for transient performance
Control [14], Maintenance
convolutional (also termed as 1-D improvement for specific applications
[114], [115]
structure convolutional NN Compared to RNN:
(CNN)) – Capability of time-series modeling is weaker
Probabilistic

Compared to NN methods:
graphical
method

– Better interpretability
N/A Bayesian Networks Maintenance [96], [116]–[118]
– Computationally intensive
– Probabilistic output with uncertainty quantification
Compared to NN methods:
Conventional
Gaussian processes – Probabilistic output with an uncertainty Maintenance [119], [120]
kernel method
quantification
Kernel method

Compared to conventional kernel method: Control [121], [122],

Support vector
– Better approximation capability with small dataset Maintenance [7], [74], [96],
machine (SVM)
– Better computational efficiency [118], [123]–[126]
Sparse kernel Compared to SVM:
method – Much sparser than SVM while maintaining
Relevance vector
comparable generalization capability Maintenance [127], [128]
machine (RVM)
– Probabilistic output with uncertainty quantification
– Training time is generally longer than SVM

a probabilistic way. Thus, the interpretability of the model is methods (e.g., Gaussian processes) are computationally inten-
much better compared to neural network methods. Besides, sive due to the whole training dataset is applied to the testing
the probabilistic graphical model is superior in dealing with stage. To avoid the excessive computational burden, sparse
uncertainty and incomplete knowledge. One of the typical solutions are proposed as support vector machine (SVM)
probabilistic graphical methods is the Bayesian network [117]. and relevance vector machine (RVM), where the parameter
More theoretical details of the probabilistic graphical methods estimation is improved based on Bayesian methods. With the
are given in Chapter 8 of [1]. sparse solution, only a subset of the training dataset is applied
For the neural network methods and the graphical methods, to the testing stage and thus it is more efficient compared to
the training dataset is discarded when the training is com- the conventional kernel methods. More theoretical details of
pleted. While the training dataset in kernel methods is kept the kernel methods are discussed in Chapters 6 and 7 of [1].
and used in the testing stage, and the learned knowledge is Generally, the requirement of the training dataset for the kernel
facilitated as the identification of critical data points (e.g., methods is lower than the neural network methods. Therefore,
support vectors in support vector machine [126]) or subset the kernel methods are more suitable for the cases with a small
in the training dataset. One typical kernel method is Gaussian dataset. While due to the training dataset is needed in the
processes, which has been applied to the remaining useful life testing stage, the memory requirement of the kernel methods
prediction of IGBTs in [119]. Note that the conventional kernel

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
7

Table III: Unsupervised learning methods and the applications to power electronics.

Function Method Advantages and Limitations Exemplary Applications

– Simple implementation Control [129], Maintenance
k-means
– Sensitive to outliers [118], [130]–[132]
Clustering Compared to k-means
Self-organizing – Better interpretability Maintenance [133], Control
maps (SOMs) – Less sensitive to initial parameter selection [99]
– Less sensitive to outliers
Principal
– Flexible framework with various improvements, Control [134], Maintenance
Data compression component analysis
e.g., kernel PCA, Bayesian PCA, etc [117], [127], [132], [135], [136]
(PCA)

is higher than the neural network methods. The involvement

of the training dataset also limits the speed performance at the
testing stage. It should be considered for online applications
where the execution time is critical, e.g., control application.
As a result, Table II shows a summary of the supervised
learning methods and their variants in power electronics, in
terms of the advantages, limitations, and exemplary applica-
tions.
2) Unsupervised Learning: Compared to the supervised
learning where the dataset is input-and-output pairs, unsu-
pervised learning has no output data for the learning target
during the learning process. Generally, the tasks of unsuper-
vised learning in applications of power electronics can be Fig. 5. Usage statistics of machine learning methods in power electronic
categorized as data clustering and data compression. systems. The statistical results are obtained based on the data in Fig. 1.
For the data clustering, it explores the regularities from
the smeared dataset and partitions the dataset into several
different groups or clusters according to their similarities. in terms of an action selection policy, which can maximize
In this way, the data characteristics within the same cluster the total expected rewards over the future. The Q-table is an
are similar to each other and different from the ones in informative policy matrix that records the optimal action to
other clusters. One typical data clustering application is the be taken given the particular condition variables. More theo-
identification of the discrete health state from the continuous retical details of RL can be found in [139]. One application
degradation data [131] in the condition monitoring of power example is the MPPT [5], [6], [140]. Note that RL obtains the
electronic converters. The purpose of the data compression is experience from the interactions between systems instead of
to eliminate excessive information in the dataset to reduce the existing datasets. It is thus more favorable for the cases where
number of features of the dataset. For example, using principal the system is with less knowledge or its model is challenging
component analysis (PCA) [127], a reduced representation of to formulate.
the dataset is obtained with a much fewer number of features, As a summary, Fig. 5 presents the usage statistics of the
which yet maintain the integrity of the dataset. machine learning methods. Supervised learning is dominantly
Generally, these unsupervised learning algorithms serve applied to power electronics. The reason is that the supervised
as the data-preprocessing before it goes to the subsequent learning is a versatile tool, which is typically the central part of
analytics (e.g., fault diagnosis). Although this step is optional, the majority of machine learning-related applications in power
it is beneficial to reducing the computational burden and electronic systems.
improving the analytics accuracy. Table III gives a summary
of typical unsupervised learning methods for power electronic E. Timeline of Relevant AI methods and Applications in Power
applications. More unsupervised learning methods and theo- Electronics
retical details can be found in [137]. Fig. 6 summarizes the milestones of the relevant AI methods
3) Reinforcement Learning : In contrast to the supervised and their applications in power electronics. It includes the year
learning and the unsupervised learning, reinforcement learning when the algorithm is first proposed, the first application in
(RL) does not require a training dataset. Instead, it aims to power electronics, the milestones of relevant AI algorithms,
find a suitable action strategy that maximizing the reward for and applications in terms of each method. It should be noted
a specific task, which is essentially a dynamic programming that the information is to the best knowledge of the authors.
or optimization task. This goal-oriented strategy is formulated Also, the timeline is not extensive to include all of the existing
from interactions with systems or simulation models by a trial- AI algorithms. Instead, only the ones that show great potentials
and-error process [138]. In this way, it accumulates experience in power electronics are included. According to Fig. 6, it can
progressively and learns a specific strategy that maximizes be noted that:
the predefined goal. Theoretically, RL is a Markov decision 1) The application of both expert system and fuzzy logic
process [139]. The training of RL aims to develop a Q-table is moderate nowadays, especially for the expert system.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
8

Expert system (ES): C5. (Real-coded GA)-Maintenance-2018, [69] E4. Control-2004, [82] G1. (SVM)-Maintenance-2008, [126]
A0. Design-1988, [24] C6. Design-2019, [3] E5. (RBFN)-Control-2008, [50] G2. (RVM)-Maintenance-2013, [128]
A1. Design-1995, [25] C7. (Multi-objective)-Design-2020, [63] E6. Maintenance-2009, [93] G3. (SVM)-Maintenance-2013, [96]
A2. Design-2008, [26] Particle swarm optimization (PSO): E7. (RNN)-Maintenance-2010, [108] G4. (SVM)-Maintenance-2017, [74]
A3. Maintenance-2008, [28] D0. Control-2003, [141] E8. (RFNN)-Control-2010,[99] G5. (GPR)-Maintenance-2018, [119]
Fuzzy logic (FL): D1. Control-2008, [50] E9. (FNN)-Control-2015, [46] G6. (SVM)-Maintenance-2019,[125]
B0. Control-1990, [29] D2. Design-2008, [45] E10. (TDNN)-Maintenance-2016, [114] k-means:
B1. (Mamdani)-Control-1997, [30] D3. Control-2010, [56] E11. Maintenance-2017 [86] H0. Control-2007, [129]
B2. (TSK)-Mainenance-2008, [38] D4. Control-2012, [55] E12. (ANFIS)-Maintenance-2018, [103] H1. Maintenance-2008, [130]
B3. (Mamdani)-Control-2011, [35] D5. Design-2014, [48] E13. Design-2018, [80] H2. Maintenance-2019, [131]
B4. (TSK)-Control-2013, [37] D6. Maintenance-2016, [57] E14. (ESN)-Maintenance-2019, [112] Principal component analysis (PCA)
B5. (Mamdani)-Maintenance-2017, [33] D7. Control-2017, [58] E15. (CNN)-Maintenance-2019, [115] I0. Control-1995, [134]
Genetic algorithm (GA): D8. Design-2019, [49] Bayesian network (BN): I1. Maintenance-2015, [127]
C0. Control-1995, [65] Neural network (NN) method: F0. Maintenance-2008 [116] I2. Maintenance-2020, [136]
C1. Design-2001, [41] E0. Control-1989, [142] F1. Maintenance -2013 [96], 2011 Reinforcement learning (RL)
C2. Maintenance-2007 [143] E1. (RNN)-Control-1999, [105] F2. Maintenance-2017 [117], 2017 J0. Control-2008, [144]
C3. Design-2008, [61] E2. (ANFIS)-Control-2000, [101] Kernel method (KM): J1. Control-2016, [6]
C4. (Mixed-integer)-Control-2009, [68] E3. Control-2002, [83] G0. (SVM)-Control-2007, [121] J2. Control-2017, [140]
Fig. 6. Timeline of relevant AI methods and applications in power electronics. The milestones are identified considering the significant algorithm variants and
the relevant applications. It is organized as the form of (significant variants)-application-year. Significant variant is specifically indicated. Otherwise, it is a
standard algorithm.

Before the 2000s, their practical implementations are tronic systems. Secondly, the structure of neural network
developed in the presence of the limited performance is quite flexible to incorporate other AI methods for
of computing hardware, which has been significantly performance improvement, implying numerous method
improved to date. This rapid development of computing variants.
hardware facilitates and accelerates the implementation 4) There is an increasing trend of applications with kernel
of other more powerful AI methods for replacing expert methods and probabilistic graphical models. It is be-
system and fuzzy logic. cause most of these methods are formulated within the
2) Metaheuristic methods are continuously evolving and Bayesian framework, which possesses better generaliza-
applied to power electronics. They are used for a com- tion and interpretability. Moreover, their computational
plete task or a key step jointly with other machine burden can be well tackled with the platforms to date.
learning methods. 5) RL is the latest frontier of the machine learning methods
3) Neural network methods are the most active area for applied to power electronics, facilitated by the rapid
AI applications for power electronics. The reason is development of computing hardware.
twofold. Firstly, the significant development of comput-
It can be noted from Figs. 2, 3, and 6 about the comparisons
ing hardware unleashes the potentials of neural network
for different AI methods:
methods in dealing with complex tasks in power elec-
1) Both metaheuristic methods and machine learning can

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
9

be applied to optimization tasks. Specifically, machine 2) Constraint space: The constraint space defines feasible
learning-based optimization (i.e., reinforcement learn- space, boundary, relationship, and limitation that the ob-
ing) focuses on the dynamic optimization involved with jective function is subjected to. These constraints include
the decision-making (e.g., MPPT). Metaheuristic method either linear or nonlinear equalities and inequalities.
is generally applied to the static optimization (e.g., They are derived from the practical design requirements,
heatsink design). e.g., geometry, volume, lifetime characteristics, cost, etc.
2) Both fuzzy logic and machine learning can be exploited 3) Solution exploration: The defined optimization problem
for classification tasks. Generally, machine learning is is to maximize (or minimize) objective functions by
more accurate and flexible than fuzzy logic. adjusting the decision variables in the constraint spaces.
3) The regression task can be implemented with expert AI methods, especially the metaheuristic methods, can
system, fuzzy logic, and machine learning. The imple- be applied to this step.
mentation of expert system is simple but less powerful 4) Performance evaluation: The candidate solution can be
compared to fuzzy logic and machine learning. The tested against the predefined objectives by using simula-
implementation of fuzzy logic needs expert experience. tion, hardwire-in-the-loop testing, prototype experiment,
Machine learning is the most popular method and var- etc. The results can be returned to previous steps for
ious algorithm variants have been developed. It can be further performance improvement and optimization.
incorporated with fuzzy logic for performance improve- Instead of a sequential procedure, the design task is an
ment. iterative trial-and-error process. Based on the evaluation at
4) Only machine learning can be applied to the task of data each step, the task may be reformulated, e.g., adjusting the
structure exploration. objectives, modifying the constraint space, reconfiguring the
The following three sections discuss the applications of the programming methods, etc. For conventional design in power
above introduced AI methods in the design, control, and electronics, it is time-consuming and needs multiple iterative
maintenance phases of power electronic systems, respectively. steps. For example, the component alignment and the model
selection rely on expert experience and intuition without ample
III. D ESIGN quantitative reference. In this way, the design performance will
Design in power electronics encompassing topology selec- converge slowly to the required standards. This drawback can
tion, component sizing, circuit synthesis, reliability consid- be mitigated by AI methods. They can be applied to Step 1)
erations, etc., is essentially an optimization task [145]. A objective formulation for the design time reduction, and Step
typical procedure for the design of power electronic systems 3) solution exploration for the modeling and optimization.
comprises four steps:
1) Objective formulation: Objective functions are desirable A. Design Time Reduction
design goals to be maximized or minimized. Generally, The formulation of design objective needs to be improved if
the design goals in power electronics include component its evaluation is computationally intensive. One application of
parameter [41], weight [146], volume [147], cost [146], AI methods is a surrogate model in the objective formulation to
heatsink pattern [3], area [148], power loss [62], etc. reduce the computational effort. The surrogate model yields an
It is crucial for formulating the required or desired identical behavior to the system dynamics that are challenging
design requirements to several explicit mathematical to formulate or need intensive computational efforts to char-
expressions as a single objective as given in (1) or acterize. In the iterative design process, AI-based surrogate
multiple objectives as given in (2) [12], [145]: model serves as a replacement that significantly reduces the
computational effort.
max f (x), (1) As an application of Design for Reliability (DfR), in [80],
x two feed-forward neural networks are applied to the automated
max wT f (x), max f (x), (2) reliability design of power electronic systems. The first feed-
x x
forward neural network serves as a surrogate model emulating
s.t. g(x) ≤ 0, h(x) = 0, x ∈ [xl , xu ].
thermal characteristics of power converters, by which the de-
where g(x) and h(x) are inequalities and equalities, sign parameters can be mapped to the information of junction
respectively. xl , xu are the lower and the upper bound- temperature variations. Subsequently, the second feed-forward
aries for decision variables x, respectively. Here the neural network is applied to map the annual mission profiles
maximization is the goal, which can simply be applied to (e.g., annual solar irradiation and ambient temperature) to
the minimization case. Note that for multiple objectives the annual lifetime consumption. In this way, the nonlinear
in (2), it can be either solved by maximizing a scalar relationship between the designed parameters and the annual
function wT f (x) by weighting multiple objectives to- lifetime consumption is quantitatively characterized, which can
gether or by optimizing objective vector f (x) directly, accelerate the iterative design process.
where Pareto front [62] can be applied to determine the Another example of AI for DfR of power electronic systems
optimal solution, e.g., the non-dominated sorting genetic is given in [109]. With superior capability in tackling time-
algorithm method for multiobjective design optimization series data, a nonlinear autoregressive network with exogenous
of power modules in [60]. inputs (NARX) is applied to the thermal modeling of power

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
10

voltage, and number of panels per string. The GA combining

with the Pareto front is applied to solve the multiobjective
optimization task. Besides, there is a specifically improved
variant of GA for the multiobjective optimization task, i.e.,
non-dominated sorting genetic algorithm II (NSGA-II) [63].
In [45], the particle swarm optimization (PSO) is applied
to the circuit synthesis of a power electronic circuit, where
Fig. 7. Nine different cell patterns for each blank cell [3]. A genetic algorithm
is applied to determine the optimal combination of different cell patterns for the optimal values of components are explored to fulfill the
the heatsink design for minimizing the junction temperature. design goals of better static and dynamic performance. For
this specific case, the simulation indicates that the PSO yields
a superior solution with less computational effort compared to
electronic systems considering the thermal cross-coupling ef- GA.
fects. The proposed NARX-based thermal model can be com- In [70], the ant colony optimization (ACO) is applied to
pleted within around 109 s, which is a significant efficiency determine the optimal component values in a power electronic
improvement compared to the 1005 s of the conventional circuit, where the conventional ACO is extended to facilitate
model. The error between the temperature estimated by the the optimization with continuous component values and ac-
NARX-based thermal model and the actual measurement is celerate the optimization process. Moreover, the component
less than 1◦ C. Experimental results indicate that the NARX- tolerance is incorporated into the optimization, which makes
based thermal model can replace the conventional model with the proposed method more beneficial to practical applications.
less testing efforts and much less computational burden.
In [79], considering the electrothermal interactions, a feed-
forward neural network is applied to construct the component IV. C ONTROL
behavior model of MOSFETs without any in-depth knowledge Essentially, control applications with AI methods in power
of the device structure. Under the static state, the complicated electronic systems can be categorized as the optimization and
nonlinear and temperature-dependent characteristics between the regression. Similar to the optimization in the design phase,
the variables, including drain-to-source voltage VDS , gate-to- the optimization-related tasks in control applications are also
source voltage VGS , junction temperature Tj , and the output dealing with metaheuristic methods. Several representative
current ID are established by using the neural network. This applications are given below.
compact model can drastically accelerate the design simulation In [64], a GA is applied to the PID tuning of a programming
process with a comparable accuracy. logic controller, where the optimization goal is to minimize
the error between the ideal step and ramp responses and
B. Modeling and Optimization the ones initialized with proportional term Kp , integral term
The modeling and optimization of power electronic systems KI , and derivative term KD found by GA. Experimental
is about specifying circuit topology, component model, com- analysis indicates that the output performance of the optimized
ponent parameter, etc, such that system dimension, weight, controller is very close to the ideal step and ramp responses.
operating frequency, etc., will result as optimal characteristics In [42], to overcome the challenges of multiple maximum
(e.g., power loss, power density) given design constraints [12]. power points in partially shaded situations for PV systems, an
Specifically, the optimization method is applied to the solution ACO-based MPPT method is proposed. It is compared with
exploration to provide an overall optimal configuration, where conventional methods, including constant voltage tracking,
metaheuristic methods in AI can be exploited. As mentioned, perturb & observe, and PSO. The experimental results indicate
the selection of a suitable metaheuristic method depends on the that the ACO-based MPPT method is superior in global
specific application. Several exemplary applications are given convergence and robustness to various shading patterns.
as follows. In [47], in a single-phase full bridge inverter, an IA is
In [3], genetic algorithm (GA) is combined with finite applied to find the optimal sinusoidal pulse-width modulation
element analysis for the automated heatsink design of a 50 (PWM) control sequences of four switches minimizing the
kW three-phase inverter. As shown in Fig. 7, GA is applied total harmonic distortion (THD) of the output waveforms. The
to optimize the combination of nine customized patterns to experiment indicates that the THD by using IA is 0.79%,
formulate a complex cell pattern of heatsink. The goal is to which is superior to that of the conventional control method
minimize the junction temperature of power semiconductor de- of hysteresis current PWM with 1.23% and the GA solution
vices. Compared to the conventional design with a regular cell with 0.99%. Moreover, the IA is superior to the GA in
pattern, the proposed method formulates a heatsink solution convergence speed. More examples of optimization-related
with 27% less in size and 6% lower in junction temperature. control applications can be found in [12].
In [62], the design of a 500 kW solar power-based micro- The regression-related tasks in control applications are deal-
grid system is formulated as a multiobjective optimization ing with the nonlinear mapping of system inputs and outputs
task, which maximizes the average power distribution and in a static or dynamic way. Specifically, it is concerned with
minimizes the system weight simultaneously. It explores the regulating systems to ensure intended performance output with
optimal values of four microgrid parameters, including battery system principles. Several limitations of conventional methods
voltage, PV maximum power, PV maximum power point are identified:

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
11

Fig. 8. Fuzzy logic-based controller for a variable-speed wind generation system [30]. MFs: Membership functions. In the rule matrix table, P: positive, V:
very, B: big, M: medium, ZE: zero, N: negative.

1) The controller configuration requires in-depth knowl- system stability. The experimental analysis demonstrates that
edge of system control principles, which are challeng- the developed adaptive TSK-type controller outperforms the
ing and even infeasible for complex cases. It is time- conventional fuzzy logic controllers and the PI controller. A
consuming for complex systems to consider the time- similar TSK-type controller can be found in [31] for approx-
varying and piecewise-linear characteristics, where the imating the typical sliding mode control curve for integrated
controller is generally optimized at several critical op- LED drivers. It is computationally efficient and implemented
erational points rather than the full operational area, on a low-cost platform.
resulting in a sub-optimal solution. Although the fuzzy logic controller can handle the system
2) Once the controller is installed, it operates in a static uncertainty, similar to conventional methods such as PID, there
way with limited adaptability, suggesting that it is only is no internal updating mechanism and thus the adaptability is
applicable to time-invariant systems. Nevertheless, when limited [50]. Also, it can be seen that the design of membership
environmental and operational conditions change, the functions and fuzzy rules require expert experience, which
controller will be less robust to system parameter shifts highly limits the method practicality. Thus, such a method is
and the control performance is likely to deteriorate. applicable to experts only in most cases. Nevertheless, from
3) From the efficient control perspective, an ideal controller this perspective, the expert experience can be coped with fuzzy
must be able to cope with parameter tolerances with logic and then incorporated with other AI techniques as a
a fast transient response to maintain system stability. hybrid method, as discussed later.
However, such a desired feature cannot be well fulfilled.
These limitations can be mitigated with AI methods. For the B. Neural Network-based Controller
regression-related task in control applications, it is organized As a black-box technique, neural network can approximate
in terms of fuzzy logic, neural network, and reinforcement a wide range of nonlinear functions to arbitrary accuracy.
learning, respectively. With few requirements on system knowledge, the NN-based
controller possesses several advantages such as robustness,
model-free, dynamic, adaptive, universal approximation, etc.
A. Fuzzy Logic-based Controller
1) Conventional Neural Network: The most widely used
Fuzzy logic-based methods have been widely applied to the neural network in power electronics is the feed-forward neural
control of power electronic systems, e.g., speed control [30], network (FFNN) (or backpropagation neural network) with a
MPPT [35], energy management [149], to name a few. feed-forward multilayer and a backpropagation topology [14].
In [30], a control strategy with three fuzzy logic controllers The respective applications essentially exploit the property of
is developed for a variable speed wind generation system. The static nonlinear mapping of the FFNN.
structure of the generator speed programming controller is In [82], an FFNN is applied to the waveform processing and
given in Fig. 8. The control variables include the increment of delayless filtering. With two cases of variable frequency and
the output power ∆Po and the last variation of speed L∆wr∗ . variable magnitude, it indicates that the FFNN can convert m-
The controller outputs the variation of speed ∆wr∗ to adjust phase waveform with an arbitrary shape into the n-phase wave-
the generator speed for a maximum wind power output. The form with various characteristics of magnitude and frequency.
Mamdani-type fuzzy logic is applied and the information is The FFNN-based waveform processing method provides a
aggregated according to the rule matrix table, e.g., “IF ∆Po is simplification of the hardware implementation. Moreover, ad-
PS AND L∆wr∗ is ZE, THEN ∆wr∗ is PM”. The membership ditional single processing functions can be embedded easily
functions are iteratively tuned by the system simulation and due to the structure flexibility.
experiment. Similar Mamdani-type fuzzy logic controller for In [83], the space vector PWM (SVPWM) for a three-
the primary frequency regulation of a wind farm can be found level voltage-fed inverter is implemented with an FFNN. The
in [34]. input of the neural network is the sampled command phase
In [36], a fuzzy logic controller is proposed for regulating voltages and the output is the pulse width patterns of SVPWM.
the speed of a switched reluctance motor based on TSK fuzzy The training data are generated by the simulation with an
logic by approximating an ideal control law. The parameter SVPWM algorithm. By comparing with a conventional DSP-
is tuned by using the Lyapunov stability theorem to ensure based SVPWM solution, the performance of the FFNN-based

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
12

Fig. 9. Structure of a radial basis function network (RBFN) with three layers
[50]. x1i is the input of the input layer node i and yi1 is its output. yj2 is the (a) Block diagram of the FNN-based controller for a boost converter.
output of the hidden layer node j. yk3 is the output of the output layer node
k. The input layer and the hidden layer are fully and directly connected with
no weights.

SVPWM is verified and it can be flexibly implemented on a

dedicated IC chip.
In addition to FFNN, another conventional NN structure is
radial basis function network (RBFN). In FFNN, the weights
of input-to-hidden and hidden-to-output are simultaneously
determined. For RBFN, the input layer is directly and fully
connected to the hidden layer without weights. The hidden
layer is connected to the output layer by weights Wj , which are
the only weight parameters to be determined in the training,
as shown in Fig. 9. Typically, the generalization of RBFN
is better than FFNN and the training speed and the execution
speed are faster. An exemplary application of RBFN in a three-
phase induction generator to regulate the DC-link voltage and (b) An FNN with a four-layer structure.
the AC line voltage can be found in [50]. Fig. 10. A fuzzy neural network (FNN)-based controller for a boost converter
Regarding the number of neurons, there are few principles [100]. x1 is the sliding surface S(x) and x2 is its differentiation, n = 2. µji
is the j th membership function for input xi . w is the weight between layers.
to determine the optimal number. A generic method is to
start with a relatively small number of neurons and then
gradually increase it according to the training error. For the
activation function in the hidden layer, there are various ena. The block diagram of the controller is given in Fig. 10(a)
options, including sigmoid [4], [51], [52], [83], radial basis and the FNN structure with four-layer is given in Fig. 10(b).
function [50], [150], hyperbolic tangent function [106], [151], The inputs of the FNN include the sliding surface S(t) and its
wavelet [46], [53], [84], [152], etc. It is worth mentioning differentiation Ṡ(t), which are obtained based on tracking the
that the wavelet activation function possesses the superior errors of the average output voltage ev and inductor current ei ,
capabilities of convergence speed and generalization. given the reference voltage Vref and current iref . The output
2) Neural Network with Fuzzy Logic: In control applica- control signal is the duty cycle u of PWM.QThe fuzzy inference
n k j
tions, parameter uncertainty and external disturbance should is implemented by the rule layer as lk = i=1 wji µi (xi ). The
PNy
be well considered for system stability and robustness. As a network output is obtained as u = f ( k=1 wk lk ). For the
result, an improved variant of NN, i.e., fuzzy neural network voltage control, the voltage tracking performance is evaluated
(FNN), or neuro-fuzzy system, which is a hybridization of NN by the mean-square error (MSE) of the output voltage:
and fuzzy logic, is proposed. FNN has the merits from both
T
aspects [100], i.e., the human-like IF-THEN reasoning rules of 1X 2
MSE = ev (d), (3)
fuzzy logic that incorporates expert knowledge and cognitive T
d=1
uncertainty, and the strong capabilities of approximation and
generalization to any nonlinear systems by the neural network. where T is the number of sampling instants. The network tun-
More theoretical details of FNN can be found in [39]. ing aims to reduce the MSE as much as possible to output an
In [100], an FNN is applied to simulate the sliding-mode accurate and stable voltage. The performance of the FNN can
control of a boost converter to alleviate the chattering phenom- be significantly improved if the membership function is well

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
13

Fig. 11. An adaptive neuro-fuzzy inference system (ANFIS)-based controller

for a PWM-inverter-fed induction motor drive [101]. It is a five-layer network
structure with the capability of automatic identification of fuzzy rules.

designed. For example, in [46], an asymmetric membership Fig. 12. A recurrent fuzzy neural network (RFNN) controller for the high-
function (AMF) is applied to the controller of a six-phase precision trajectory tracking control of a linear microstepping motor driver
[99]. A memory unit of time-delayed feedback connection Z −1 is added to
permanent magnet synchronous motor. It indicates that the enable the dynamic capability of neural network controller.
learning speed can be improved and the network structure can
be simplified compared to conventional membership functions,
e.g., Gaussian function [71], [99], [100]. characterization. The dynamic performance of the controller
One of the challenges of FNN is the design of the fuzzy rule, is critical for the transient response. To enable the dynamic
where extensive expert experience is usually needed [100]. capability of the neural network controller, a memory unit of
To overcome this challenge, another typical and effective time-delayed feedback connection Z −1 is usually inserted to
framework incorporating fuzzy logic and neural network is an formulate recurrent neural network (RNN) [107], as shown in
adaptive neuro-fuzzy inference system (ANFIS), which can be Fig. 12. The outputs of the network not only depend on the
extended from the four-layer structure in Fig. 10 as a five- present inputs but also on the previous ones. As a result, the
layer topology [101], as shown in Fig. 11. In the ANFIS, network structure can tackle the time series data to facilitate
the IF-THEN fuzzy rules, which require the involvement of the better performance of dynamics and sensitivity.
experts, can be generated automatically in the training. For In [106], a robust controller based on RNN is proposed
example, in [101], a direct-torque neuro-fuzzy control scheme for single-phase grid-connected converters for better control
is developed for a PWM-inverter-fed induction motor drive performance in the presence of system parameter changes.
based on an ANFIS. As shown in Fig. 11, the inputs of the The training of the RNN is completed by the Levenberg-
ANFIS-based controller include the flux error εm and the Marquardt (LM) method [13], [82], [106]. The harmonics
torque error εΨ . Layer 1 is the membership layer with the can be significantly reduced by using the proposed RNN-
input weights wm and wΨ . Layer 2 chooses the minimum based controller, and the requirements of the high sampling
from the inputs. Normalization is performed in layer 3. In and switching frequency and the damping policies for the
layer 4, the outputs oi is linearly combined with the network conventional control methods can be mitigated. A similar
inputs ud = (εm , εΨ ). Layer 5 is the network outputs of RNN structure, which is also termed as Elman neural network
the stator voltage command vectors in polar coordinates Vc (ENN), can be found in [52].
and ϕVc . ∆γi is the increment angle and γs is the actual In addition to the performance of dynamics, fuzzy logic is
angle of the stator flux vector. In contrast to the conventional also incorporated into RNN in order to improve the perfor-
training schemes, the parameter tuning of the ANFIS is mance of robustness. For example, in [99], a controller based
completed interactively with the backpropagation algorithms on a TSK-type self-organizing recurrent fuzzy neural network
(for membership functions) and the least square method (for (RFNN) is proposed for a high-precision trajectory tracking
parameters in 4th layer). More theoretical details of the training control of a linear microstepping motor driver. The network
methods of the ANFIS can be found in [153]. structure is given in Fig. 12. The TSK-type self-organizing
3) Neural Network with Recurrent Units: The NN struc- RFNN is applied to model the inverse dynamics of the driver.
tures in Section IV-B1 and FNN in IV-B2, however, are only Compared to the FNN in Fig. 10(b), the key of the RFNN is
applicable to the static relationship mapping and behavior the insertion of a recurrent layer, where the delayed neuron

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
14

output hi (k) is returned as the neuron input to facilitate the

network dynamics. The network diagram and size are adjusted
by the self-organizing method, and the respective network
parameters are tuned with the method of recursive least square.
As a result, the network diagram and its parameters can be
optimized simultaneously.
4) Training Methods of Neural Network: Essentially, the
training of the neural network is an optimization task. Of
course, it can be completed with conventional optimization
methods, e.g., PSO [51], recursive least square [99], Kalman Fig. 13. Framework of reinforcement learning in the maximum power point
filter [105], etc. Considering a large number of parameters in tracking (MPPT) controller of wind energy conversion systems [5], [138].
the neural network, these conventional optimization methods A Q-table is formulated to save the optimal generator rotor speed wr∗ to be
performed given the current system state st , including the current electrical
are generally inefficient. As a result, an elaborate training output power Pe and the generator rotor speed wr .
scheme is developed, i.e., backpropagation algorithm [4], [50],
[52], [53], [71], [83], [84], [150]. More theoretical details of
the backpropagation algorithm can be found in Chapter 5 of learning process is incrementally completed. This feature is
[1]. especially useful for the case where the training data can only
The backpropagation algorithm is based on the idea of be sequentially obtained in field application. The intelligent
steepest gradient descent. One of the key steps in the back- controller [53] is a typical case of a sequential training scheme
propagation algorithm is the iteration of the weight update: since the input data of the neural network can only be available
sequentially by interacting with the output of the control
wk+1 = wk − ηk gk , (4) command and the system. With this adaptive capability, the
neural network can be re-parameterized and reconfigured for
tracking the system parameter shifts. One of the key steps for
where wk is the current weight, gk is the current gradient, the sequential learning is determining a suitable learning rate
ηk is the learning rate, and wk+1 is the weight of the next ηk in (4), since a larger ηk will result in system instability
iteration. To calculate the gradient gk and find the steepest and a smaller ηk will lead to slow convergence. The optimal
direction of gradient descent efficiently, various improved learning rate ηk can be determined by using the metaheuristic
variants of the backpropagation algorithm have been proposed, methods in the training, e.g., PSO in [50], [52], [53] and
e.g., Levenberg-Marquardt method [13], [82], [106], resilient differential evolutionary in [46]. As a result, the sequential
backpropagation algorithm, conjugate gradient algorithm, one- learning process can be stable and converges fast.
step secant algorithm, etc. Note that it is challenging to deter-
mine the most suitable training algorithm for a specific task.
It depends on multiple factors, including problem complexity, C. Reinforcement Learning-based Controller
dataset size, number of parameters, task types of classifica- With reinforcement learning, the controller learns a goal-
tion or regression, etc. A useful reference can be found in oriented control strategy by interacting with the physical
Matlab Manual of Neural Network Toolbox [40], where the system or its simulation model [138]. It accumulates expe-
theoretical details, advantages, limitations, and comparisons of rience progressively and learns a specific control strategy that
these training algorithms are thoroughly analyzed with several maximizes predefined goals.
benchmark examples. It is worth mentioning that Levenberg- One of the relevant applications of RL-based controller is
Marquardt method is one of the most widely used methods for the MPPT in renewable energy systems [5], as shown in Fig.
the applications in power electronics with a fast convergence 13. Specifically, a real-time intelligent MPPT algorithm using
speed and a high accuracy. RL is proposed for a wind energy conversion system. With
Considering whether the training dataset is available in a the online learning capability of RL by interacting with the
batch form or in a sequential form, the training scheme of environment, an optimum control strategy is formulated in the
the neural network can be completed in either batch learning, Q-table. The Q-table consists of elements of state transition
which is also termed as offline learning, or sequential learn- probability q(st , at ), which can facilitate the maximized power
ing, which is also termed as online learning or incremental output (or reward) if action at , i.e., the expected generator
learning. rotor speed wr∗ , is performed given the current system state
For batch learning, the gradient gk in (4) is calculated based st , including the current electrical output power Pe and the
on all the data points in the dataset for the parameter updates. generator rotor speed wr . As a highlight, the wind turbine
It generally applies to the case where the whole dataset parameter and the wind speed are not required. This work is
is available before the neural network is implemented for further extended by integrating an NN into the Q-learning of
field application, e.g., the waveform processing and delayless RL [6]. In this way, the challenges in the determination of the
filtering in [82]. state space are avoided. The online learning process can be
For sequential learning, the gradient gk in (4) is calculated reactivated once the learned optimal relationship is destructed
based on every newly available data point or several newly by the system aging behaviors. It significantly improves the
available data points forming a mini-batch. Therefore, the autonomous capability of the wind energy conversion system.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
15

Table IV: The advantages and limitations of AI algorithms in control applica-

tions. FFNN – Feed-forward Neural network and its variants, FNN – Fuzzy 1) Offline training and knowledge learning: It integrates
neural network and its variants, RNN – Recurrent neural network and its various aspects of knowledge including historical mon-
variants, RFNN – Recurrent fuzzy neural network and its variants. Superior: itoring data, simulation data, accelerated aging test ex-
+++, intermediate: ++, inferior: +
periment, failure mode and effects analysis (FMEA),
Fuzzy Machine learning Reinforcement etc. Moreover, ensemble methods or fusion techniques
Performance
logic FFNN FNN RNN RFNN learning are typically applied to this part for performance im-
Approximate
capability
+ ++ +++ +++ +++ +++ provement. As a result, physical system dynamics and
Robustness ++ + +++ ++ +++ +++ behaviors (e.g., degradation behavior) can be accurately
Computational characterized as offline models based on the information
+++ +++ ++ + + +
burden of the unit population.
Dataset 2) Condition monitoring and health assessment: This part
++ +++ ++ + + N/A
requirement
deals with the health assessment of the unit in service
Dynamics No No No Yes Yes Yes
Expert
subjected to the online condition monitoring in field
knowledge applications. The offline model is tailored and individu-
Yes No Yes No Yes No
embedded alized to the unit in service through the model parameter
capability
tuning layer by adapting to field operational environment
and workload. The functions include the noninvasive
parameter identification, data preprocessing (e.g., data
A similar example can be found in [140], where RL is applied cleaning), feature mining, anomaly detection, fault di-
to the MPPT control of a buck converter of photovoltaic arrays. agnosis, and RUL prediction. In this way, insightful
For the neural network-based controller, the learning process knowledge for decision-making can be extracted from
is completed from examples provided by an external super- the continuous condition monitoring information.
visor. While the RL controller can learn the experience by 3) Management and decision-making: In this part, the sup-
directly interacting with the environment through actions and portive knowledge of health assessment are returned for
rewards. It is worth mentioning that the training of the RL optimal decision making. With this feedback, control
controller is based on the interactions between the controller policies (e.g., power routing) can be adjusted to max-
and the system, and the offline dataset is unnecessary in this imize the system performance given the real-time health
case. As a result, the RL-based controller is beneficial to new status. Moreover, economical maintenance policy can
systems without existing datasets. be made to facilitate the condition-based and predictive
maintenance.
D. Discussions Subsequently, the relevant applications of AI in maintenance
in terms of these three parts are discussed in detail.
A summary of the advantages and limitations of AI al-
gorithms in control applications is given in Table IV. It is
worth mentioning that the dynamic performance, robustness,
A. Condition Monitoring
generalization, and convergence speed of AI algorithms are
critical in control applications. The algorithm complexity Condition monitoring [20], [157], [158] in power electronics
and computational burden are the major challenges. Thus, includes system parameter identification, data preprocessing,
high-performance Digital Signal Processor (DSP) or Field and feature mining. The condition monitoring information is
Programmable Gate Array (FPGA) is necessary for practical applied to uncover hidden and informative insights, which
implementations. serve as a basis for the subsequent PHM applications.
1) System Parameter Identification: The system parameter
identification [159] deals with information acquisition for crit-
V. M AINTENANCE
ical components. Developing specific hardware for parameter
Although reliability characteristics have been elaborately identification (e.g., temperature-sensitive electrical parameters
considered in design and control, power electronic systems still of IGBTs [158]), however, is quite a challenging task due to
undertake various risks and even catastrophic failures due to features of power electronic systems, e.g., very tight space
complex and severe working environments [18], [154], [155]. in a power module, very fast switching frequency, relatively
The reliability and safety of power electronic components, insignificant parameter changes in terms of aging [157], etc.
converters, and systems are of great importance for field One of the promising solutions is noninvasive method without
applications. In maintenance, preventive activities, including any extra hardware implementation, where information of
condition monitoring, anomaly detection, fault diagnosis, RUL interest can be inferred or estimated indirectly from available
prediction, etc., are effective approaches to ensure that in- physical signals. As a result, the condition monitoring can
tended functions can be properly executed. These activities are be implemented with a sensorless and cost-efficient solution,
aligned with the IEEE standard framework of PHM for elec- which is favorable for industrial practitioners. Generally, the
tronic systems [156]. Fig. 14 presents a systematic flowchart of system parameter identification can be categorized into model-
maintenance activities in power electronic systems. Generally, free and model-based methods, considering whether the sys-
it consists of three parts: tem dynamics and models are required.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
16

Fig. 14. Flowchart of maintenance in power electronic systems.

For the model-free method, no prior knowledge of the

system dynamics is required. Essentially, it deals with the re-
gression capability of AI algorithms to construct a relationship
between the inputs and outputs. For example, in a three-phase
front-end diode bridge motor drive, the current ia,out in a-phase
and the DC-link ripple voltage ∆vdc are considered as the
inputs, and the capacitance C is applied as the output for the
training of an FFNN [86]–[88]. In this way, the relationship
between the input signals and the capacitance is established
and thus the capacitance can be inferred indirectly. Similarly,
it is demonstrated that the capacitance can be estimated by
the FFNN constructed by the frequency domain information
of DC-link voltage ripple. The potentials of FFNN in the
capacitance estimation are illustrated in a hardware prototype
Fig. 15. Examples of model-free methods of system parameter identification
[88]. with AI. (a) Capacitance identification of DC-link capacitor [88]; (b) a-
phase current estimation for calculating the impedance measurement of power
In [108], considering the dynamic capability of RNN, an electronic system [108]; and (c) Equivalent series resistance (ESR) estimation
impedance identification method is proposed based on RNN to in future p steps for supercapacitors [103].
enable the stability analysis for power electronic systems over
a wide frequency range. The RNN is applied to build a model
that can produce identical outputs as the physical system A summary for the framework of model-free parameter
given the same inputs. The inputs of RNN include three-phase identification methods is given in Fig. 15. It can be seen that AI
voltages va , vb , vc . The output is the a-phase current ia . As a methods serve as the regression tool f (·) between the available
result, the RNN-based model possesses the same frequency input signals and the parameter to be monitored.
characteristics as the physical one. It can be performed for The model-free method is attractive for industrial appli-
the impedance identification without interrupting the system cations due to less hardware cost. However, it is typically
operation. sensitive to external noise and disturbance due to the lack
In [103], an improved ANFIS is applied to estimate the of system model. Thus, its robustness should be carefully
capacitance and equivalent series resistance (ESR) of the considered. This issue can be possibly mitigated with a large
supercapacitor. At monitoring time t, the inputs of the ANFIS amount of data in the training stage [159] to cover situations in
include the supply voltage Vt , the supercapacitor temperature field applications as much as possible. Nevertheless, the data
θt , and a time series ESRt−400:100:t consisting of 5 previous collection is time-consuming and costly.
ESR data points. The output of the ANFIS is the ESR Another category of the system parameter identification is
estimations in future p steps. Experimental analysis indicates the model-based method. As the name implies, for a model-
that ESR of supercapacitor can be accurately estimated and based method, system physics and models are partially known
the normalized root mean square error of the ESR estimation in advance and the identification model is formulated with un-
is as small as 0.025 at condition monitoring time of 2600 h. known model parameters. In this way, the system identification

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
17

Fig. 16. Dynamic model of a PV panel for parameter identification with

model-based method [69]. System parameters includes the input current Iph ,
output current Io (i(t) ), voltage vsh across capacitor Csh , resistor Rsh , p-n
junction capacitance Csh , and resistor Rs .

task is equivalent to the exploration of optimal parameters of

the model, which is essentially an optimization task. In this
case, AI, especially the metaheuristic methods, is utilized as an
optimizer to find the optimal solutions. Numerous approaches
such as PSO [57], crow search algorithm [73], GA [69], etc, Fig. 17. k-means clustering method for discretization of filtered degradation
paths of increment of drain-to-source on-state resistance ∆RDS(on) of different
or their improved variants, can be exploited. power MOSFET devices #26, #29, #32, #35, #36, #37, #38 [131].
In [69], a parameter identification method for the health
diagnostic of a PV panel is developed. The equivalent circuit
of the PV panel is given in Fig. 16, and its system model is should be considered. For example, for the condition moni-
explicitly derived as toring of power MOSFETs in [131], the device is considered
as failed if there is an increase of 0.08 Ω for the degradation
 vsh vsh − v indicator of drain-to-source on-state resistance RDS(on) . Such a
iCsh (vsh , v)= Iph − iD (vsh ) −
 − ,
 R sh Rs  tiny increment is challenging to observe. Thus, more research
dv (v , v) 1 vsh vsh − v efforts are necessary to improve the sensitivity of the AI-based
 sh sh
 = Iph − iD (vsh ) − − ,
dt Csh Rsh Rs parameter identification methods. Moreover, it is worth men-
(5) tioning that computational burden and embedded capabilities
where Iph is the input current, Io is the output current, vsh is should be considered for field applications.
the voltage across the capacitor Csh , Rsh is the resistance, 2) Data Preprocessing and Feature Mining: Data prepro-
and Csh is the p-n junction capacitance. As a result, the cessing and feature mining are concerned with refining the
parameter identification is equivalent to find a parameter set raw data to better serve the applications, e.g., fault diagnosis.
G = {Iph , Io , vsh , Rsh , Csh , Rs } that ensures an identical By exploring dataset structure, it includes data cleaning to
output as the physical system. By injecting large signal distur- reduce noise, data clustering to discover groups of similar
bances to the panel voltages in the testing stage, the dynamic data points, density estimation to identify the data distribution,
response of the current-voltage characteristics is sampled to data compression that projects high-dimensional data down to
calculate the objective function as low-dimensional data to reduce the number of features, data
N2 fusion to integrate multiple information sources, etc. Typically,
1 X 2 the performance of the subsequent PHM application, e.g., the
fobj (G) = (ip [k] − i[k]) , (6)
N1 − N2 + 1 diagnostic accuracy, can be significantly improved if the data
k=N1
preprocessing and feature mining are properly conducted.
where ip [k] and i[k] are the current output of the model and In [131], a reliability assessment method for power MOS-
the physical system, respectively, and N1 and N2 are the start FETs based on a continuous-time Markov chain is proposed.
index and the end index for the sampling. Subsequently, an To discretize the continuous degradation path of power MOS-
improved GA method is used to explore an optimal solution FETs without breaking the inherent monotonicity, a k-means
minimizing fobj (G) in (6). A similar investigation can be method is applied to divide the evolution of drain-to-source
found in [57], where a modified PSO algorithm is applied on-state resistance RDS(ON) into 11 discrete states, as shown
to the internal parameter identification of a PV panel. in Fig. 17.
Due to the involvement of system dynamics and models, the In [133], a health state identification method for IGBTs
amount of data required for the estimation can be significantly based on self-organizing maps (SOMs) is proposed. It is
reduced for the model-based methods. Also, the overfitting risk essentially a clustering task. The states of the device are
in the model-free methods can be mitigated. It exhibits better clustered as the healthy state, partially degraded state, heavily
dynamics to handle unexpected disturbance and switchable degraded state, and failure state, considering the distance
working modes. However, due to the system complexity, the between the input measurements (including collector current
system dynamics and models are challenging to formulate in Ic , collector-emitter voltage Vce , and case temperature T ) and
most cases. the best matching unit of the trained SOMs.
For parameter identification methods in power electronics, In [160], a composite failure precursor of SiC MOSFETs
the accuracy and robustness under the complex environment is developed with a data fusion technique of genetic pro-

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
18

gramming, which is a variant of GA. It integrates multiple

degradation signals of a power semiconductor device in a
nonlinear way. Since the composite failure precursor is directly
optimized in terms of the RUL prediction model, the prediction
accuracy is improved by 35.3% and the prediction uncertainty
is reduced by 16.3%. It indicates that data fusion in condi-
tion monitoring is potentially useful especially for system-
level applications (e.g., converters), where multiple physical
degradation signals exist. Fig. 18. Feed-forward neural network for anomaly detection of a full-bridge
diode rectifier. The inputs of the neural network include input voltage vi (t),
An integrated toolbox “Diagnostic Feature Designer” for input current ii (t), and output current io (t); the output of the neural network
the feature identification is available in Matlab [161], which is output voltage vo (t) [93].
can be applied to the data preprocessing and feature mining
as an automatic tool.
ANFIS to determine the severity levels of a capacitor in the
DC-link filter [102].
B. Anomaly Detection and Fault Diagnosis In [112], a multi-switches fault diagnosis algorithm for
The anomaly detection makes a binary decision and focuses voltage-source inverters is proposed. An echo state network
on the abnormal behavior identification. It provides an indica- (ESN) is used as a diagnostic classifier given small low-
tion when the rated system characteristics or nominal param- frequency data. Note that ESN is an improved variant of RNN
eters exceed the predefined safety range. Once the anomaly to avoid gradient exploding and vanishing in the training. In
behavior occurs, the fault diagnosis [19] identifies and locates this work, the diagnostic performance of ESN is compared
the detailed failure modes subsequently. Essentially, anomaly with the FFNN, the FFNN with a wavelet activation function,
detection and fault diagnosis are the classification, regression, and the RBFN. It indicates that the ESN is superior in the
or clustering tasks. Based on the learned relationship from sensitivity, design process, and training speed.
the training stage, it determines the fault label when a new In [115], an 1-D convolutional neural network (CNN) is
fault signature becomes available. Note that the feasibility applied to the fault diagnosis of a modular multilevel converter.
of AI-based anomaly detection and fault diagnosis is based One advantage of 1-D CNN is that the feature extraction
on two assumptions [33]: firstly, the fault occurrence in any and diagnostic classification can be integrated together, which
components has an impact on the fault signature; secondly, enables the fault diagnostics on the raw data directly. In
the impact on these signatures varies with different fault this way, the feature extraction, which is usually experience-
modes and fault locations. The methods of anomaly detection intensive, can be avoided. The experimental results indicate
and fault diagnosis can be categorized as supervised learning that the proposed method is highly reliable and provides a
methods and unsupervised methods. detection accuracy of 98.9% and a fault diagnostic accuracy
1) Supervised Learning Methods: In [93], an FFNN is of 99.7% within 100 ms.
applied to establish the nonlinear relationship of the inputs In addition to the above neural network-based methods,
and outputs of a full-bridge diode rectifier. The training of kernel methods, including the support vector machine and
the FFNN is completed at the normal operation mode of the relevance vector machine, are also applied for anomaly
the rectifier, as shown in Fig. 18. As a result, the principles detection and fault diagnosis. One advantage of the kernel
and mapping relationship between the inputs, including input methods is that the dataset size requirement is relatively lower
voltage vi (t), input current ii (t), and output current io (t), and than the neural network-based methods.
the output signal of output voltage vo (t) are characterized, In [7], based on the time-domain fault features, a support
considered as a digital emulator indicating the normal oper- vector machine-based fault diagnosis method is proposed for
ational mode of the rectifier. This digital emulator and the incipient yet progressive faults of IGBTs in an inverter. The
physical rectifier are simultaneously operated and their outputs training of SVM can be completed by metaheuristic methods
are compared in real-time. Once the monitored output voltage (e.g., PSO, GA, etc.). For a total of 41 fault classes, it achieves
of physical rectifier significantly deviates from the output of an average accuracy of 94.82% being robust to both load
FFNN, it suggests that the rectifier runs into an abnormal variations and motor parameter shifts.
mode, which facilitates the anomaly detection. In this case, In [127], a relevance vector machine (RVM) is applied
the FFNN essentially serves as the regression tool. for the fault diagnosis of a cascaded H-bridge multilevel
In [90], an open-circuit fault diagnosis algorithm is proposed inverter. Principal component analysis (PCA) is applied to
for the inverter in a microgrid system subjected to varying load extract the fault signal feature. Experimental analysis indi-
conditions. A signal processing method is proposed to reduce cates that the RVM outperforms the FFNN and the SVM,
the required amount of information for the fault representation with 100% diagnostic accuracy in this specific case study.
and suppress the impact of the load change. An FFNN is Compared to SVM with the direct fault label as its output,
used as a diagnostic classifier. The computational burden RVM is formulated under the Bayesian framework. It makes
of the proposed method can be reduced to 10% of that of probabilistic outputs of the fault information, which possesses
the existing algorithms. In this case, the FFNN serves as a good theoretical guidance and is favorable to the uncertainty
classification tool. Similar fault diagnosis ideas include the analysis on diagnostic results. Generally, for the same task,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
19

the RVM is sparser than SVM, indicating faster speed for field
applications. However, the training time of RVM is generally
longer than SVM.
2) Unsupervised Learning Methods: In [136], principal
component analysis (PCA) is applied to the anomaly detec-
tion of SiC MOSFETs. Multiple statistical features, including
kurtosis, skewness, etc., are considered as the inputs of the
PCA algorithm. The output is compact with fewer features
and a transformation matrix. For field applications, the newly
available data is applied to the transformation matrix for the
calculation of an anomaly index. Abnormal behavior is notified
when the anomaly index exceeds a predefined threshold. The
method is verified by a processor-in-the-loop experiment. This
detection mechanism is similar to [93]. Other unsupervised
learning methods in anomaly detection and fault diagnosis,
including k-means and SOMs, can be found in [118]. Fig. 19. Flowchart and procedures of AI methods for remaining useful
3) Discussions: Table V summarizes the features of typical life (RUL) prediction of power electronic systems. PDF: probability density
function.
AI algorithms and their variants for anomaly detection and
fault diagnosis. It can be seen that each AI algorithm possesses
advantages and limitations. To fully exploit the advantages of ample data of the normal operation case and the scarcity
each algorithm, it is effective to combine multiple algorithms of data with fault labels due to catastrophic failures.
for a decision-level fusion to improve the diagnostic accuracy Thus, the algorithm applicability given limited size of
and robustness. An example of decision-level fusion for fault dataset and poor quality dataset should be investigated.
diagnosis of IGBTs can be found in [96]. More ensemble 3) The practicality, including computational burden, adap-
methods to combine multiple algorithms can be found in tive capability, robustness, difficulty of algorithm design
Chapter 14 in [1]. From the AI perspective, there is a negligible and debugging [112], implementation cost, etc, should
difference between power electronics and other engineering also be comprehensively considered.
areas (e.g., electromechanical applications) in terms of the
anomaly detection and fault diagnosis tasks. Two reviews of C. Remaining Useful Life Prediction
AI methods in anomaly detection and fault diagnosis can be
found in [162], [163]. Lifetime prediction in the design phase is to support the
Note that various AI methods and their variants have been DfR, which refers to the feature of a population of units. As
successfully applied to anomaly detection and fault diagnosis. one of the critical aspects of Prognostics and Health Manage-
There are differences in terms of how the data are collected and ment [165], the RUL prediction is not to predict the lifetime
types of available data in different applications, which is an of a population of units. It predicts the residual lifetime of an
important aspect of practical applications of AI. An integrated individual unit in service based on the condition monitoring
platform “Predictive Maintenance Toolbox” is available in information. There are associated uncertainties in the lifetime
Matlab [164], which includes various algorithms of anomaly prediction, including model calibration errors, manufacturing
detection and diagnostics. It is beneficial for the method tolerances, variations of operational environments and work-
development and benchmark analysis. From the AI perspec- load, etc. These uncertainties result in inaccurate reliability
tive, most of the methods can be interchangeably applied estimates for a specific unit in field operation [166]. RUL
with a comparable performance in terms of the evaluation prediction is applied as an additional tool to reduce the un-
accuracy. Although the accuracy can be further improved certainties for reliability-critical, safety-critical, or availability-
by advanced algorithms (e.g., deep learning methods), the critical applications.
accuracy improvement after a high score, e.g., 90%, is rela- The flowchart and procedures for RUL prediction are given
tively less significant compared with other practical concerns. in Fig. 19. The regression model can be established based on
More considerations should be devoted to the gap between historical dataset. The probability density function (PDF) of
theoretical algorithms and practical implementations, where degradation level at any specific condition monitoring time
the practical considerations include can be estimated based on the regression model. The PDF
of the RUL can be derived from the PDF of the degradation
1) In addition to the single component fault, the failure
level. Given the fact that the system is properly functioning
mode of multiple components failed simultaneously
at condition monitoring time t, its RUL l is defined as the
should be considered. The dependence and coupling
residual lifetime when the degradation process D(t) exceeds
effects among the component failures should be incor-
the failure threshold w, i.e.,
porated into the diagnostic algorithms.
2) Considering the challenges in the data acquisition of
l = inf {l : D(t + l) ≥ w | D(t) < w, D1:j } , (7)
power electronic systems, the training dataset for practi-
cal application is typically limited. This situation is even where D1:j is the cumulative CM information up to time
worse for a dataset with unbalanced fault labels, i.e., the t. Note that RUL l is a random variable. In addition to its

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
20

Table V: Algorithm comparisons of anomaly detection and fault diagnosis in power electronic systems. N/A: not available

References Methods Number of faults Speed Accuracy Advantages & Limitations

Feed-forward neural - Training may result in a local solution
[94] 5 N/A >85.0%
network - Sensitive to weights
- Training solution is globally optimal
[112] Echo state network 4 - Training time: 0.0626 s N/A - Less sensitive to weights
- Fast training speed
1-D convolutional
[115] 9 - Testing time: <0.1 s 99.70% - Feature mining is unnecessary
neural network
[7] Support vector machine 41 N/A 94.82% - Lower requirement for feature mining
Relevance vector - Probabilistic outputs
[127] 33 - Testing time: <0.14 s >97.30%
machine - Higher model sparsity and speed
- Probabilistic outputs
[117] Bayesian network 22 - Training time: 0.172 s 98.99%
- Better toleration for sensor noise and bias
[33] Fuzzy logic 12 - Testing time: <0.017 s N/A - Fast testing speed
[96] Ensemble model 3 N/A >97.0% - Better diagnostic accuracy and robustness

Fig. 20. Remaining useful life prediction of power MOSFETs based on echo
state network [111]. For the network training, the input weights W in and the
recurrent weights W are randomly generated. The output weights W out are
estimated by least-square methods.
Fig. 21. Gaussian processes regression for the remaining useful life prediction
of IGBTs [119].

expected value, the uncertainty metrics with the lower and up-
per confidence interval (llo , lup ) are also of great importance. time is established by the Gaussian processes regression. Since
AI methods in RUL prediction is typically dealing with a Gaussian process is formulated with the Bayesian framework,
nonlinear regression between the degradation information and it is able to predict the uncertainty of variation ∆Vce,on
the corresponding RUL based on the training dataset [167]. intrinsically. It can be seen from Fig. 21 that the error bar of the
In this way, degradation patterns can be characterized. Once evolution of ∆Vce,on is explicitly derived, which can be further
the degradation patterns have been learned, it can be directly utilized for the calculation of the confidence interval of RUL.
projected based on the regression model to facilitate the future Another example of kernel method for RUL prediction can be
degradation level prediction. As a result, the RUL can be found in [74], where a support vector machine is applied to
estimated. the degradation modeling of a buck converter.
In [111], an echo state network is applied to the RUL To make AI-based methods of the RUL prediction more
prediction of power MOSFETs, as shown in Fig. 20. The input practical for field applications, more efforts should be devoted
of the echo state network is the degradation indicator drain-to- to the following aspects:
source on-state resistance RDS,(on) at times k−1 and k, and the 1) Uncertainty quantification: Compared to other
output is the RDS,(on) at time k + 1. To facilitate the adaptation regression-related tasks, e.g., control applications,
of the echo state network, a particle filter is exploited to the capability of uncertainty quantification is more
recursively update the output weights when new condition critical for RUL prediction. As shown in Fig. 19, the
monitoring data of the in-situ device becomes available. In RUL is a random variable and thus quantification of
this way, the degradation model is adaptive to varying external the confidence interval is essential for the optimal
environments and operational modes. Another neural network decision-making. These uncertainties come from
method involving time-delayed neural network for the RUL the population heterogeneity, measurement noise,
prediction of IGBTs can be found in [114]. varying operational settings, etc, which should be
In [119], Gaussian processes regression is applied to the comprehensively considered for a practical solution.
RUL prediction of IGBTs. For the degradation modeling, AI methods are rather challenging for the uncertainty
the nonlinear relationship between the decrement of on-state quantification of prediction results considering the
collector-emitter voltage ∆Vce,on and the condition monitoring black-box feature. Several feasible approaches include

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
21

Table VI: Requirements of AI for exemplary applications in design, control,

the Monte-Carlo methods [114], incorporating particle and maintenance. High: +++, moderate: ++, low: +
filter in the neural network [111], and Bayesian-based
Requirements Heatsink Intelligent RUL
AI methods (e.g., Gaussian process, relevance vector design controller prediction
machine). Another promising direction is the stochastic (Design) (Control) (Maintenance)
data-driven methods [154], [160], [168], which can Computation Effort +++ ++ ++
intrinsically provide the PDF of the RUL for calculating Algorithm Speed + +++ ++
the confidence interval. Accuracy ++ +++ +++
2) Adaptive capability: It is concerned with the the model Dataset requirement + + +++
Interpretability + +++ +++
parameter tuning layer in Fig. 14 for connecting the
offline models and the online models, which is a key
step for practical applications. If a specific AI method
lacks an adaptive capability, its application is limited stability needs to be theoretically ensured and thus the in-
since one prerequisite is that the training data and the terpretability is critical. The intelligent controller is generally
test data should be generated under similar situations tuned online, it is unnecessary to prepare the dataset for the
(e.g., external environments and operational modes) and model training.
share a high-level similarity [95]. It is challenging for For the RUL prediction of switching devices in a converter
power electronics since operational settings of the in- system, the requirement of the algorithm speed is moderate
situ system (i.e., the test data) are quite different from since the device degradation is slow and the long time span of
that of the training dataset, which is generally obtained decision making is acceptable. The degradation model for the
with accelerated testing experiments. The majority of the RUL prediction can be prepared in offline mode and efficiently
research [74], [114], [119] assumes that the operational tuned in online mode, and the computational effort in this
settings of the in-situ system are identical to the training application is moderate. Since the model accuracy is highly
dataset (e.g., accelerated aging experiments), which may dependent on the dataset, the dataset requirement, e.g., dataset
not be the case in field applications. Thus, the adaptive quality, dataset size, label balance (e.g., limited abnormal data
capability of the AI-based RUL prediction method is in the training dataset), etc., is the most critical. Moreover, the
critical to bridge academic research and industrial appli- interpretability of the RUL prediction results with uncertainty
cations. Other promising directions of model parameter is critical as well. As a result, a comparison of AI algorithms
tuning include the explicit mapping relationship deriva- in each phase of the life-cycle of power electronic systems is
tions [169] and transfer learning [170], [171] of degra- provided in Table VI.
dation characteristics under various operational settings It is concluded that AI possesses immense potentialities in
(temperature, voltage, humidity, etc.). This may, how- power electronic systems. Many opportunities and issues are
ever, imply intensive investigations of system models. yet to be explored as follows:
1) Motivations and Justifications of AI Applied to Power
Electronic Systems: Although there are numerous studies
VI. O UTLOOK ON AI FOR P OWER E LECTRONIC S YSTEMS
on AI for power electronic systems in the literature since
From the algorithm perspective, it is necessary to investigate the 1990s, the practical implementations in industry are
the features of AI when it applies to different life-cycle phases. still limited, which is a sharp contrast compared to the
By using a power converter system, three specific examples claimed AI potentials. It is necessary for deeper inves-
are applied to illustrate the requirements of AI methods for tigations into tasks where AI can essentially outperform
each life-cycle phase. conventional methods. The justifications of AI-based
For the heatsink design of a converter system, a large num- solutions should be clearly identified by comparing to
ber of decision variables, e.g., weight, volume, pattern, need conventional methods from the industrial perspectives,
to be determined, which is essentially an optimization task. e.g., implementation complexity, algorithm accuracy
The metaheuristic methods are applied to the optimization that and robustness, algorithm accountability, extra hardware
involves an iterative trial-and-error procedure. Although the cost, computational energy consumption, embedded ca-
computational effort is intensive, the design task is typically pability, etc.
performed offline. There is less requirement on the algorithm 2) Interwoven AI implementations through Life-cycle
speed in this case. Although the metaheuristic method-based Phases: Implementations of AI in each life-cycle phase
optimization does not ensure a global solution, the sub-optimal of design, control, and maintenance, will facilitate flex-
heatsink design is still superior and satisfactory in most cases. ible functional interactions. This feature is beneficial to
Thus, the algorithm accuracy is not critical as well. The overall performance optimization and procedure simpli-
training dataset and interpretability of the optimization process fication. It enables the system capability in managing
are not required. data flow between electrical and other disciplines (e.g.,
For the intelligent controller of a converter system, the real- mechanical area) [13] as well. For example, aging in-
time control errors, e.g., the voltage error, the current error, formation obtained by the AI-based system parameter
need to be returned to the controller for the adaptive updating identification can be flexibly incorporated into the AI-
in an online mode. Thus, the requirements of algorithm speed based controller for reliability improvement. Therefore,
and accuracy are the most critical. In addition, the controller more attention should be paid to the interwoven inter-

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
22

actions powered by AI. tion may be not feasible in the future. Thus, for power
3) Multilevel Information Fusion: Robustness is essential electronics applications, it is promising to develop a
for safety-critical power electronic systems. Multiple collaborative learning scheme for AI algorithms without
sources of information and models are available in collectively aggregating data from different locations,
most cases for a specific application of power electronic e.g., federated learning [174]. It is well aligned with the
systems. If these information sources and models are si- trend of data privacy regulations for the implementation
multaneously exploited, possible biases can be mitigated of AI solutions.
to improve robustness. Multilevel information fusion can 8) Power Electronics Database: Due to the complexity
be performed at the data-level [160], [172], feature-level, of system dynamics of power electronics, extensive
decision-level [96], and their combinations, in order to datasets are required for the model training, especially
exploit the insights of each information sources. For for the maintenance applications. While the experi-
example, the well-established differential equations of mental testing for data collection is generally time-
power converter system can be integrated with AI as a consuming and expensive. There is a compelling demand
hybrid solution for condition monitoring. As a result, for building up common power electronics data and
the advantages from both the model-driven side and the knowledge base. These open-source datasets are critical
data-driven side can be gained for better accuracy and to benchmark algorithm performance and accelerating
robustness. application development. It will benefit the global power
4) Computation-light AI: Compared to other industrial ar- electronics communities in academia and industry.
eas (e.g., image recognition), one of the key features of
power electronic systems is that there is no powerful
VII. C ONCLUSIONS
computation unit. While real-time applications, e.g.,
control, impose a rigid requirement on the algorithm Existing AI methods in power electronic systems are com-
speed. Although complex deep learning techniques [170] prehensively reviewed in this paper. New findings are identi-
can provide superior performance, it is computationally fied as
intensive for power electronic systems. A prospective 1) From the application perspective, the AI methods ap-
direction is the computation-light AI algorithms that plied in power electronic systems can be categorized
can be implemented on cost-effective units but provide as the design, control, and maintenance. The usage
comparable performance with deep learning algorithms. percentage, application trend, features, and requirements
5) Data-light AI: One of the bottlenecks of AI implemen- of AI in each life-cycle phase are discussed.
tation on power electronic systems is the dataset. For 2) From the method perspective, the AI methods applied in
example, AI-based solutions for remaining useful life power electronic systems can be categorized as expert
prediction requires the dataset to be versatile enough system, fuzzy logic, metaheuristic methods, and machine
for accurate degradation behavior learning. However, learning. The usage percentage, advantages, and limi-
the dataset size is generally small since the degradation tations of relevant AI algorithms in each category are
experiments are resource-consuming. This situation is comprehensively compared.
even worse for safety-critical cases. Thus, developing 3) From the function perspective, the AI-related appli-
AI algorithms with lower dataset requirement, i.e., data- cations are essentially dealing with the optimization,
light AI solutions that can provide acceptable perfor- classification, regression, and data structure exploration.
mance in the presence of poor datasets, is a prospective 4) The milestones of relevant algorithm variants and appli-
direction. cations are identified and organized as a timeline map.
6) Explainable AI: Most of the AI algorithms in power 5) For each life-cycle phase, illustrative examples are dis-
electronics suffer from the “black-box” feature. For cussed and the challenges and future research opportu-
example, most of the AI-based solutions for remaining nities are identified.
useful life prediction can only provide a point estimation
without sensitivity analysis and uncertainty quantifi-
cation. It makes AI-based solutions opaque and less R EFERENCES
convincing for practitioners to implement in industry [1] C. M. Bishop, Pattern Recognition and Machine Learning. Springer,
applications, especially for safety-critical cases. There is 2006.
a pressing need to improve the algorithm transparency [2] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
for explainable AI with better interpretability. Under- press, 2016.
[3] T. Wu, Z. Wang, B. Ozpineci, M. Chinthavali, and S. Campbell,
standing how models come up the decisions is critical for “Automated heatsink optimization for air-cooled power semiconductor
model simplification and safety, with which AI solutions modules,” IEEE Trans. Power Electron., vol. 34, no. 6, pp. 5027–5031,
can be implemented with confidence. Jun. 2019.
[4] X. Zhan, W. Wang, and H. Chung, “A neural-network-based color
7) Dataset Privacy: An increasing attention has been paid control method for multi-color LED systems,” IEEE Trans. Power
to the data privacy, e.g., General Data Protection Reg- Electron., vol. 34, no. 8, pp. 7900–7913, Aug. 2019.
ulation (GDPR) [173] in the European Union. With [5] C. Wei, Z. Zhang, W. Qiao, and L. Qu, “Reinforcement-learning-based
intelligent maximum power point tracking control for wind energy
these critical regulations, the training of standard AI conversion systems,” IEEE Trans. Ind. Electron., vol. 62, no. 10, pp.
algorithms is challenging since a centralized data collec- 6360–6370, Oct. 2015.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
23

[6] C. Wei, Z. Zhang, W. Qiao, and L. Y. Qu, “An adaptive network- [29] Y. Izuno, R. Takeda, and M. Nakaoka, “New fuzzy reasoning-based
based reinforcement learning method for MPPT control of PMSG wind high-performance speed/position control schemes for ultrasonic motor
energy conversion systems,” IEEE Trans. Power Electron., vol. 31, driven by two-phase resonant inverter,” in IEEE Industry Appl. Soc.
no. 11, pp. 7837–7848, Nov. 2016. Ann. Meeting, 1990, pp. 325–330.
[7] I. Bandyopadhyay, P. Purkait, and C. Koley, “Performance of a clas- [30] M. G. Simoes, B. K. Bose, and R. J. Spiegel, “Design and performance
sifier based on time-domain features for incipient fault detection in evaluation of a fuzzy-logic-based variable-speed wind generation sys-
inverter drives,” IEEE Trans. Ind. Informat., vol. 15, no. 1, pp. 3–14, tem,” IEEE Trans. Ind. Appl., vol. 33, no. 4, pp. 956–965, Jul./Aug.
Jan. 2019. 1997.
[8] A. E. Mejdoubi, H. Chaoui, J. Sabor, and H. Gualous, “Remaining [31] R. Osorio, J. M. Alonso, N. Vazquez, S. E. Pinto, F. D. Sorcia-Vazquez,
useful life prognosis of supercapacitors under temperature and voltage M. Martinez, and L. M. Barrera, “Fuzzy logic control with an improved
aging conditions,” IEEE Trans. Ind. Electron., vol. 65, no. 5, pp. 4357– algorithm for integrated LED drivers,” IEEE Trans. Ind. Electron.,
4367, May 2018. vol. 65, no. 9, pp. 6994–7003, Sep. 2018.
[9] F. Tao, H. Zhan, A. Liu, and A. Y. C. Nee, “Digital twin in industry: [32] A. Bubshait and M. G. Simoes, “Design of fuzzy logic-based dynamic
State-of-the-art,” IEEE Trans. Ind. Informat., vol. 15, no. 4, pp. 2405– droop controller of wind turbine system for primary frequency support,”
2415, Apr. 2019. in IEEE Industry Appl. Soc. Ann. Meeting, 2018, pp. 1–7.
[10] X. He, W. Shi, W. Li, H. Luo, and R. Zhao, “Reliability enhancement [33] W. Q. Chen and A. M. Bazzi, “Logic-based methods for intelligent
of power electronics systems by big data science,” Proc. CSEE, vol. 37, fault diagnosis and recovery in power electronics,” IEEE Trans. Power
no. 1, pp. 209–221, Jan. 2017. Electron., vol. 32, no. 7, pp. 5573–5589, Jul. 2017.
[34] M. G. Simoes and A. Bubshait, “Frequency support of smart grid
[11] K. L. Tsui, Y. Zhao, and D. Wang, “Big data opportunities: system
using fuzzy logic-based controller for wind energy systems,” Energies,
health monitoring and management,” IEEE Access, vol. 7, pp. 68 853–
vol. 12, no. 8, pp. 1–15, Apr. 2019.
68 867, May 2019.
[35] B. N. Alajmi, K. H. Ahmed, S. J. Finney, and B. W. Williams,
[12] S. E. De Leon-Aldaco, H. Calleja, and J. A. Alquicira, “Metaheuristic
“Fuzzy-logic-control approach of a modified hill-climbing method for
optimization methods applied to power converters: A review,” IEEE
maximum power point in microgrid standalone photovoltaic system,”
Trans. Power Electron., vol. 30, no. 12, pp. 6791–6803, Dec. 2015.
IEEE Trans. Power Electron., vol. 26, no. 4, pp. 1022–1030, Apr. 2011.
[13] M. R. G. Meireles, P. E. M. Almeida, and M. G. Simoes, “A compre- [36] C. L. Tseng, S. Y. Wang, S. C. Chien, and C. Y. Chang, “Develop-
hensive review for industrial applicability of artificial neural networks,” ment of a self-tuning TSK-Fuzzy speed control strategy for switched
IEEE Trans. Ind. Electron., vol. 50, no. 3, pp. 585–601, Jun. 2003. reluctance motor,” IEEE Trans. Power Electron., vol. 27, no. 4, pp.
[14] B. K. Bose, “Neural network applications in power electronics and mo- 2141–2152, Apr. 2012.
tor drives-an introduction and perspective,” IEEE Trans. Ind. Electron., [37] R. C. Garcia, W. I. Suemitsu, and J. O. P. Pinto, “Takagi-Sugeno fuzzy
vol. 54, no. 1, pp. 14–33, Feb. 2007. model and control of a boost converter using Type-I internal model
[15] B. K. Bose, “Artificial intelligence techniques in smart grid and control,” in 39th Annu. Conf. the IEEE Ind. Electron. Soc., 2013, pp.
renewable energy systems-some example applications,” Proc. IEEE, 3794–3799.
vol. 105, no. 11, pp. 2262–2273, Nov. 2017. [38] F. Zidani, D. Diallo, M. E. H. Benbouzid, and R. Nait-Said, “A fuzzy-
[16] M. Seyedmahmoudian, B. Horan, T. K. Soon, R. Rahmani, A. M. based approach for the diagnosis of fault modes in a voltage-fed PWM
Than Oo, S. Mekhilef, and A. Stojcevski, “State of the art artificial inverter induction motor drive,” IEEE Trans. Ind. Electron., vol. 55,
intelligence-based MPPT techniques for mitigating partial shading no. 2, pp. 586–593, Feb. 2008.
effects on PV systems-a review,” Renewable Sustainable Energy Rev., [39] R. Fullér, Introduction to Neuro-Fuzzy Systems. Springer Science &
vol. 64, pp. 435–455, Oct. 2016. Business Media, 2000, vol. 2.
[17] A. Mellit and S. A. Kalogirou, “Artificial intelligence techniques for [40] Matlab, “Manual of Neural Network Toolbox,” http:
photovoltaic applications: A review,” Progress Energy Combustion Sci., //matlab.izmiran.ru/help/toolbox/nnet/, 2005.
vol. 34, no. 5, pp. 574–632, Oct. 2008. [41] J. Zhang, H. S. H. Chung, W. L. Lo, S. Y. Hui, and A. K. M. Wu,
[18] H. S.-H. Chung, H. Wang, F. Blaabjerg, and M. Pecht, Reliability of “Implementation of a decoupled optimization technique for design of
Power Electronic Converter Systems. Institution Eng. Technol., 2015. switching regulators using genetic algorithms,” IEEE Trans. Power
[19] M. Riera-Guasp, J. A. Antonino-Daviu, and G. Capolino, “Advances Electron., vol. 16, no. 6, pp. 752–763, Nov. 2001.
in electrical machine, power electronic, and drive condition monitoring [42] L. L. Jiang, D. L. Maskell, and J. C. Patra, “A novel ant colony
and fault detection: State of the art,” IEEE Trans. Ind. Electron., vol. 62, optimization-based maximum power point tracking for photovoltaic
no. 3, pp. 1746–1759, Mar. 2015. systems under partially shaded conditions,” Energy and Buildings,
[20] H. Soliman, H. Wang, and F. Blaabjerg, “A review of the condition vol. 58, pp. 227–236, Mar. 2013.
monitoring of capacitors in power electronic converters,” IEEE Trans. [43] Q. Wang and S. Niu, “Design, modeling, and control of a novel hybrid-
Ind. Appl., vol. 52, no. 6, pp. 4976–4989, Nov./Dec. 2016. excited flux-bidirectional-modulated generator-based wind power gen-
[21] M. Pecht and R. Jaai, “A prognostics and health management roadmap eration system,” IEEE Trans. Power Electron., vol. 33, no. 4, pp. 3086–
for information and electronics-rich systems,” Microelectronics Rel., 3096, Apr. 2018.
vol. 50, no. 3, pp. 317–323, Mar. 2010. [44] S. Lyden and M. E. Haque, “A simulated annealing global maximum
power point tracking approach for pv modules under partial shading
[22] L. Duchesne, E. Karangelos, and L. Wehenkel, “Recent developments
conditions,” IEEE Trans. Power Electron., vol. 31, no. 6, pp. 4171–
in machine learning for energy systems reliability management,” Proc.
4181, Jun. 2016.
IEEE, 2020, Early access.
[45] J. Zhang, Y. Shi, and Z.-H. Zhan, “Power electronic circuits design: A
[23] R. C. G. Joao Pinto, Burak Ozpineci, “Tutorial: Artificial intelligence
particle swarm optimization approach,” in Asia-Pacific Conf. Simulated
applications to power electronics,” in IEEE Energy Convers. Congr.
Evolution Learning, 2008, pp. 605–614.
Expo., 2019, pp. 1–139.
[46] Y. C. Hung, F. J. Lin, J. C. Hwang, J. K. Chang, and K. C. Ruan,
[24] J. Foutz, “Power supply circuit development estimating aid: an expert “Wavelet fuzzy neural network with asymmetric membership function
system application,” in Annu. IEEE Applied Power Electron. Conf. and controller for electric power steering system via improved differential
Expo., 1988, pp. 64–71. evolution,” IEEE Trans. Power Electron., vol. 30, no. 4, pp. 2350–2362,
[25] S. M. Chhaya and B. K. Bose, “Expert system aided automated design, Apr. 2015.
simulation and controller tuning of AC drive system,” in Proc. 21st [47] J. Yuan, B. Chen, B. Rao, C. Tian, W. Wang, and X. Xu, “Possible
Annu. Conf. IEEE Ind. Electron., vol. 1, 1995, pp. 712–718. analogy between the optimal digital pulse width modulation technology
[26] W. Li and J. P. Ying, “Design and analysis artificial intelligence (AI) and the equivalent optimisation problem,” IET Power Electron., vol. 5,
research for power supply - power electronics expert system (PEES),” no. 7, pp. 1026–1033, Aug. 2012.
in Annu. IEEE Applied Power Electron. Conf. and Expo., Vols 1-4, [48] S. Singh and B. Singh, “Optimized passive filter design using modified
2008, pp. 2009–2015. particle swarm optimization algorithm for a 12-pulse converter-fed
[27] D. Fezzani, H. Piquet, and H. Foch, “Expert system for the CAD in LCI-synchronous motor drive,” IEEE Trans. Ind. Appl., vol. 50, no. 4,
power electronics - application to UPS,” IEEE Trans. Power Electron., pp. 2681–2689, Jul./Aug. 2014.
vol. 12, no. 3, pp. 578–586, May 1997. [49] B. Zhao, X. Zhang, and J. J. Huang, “AI algorithm-based two-
[28] A. M. Elsaadawi, A. E. Kalas, and M. Fawzi, “Development of an stage optimal design methodology of high-efficiency CLLC resonant
expert system to fault diagnosis of three phase induction motor drive converters for the hybrid AC-DC microgrid applications,” IEEE Trans.
system,” in Int. Middle-East Power System Conf., 2008, pp. 497–502. Ind. Electron., vol. 66, no. 12, pp. 9756–9767, Dec. 2019.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
24

[50] F. J. Lin, L. T. Teng, and M. H. Yu, “Radial basis function network con- [70] J. Zhang, H. S. H. Chung, A. W. L. Lo, and T. Huang, “Extended
trol with improved particle swarm optimization for induction generator ant colony optimization algorithm for power electronic circuit design,”
system,” IEEE Trans. Power Electron., vol. 23, no. 4, pp. 2157–2169, IEEE Trans. Power Electron., vol. 24, no. 1-2, pp. 147–162, Jan. 2009.
Jul. 2008. [71] F. J. Lin, P. K. Huang, H. C. Wang, and L. T. Teng, “An induction
[51] W. Wang, H. S. Chung, R. Cheng, C. S. Leung, X. Zhan, A. W. generator system using fuzzy modeling and recurrent fuzzy neural
Lo, J. Kwok, C. J. Xue, and J. Zhang, “Training neural-network- network,” IEEE Trans. Power Electron., vol. 22, no. 1, pp. 260–271,
based controller on distributed machine learning platform for power Jan. 2007.
electronics systems,” in IEEE Energy Conversion Congr. Expo., 2017, [72] B. X. Li and K. S. Low, “Low sampling rate online parame-
pp. 3083–3089. ters monitoring of DC-DC converters for predictive-maintenance us-
[52] W. M. Lin and C. M. Hong, “A new Elman neural network-based ing biogeography-based optimization,” IEEE Trans. Power Electron.,
control algorithm for adjustable-pitch variable-speed wind-energy con- vol. 31, no. 4, pp. 2870–2879, Apr. 2016.
version systems,” IEEE Trans. Power Electron., vol. 26, no. 2, pp. [73] Q. Sun, Y. Wang, Y. Jiang, and L. Shao, “Non-invasive condition
473–481, Feb. 2011. monitoring for boost converter based on crow search algorithm,” J.
[53] F. J. Lin, L. T. Teng, and H. Chu, “A robust recurrent wavelet neural Intell. & Fuzzy Syst., vol. 34, no. 6, pp. 3661–3670, Jun. 2018.
network controller with improved particle swarm optimization for [74] L. Wang, J. Yue, Y. Su, F. Lu, and Q. Sun, “A novel remaining
linear synchronous motor drive,” IEEE Trans. Power Electron., vol. 23, useful life prediction approach for superbuck converter circuits based
no. 6, pp. 3067–3078, Nov. 2008. on modified grey wolf optimizer-support vector regression,” Energies,
[54] M. A. Hassan and M. A. Abido, “Optimal design of microgrids in vol. 10, no. 4, p. 459, Apr. 2017.
autonomous and grid-connected modes using particle swarm optimiza- [75] A. Kavousi, B. Vahidi, R. Salehi, M. K. Bakhshizadeh, N. Farokhnia,
tion,” IEEE Trans. Power Electron., vol. 26, no. 3, pp. 755–769, Mar. and S. H. Fathi, “Application of the bee algorithm for selective
2011. harmonic elimination strategy in multilevel inverters,” IEEE Trans.
[55] K. Ishaque, Z. Salam, M. Amjad, and S. Mekhilef, “An improved Power Electron., vol. 27, no. 4, pp. 1689–1696, Apr. 2012.
particle swarm optimization (PSO)-based MPPT for PV with reduced [76] M. H. Etesami, N. Farokhnia, and S. Hamid Fathi, “Colonial compet-
steady-state oscillation,” IEEE Trans. Power Electron., vol. 27, no. 8, itive algorithm development toward harmonic minimization in multi-
pp. 3627–3638, Aug. 2012. level inverters,” IEEE Trans. Ind. Informat., vol. 11, no. 2, pp. 459–466,
[56] H. Taghizadeh and M. T. Hagh, “Harmonic elimination of cascade Apr. 2015.
multilevel inverters with nonequal DC sources using particle swarm [77] K. Haghdar, “Optimal DC source influence on selective harmonic
optimization,” IEEE Trans. Ind. Electron., vol. 57, no. 11, pp. 3678– elimination in multilevel inverters using teaching-learning-based op-
3684, Nov. 2010. timization,” IEEE Trans. Ind. Electron., vol. 67, no. 2, pp. 942–949,
[57] W. Wang, A. C. Liu, H. S. Chung, R. W. Lau, J. Zhang, and A. W. Lo, Feb. 2020.
“Fault diagnosis of photovoltaic panels using dynamic current-voltage [78] J. Zhang, Z. H. Zhan, Y. Lin, N. Chen, Y. J. Gong, J. H. Zhong, H. S. H.
characteristics,” IEEE Trans. Power Electron., vol. 31, no. 2, pp. 1588– Chung, Y. Li, and Y. H. Shi, “Evolutionary computation meets machine
1599, Feb. 2016. learning: A survey,” IEEE Computational Intell. Mag., vol. 6, no. 4,
[58] Z. H. Liu, H. L. Wei, Q. C. Zhong, K. Liu, X. S. Xiao, and L. H. Wu, pp. 68–75, Nov. 2011.
“Parameter estimation for VSI-fed PMSM based on a dynamic PSO
[79] D. Chiozzi, M. Bernardoni, N. Delmonte, and P. Cova, “A neural
with learning strategies,” IEEE Trans. Power Electron., vol. 32, no. 4,
network based approach to simulate electrothermal device interaction
pp. 3154–3165, Apr. 2017.
in SPICE environment,” IEEE Trans. Power Electron., vol. 34, no. 5,
[59] M. Liserre, A. Dell’Aquila, and F. Blaabjerg, “Genetic algorithm-
pp. 4703–4710, May 2019.
based design of the active damping for an LCL-filter three-phase active
[80] T. Dragicevic, P. Wheeler, and F. Blaabjerg, “Artificial intelligence
rectifier,” IEEE Trans. Power Electron., vol. 19, no. 1, pp. 76–86, Jan.
aided automated design for reliability of power electronic systems,”
2004.
IEEE Trans. Power Electron., vol. 34, no. 8, pp. 7161–7171, Aug.
[60] B. Ji, X. G. Song, E. Sciberras, W. P. Cao, Y. H. Hu, and V. Pickert,
2019.
“Multiobjective design optimization of IGBT power modules consider-
[81] M. G. Simoes and B. K. Bose, “Neural-network-based estimation of
ing power cycling and thermal cycling,” IEEE Trans. Power Electron.,
feedback signals for a vector controlled induction-motor drive,” IEEE
vol. 30, no. 5, pp. 2493–2504, May 2015.
Trans. Ind. Appl., vol. 31, no. 3, pp. 620–629, May/Jun. 1995.
[61] F. Wang, G. Chen, D. Boroyevich, S. Ragon, M. Arpilliere, and
V. R. Stefanovic, “Analysis and design optimization of diode front- [82] J. Zhao and B. K. Bose, “Neural-network-based waveform processing
end rectifier passive components for voltage source inverters,” IEEE and delayless filtering in power electronics and AC drives,” IEEE Trans.
Trans. Power Electron., vol. 23, no. 5, pp. 2278–2289, Sep. 2008. Ind. Electron., vol. 51, no. 5, pp. 981–991, Oct. 2004.
[62] M. D’Antonio, C. Shi, B. Wu, and A. Khaligh, “Design and optimiza- [83] S. K. Mondal, J. O. P. Pinto, and B. K. Bose, “A neural-network-based
tion of a solar power conversion system for space applications,” IEEE space-vector PWM controller for a three-level voltage-fed inverter
Trans. Ind. Appl., vol. 55, no. 3, pp. 2310–2319, May/Jun. 2019. induction motor drive,” IEEE Trans. Ind. Appl., vol. 38, no. 3, pp.
[63] B. Chen, X. Liang, and N. Wan, “Design methodology for inductor- 660–669, May/Jun. 2002.
integrated Litz-wired high-power medium-frequency transformer with [84] C. M. Lin, K. N. Hung, and C. F. Hsu, “Adaptive neuro-wavelet control
the nanocrystalline core material for isolated DC-link stage of solid- for switching power supplies,” IEEE Trans. Power Electron., vol. 22,
state transformer,” IEEE Trans. Power Electron., 2020, Early access. no. 1, pp. 87–95, Jan. 2007.
[64] L. G. Junior, J. O. P. Pinto, J. A. B. Filho, and G. Lambert-Torres, [85] M. Novak and T. Dragicevic, “Supervised imitation learning of finite set
“Recursive least square and genetic algorithm based tool for PID model predictive control systems for power electronics,” IEEE Trans.
controllers tuning,” in Int. Conf. Intelligent Sys. Appl. to Power Syst., Ind. Electron., 2020, Early access.
2007, pp. 1–6. [86] H. Soliman, P. Davari, H. Wang, and F. Blaabjerg, “Capacitance esti-
[65] M. J. Schutten and D. A. Torrey, “Genetic algorithms for control of mation algorithm based on DC-link voltage harmonics using artificial
power converters,” in 26th Annu. IEEE Power Electron. Specialists neural network in three-phase motor drive systems,” in IEEE Energy
Conf., Vols I and II, 1995, pp. 1321–1326. Conversion Congr. Expo., 2017, pp. 5795–5802.
[66] M. S. A. Dahidah and V. G. Agelidis, “Selective harmonic elimination [87] H. Soliman, H. Wang, B. Gadalla, and F. Blaabjerg, “Condition
PWM control for cascaded multilevel voltage source converters: A monitoring for DC-link capacitors based on artificial neural network
generalized formula,” IEEE Trans. Power Electron., vol. 23, no. 4, algorithm,” in IEEE 5th Int. Conf. Power Eng., Energy and Electr.
pp. 1620–1630, Jul. 2008. Drives, 2015, pp. 587–591.
[67] M. J. Neath, A. K. Swain, U. K. Madawala, and D. J. Thrimawithana, [88] H. Soliman, I. Abdelsalam, H. Wang, and F. Blaabjerg, “Artificial
“An optimal PID controller for a bidirectional inductive power transfer neural network based DC-link capacitance estimation in a diode-bridge
system using multiobjective genetic algorithm,” IEEE Trans. Power front-end inverter system,” in IEEE 3rd Int. Future Energy Electron.
Electron., vol. 29, no. 3, pp. 1523–1531, Mar. 2014. Conf. ECCE Asia, 2017, pp. 196–201.
[68] A. Mehrizi-Sani and S. Filizadeh, “An optimized space vector modula- [89] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “Multilevel feature moving
tion sequence for improved harmonic performance,” IEEE Trans. Ind. average ratio method for fault diagnosis of the microgrid inverter
Electron., vol. 56, no. 8, pp. 2894–2903, Aug. 2009. switch,” IEEE-CAA J. of Automatica Sinica, vol. 4, no. 2, pp. 177–
[69] M. Garaj, K. Y. Hong, H. S.-H. Chung, J. Zhou, and A. W.-L. Lo, 185, Apr. 2017.
“Photovoltaic panel health diagnostic system for solar power plants,” [90] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “A diagnosis algorithm
in IEEE Energy Conversion Congr. Expo., 2018, pp. 1–6. for multiple open-circuited faults of microgrid inverters based on main

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
25

fault component analysis,” IEEE Trans. Energy Conversion, vol. 33, [111] Z. Li, Z. Zheng, and R. Outbib, “A prognostic methodology for power
no. 3, pp. 925–937, Sep. 2018. MOSFETs under thermal stress using echo state network and particle
[91] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “Multiple open-circuit filter,” Microelectronics Rel., vol. 88-90, pp. 350–354, Sep. 2018.
fault diagnosis based on multistate data processing and subsection [112] Z. J. Huang, Z. S. Wang, X. S. Yao, and H. G. Zhang, “Multi-switches
fluctuation analysis for photovoltaic inverter,” IEEE Trans. Instrum. fault diagnosis based on small low-frequency data for voltage-source
Meas., vol. 67, no. 3, pp. 516–526, Mar. 2018. inverters of PMSM drives,” IEEE Trans. Power Electron., vol. 34, no. 7,
[92] S. Qiang and Y. Li, “Motor inverter fault diagnosis using wavelets neu- pp. 6845–6857, Jul. 2019.
ral networks,” in 2013 IEEE Int. Conf. on Syst., Man, and Cybernetics, [113] Z. Y. Xue, K. S. Xiahou, M. S. Li, T. Y. Ji, and Q. H. Wu, “Diagnosis
2013, pp. 3168–3173. of multiple open-circuit switch faults based on long short-term memory
[93] S. Mohagheghi, R. G. Harley, T. G. Habetler, and D. Divan, “Con- network for DFIG-based wind turbine systems,” IEEE J. Emerg. Sel.
dition monitoring of power electronic circuits using artificial neural Topics Power Electron., 2019, Early access.
networks,” IEEE Trans. Power Electron., vol. 24, no. 10, pp. 2363– [114] A. Alghassi, S. Perinpanayagam, and M. Samie, “Stochastic RUL
2367, Oct. 2009. calculation enhanced with TDNN-based IGBT failure modeling,” IEEE
[94] S. Khomfoi and L. M. Tolbert, “Fault diagnostic system for a multilevel Trans. Rel., vol. 65, no. 2, pp. 558–573, Jun. 2016.
inverter using a neural network,” IEEE Trans. Power Electron., vol. 22, [115] S. Kiranyaz, A. Gastli, L. Ben-Brahim, N. Al-Emadi, and M. Gabbouj,
no. 3, pp. 1062–1069, May 2007. “Real-time fault detection and identification for MMC using 1-D
[95] Y. Cui, J. Shi, and Z. Wang, “Quantum assimilation-based state-of- convolutional neural networks,” IEEE Trans. Ind. Electron., vol. 66,
health assessment and remaining useful life estimation for electronic no. 11, pp. 8760–8771, Nov. 2019.
systems,” IEEE Trans. Ind. Electron., vol. 63, no. 4, pp. 2379–2390, [116] C. Y. Yin, H. Lu, M. Musallam, C. Bailey, and C. M. Johnson, “A
Apr. 2016. prognostic assessment method for power electronics modules,” in 2nd
[96] P. Tamilselvan, P. F. Wang, and M. Pecht, “A multi-attribute classifi- Electron. Syst.-Integration Technol. Conf., 2008, pp. 1353–1358.
cation fusion system for insulated gate bipolar transistor diagnostics,” [117] B. Cai, Y. Zhao, H. Liu, and M. Xie, “A data-driven fault diagnosis
Microelectronics Rel., vol. 53, no. 8, pp. 1117–1129, Aug. 2013. methodology in three-phase inverters for PMSM drive systems,” IEEE
[97] Z. S. Wang, Z. J. Huang, C. H. Song, and H. G. Zhang, “Multiscale Trans. Power Electron., vol. 32, no. 7, pp. 5590–5600, Jul. 2017.
adaptive fault diagnosis based on signal symmetry reconstitution pre- [118] S. S. Moosavi, A. Kazemi, and H. Akbari, “A comparison of various
processing for microgrid inverter under changing load condition,” IEEE open-circuit fault detection methods in the IGBT-based DC/AC inverter
Trans. Smart Grid, vol. 9, no. 2, pp. 797–806, Mar. 2018. used in electric vehicle,” Eng. Failure Analysis, vol. 96, pp. 223–235,
[98] D. Diallo, M. E. H. Benbouzid, D. Hamad, and X. Pierre, “Fault Feb. 2019.
detection and diagnosis in an induction machine drive: A pattern [119] S. H. Ali, M. Heydarzadeh, S. Dusmez, X. Li, A. S. Kamath, and
recognition approach based on concordia stator mean current vector,” B. Akin, “Lifetime estimation of discrete IGBT devices based on
IEEE Trans. Energy Conversion, vol. 20, no. 3, pp. 512–519, Sep. gaussian process,” IEEE Trans. Ind. Appl., vol. 54, no. 1, pp. 395–
2005. 403, Jan./Feb. 2018.
[99] C. S. Chen, “TSK-Type self-organizing recurrent-neural-fuzzy control [120] J. R. Celaya, A. Saxena, S. Saha, and K. Goebel, “Prognostics of power
of linear microstepping motor drives,” IEEE Trans. Power Electron., MOSFETs under thermal stress accelerated aging using data-driven
vol. 25, no. 9, pp. 2253–2265, Sep. 2010. and model-based methodologies,” in Annu. Conf.Prognostics Health
[100] R. J. Wai and L. C. Shih, “Adaptive fuzzy-neural-network design for Manage. Soc., 2011, pp. 1–10.
voltage tracking control of a DC-DC boost converter,” IEEE Trans.
[121] A. Lidozzi, L. Solero, F. Crescimbini, and A. Di Napoli, “SVM PMSM
Power Electron., vol. 27, no. 4, pp. 2104–2115, Apr. 2012.
drive with low resolution hall-effect sensors,” IEEE Trans. Power
[101] P. Z. Grabowski, M. P. Kazmierkowski, B. K. Bose, and F. Blaabjerg,
Electron., vol. 22, no. 1, pp. 282–290, Jan. 2007.
“A simple direct-torque neuro-fuzzy control of PWM-inverter-fed in-
[122] N. A. Ahmed and A. K. Al-Othman, “Photovoltaic system with voltage-
duction motor drive,” IEEE Trans. Ind. Electron., vol. 47, no. 4, pp.
based maximum power point tracking using support vector machine,”
863–870, Aug. 2000.
in Proc. 5th IEEE Conf. Ind. Electron. Appl., Vol 4, 2010, pp. 539–544.
[102] T. Kamel, Y. Biletskiy, and L. Chang, “Capacitor aging detection for the
DC filters in the power electronic converters using ANFIS algorithm,” [123] F. Mei, N. Liu, H. Y. Miao, Y. Pan, H. Y. Sha, and J. Y. Zheng, “On-
in IEEE 28th Canadian Conf. Elect. Computer Eng., 2015, pp. 663– line fault diagnosis model for locomotive traction inverter based on
668. wavelet transform and support vector machine,” Microelectronics Rel.,
[103] A. Soualhi, M. Makdessi, R. German, F. R. Echeverria, H. Razik, vol. 88-90, pp. 1274–1280, Sep. 2018.
A. Sari, P. Venet, and G. Clerc, “Heath monitoring of capacitors and [124] C. Delpha, H. Chen, and D. Diallo, “SVM based diagnosis of inverter
supercapacitors using the neo-fuzzy neural approach,” IEEE Trans. Ind. fed induction machine drive: A new challenge,” in 38th Ann. Conf.
Informat., vol. 14, no. 1, pp. 24–34, Jan. 2018. IEEE Ind. Electron. Soc., 2012, pp. 3931–3936.
[104] C. Schauder, “Adaptive speed identification for vector control of [125] X. X. Zheng and P. Peng, “Fault diagnosis of wind power converters
induction-motors without rotational transducers,” IEEE Trans. Ind. based on compressed sensing theory and weight constrained AdaBoost-
Appl., vol. 28, no. 5, pp. 1054–1061, Sep./Oct. 1992. SVM,” J. Power Electron., vol. 19, no. 2, pp. 443–453, Mar. 2019.
[105] L. E. B. da Silva, B. K. Bose, and J. O. P. Pinto, “Recurrent-neural- [126] A. G. Abo-Khalil and D. Lee, “DC-Link capacitance estimation in
network-based implementation of a programmable cascaded low-pass AC/DC/AC PWM converters using voltage injection,” IEEE Trans. Ind.
filter used in stator flux synthesis of vector-controlled induction motor Appl., vol. 44, no. 5, pp. 1631–1637, Sep./Oct. 2008.
drive,” IEEE Trans. Ind. Electron., vol. 46, no. 3, pp. 662–665, Jun. [127] T. Z. Wang, H. Xu, J. G. Han, E. Elbouchikhi, and M. E. H. Benbouzid,
1999. “Cascaded H-Bridge multilevel inverter system fault diagnosis using a
[106] X. G. Fu and S. H. Li, “Control of single-phase grid-connected convert- PCA and multiclass relevance vector machine approach,” IEEE Trans.
ers with LCL filters using recurrent neural network and conventional Power Electron., vol. 30, no. 12, pp. 7006–7018, Dec. 2015.
control methods,” IEEE Trans. Power Electron., vol. 31, no. 7, pp. [128] S. Q. Zhou, L. W. Zhou, and P. J. Sun, “Monitoring potential defects in
5354–5364, Jul. 2016. an IGBT module based on dynamic changes of the gate current,” IEEE
[107] J. O. P. Pinto, B. K. Bose, and L. E. B. da Silva, “A stator-flux-oriented Trans. Power Electron., vol. 28, no. 3, pp. 1479–1487, Mar. 2013.
vector-controlled induction motor drive with space-vector PWM and [129] J. Zhang, H. S. Chung, and W. Lo, “Clustering-based adaptive
flux-vector synthesis by neural networks,” IEEE Trans. Ind. Appl., crossover and mutation probabilities for genetic algorithms,” IEEE
vol. 37, no. 5, pp. 1308–1318, Sep./Oct. 2001. Trans. on Evolutionary Computation, vol. 11, no. 3, pp. 326–335, 2007.
[108] P. Xiao, G. K. Venayagamoorthy, K. A. Corzine, and J. Huang, [130] E. G. Strangas, S. Aviyente, and S. S. H. Zaidi, “Time-frequency
“Recurrent neural networks based impedance measurement technique analysis for efficient fault diagnosis and failure prognosis for interior
for power electronic systems,” IEEE Trans. Power Electron., vol. 25, permanent-magnet AC motors,” IEEE Trans. Ind. Electron., vol. 55,
no. 2, pp. 382–390, Feb. 2010. no. 12, pp. 4191–4199, Dec. 2008.
[109] Y. Zhang, Z. Wang, H. Wang, and F. Blaabjerg, “Artificial intelligence- [131] S. Zhao, V. Makis, S. Chen, and Y. Li, “Health assessment method
aided thermal model considering cross-coupling effects,” IEEE Trans. for electronic components subject to condition monitoring and hard
Power Electron., 2020, Early access. failure,” IEEE Trans. Instrum. Meas., vol. 68, no. 1, pp. 138–150, Jan.
[110] M. R. Habibi, H. R. Baghaee, T. Dragicevic, and F. Blaabjerg, “De- 2019.
tection of false data injection cyber-attacks in DC microgrids based on [132] Y. Z. Lu and A. Christou, “Prognostics of IGBT modules based on
recurrent neural networks,” IEEE J. Emerging Selected Topics Power the approach of particle filtering,” Microelectronics Rel., vol. 92, pp.
Electron., 2020, Early access. 96–105, Jan. 2019.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
26

[133] M. Rigamonti, P. Baraldi, A. Alessi, E. Zio, D. Astigarraga, and [155] S. Zhao, V. Makis, S. Chen, and Y. Li, “Health evaluation method for
A. Galarza, “An ensemble of component-based and population-based degrading systems subject to dependent competing risks,” J. Syst. Eng.
self-organizing maps for the identification of the degradation state of Electron., vol. 29, no. 2, pp. 436–444, Apr. 2018.
insulated-gate bipolar transistors,” IEEE Trans. Rel., vol. 67, no. 3, pp. [156] IEEE, “IEEE standard framework for prognostics and health manage-
1304–1313, Sep. 2018. ment of electronic systems,” IEEE Std 1856-2017, pp. 1–31, 2017.
[134] N. Femia, G. Spagnuolo, and V. Tucci, “State-space models and order [157] S. Yang, D. Xiang, A. Bryant, P. Mawby, L. Ran, and P. Tavner, “Con-
reduction for dc-dc switching converters in discontinuous modes,” dition monitoring for device reliability in power electronic converters:
IEEE Trans. Power Electron., vol. 10, no. 6, pp. 640–650, 1995. A review,” IEEE Trans. Power Electron., vol. 25, no. 11, pp. 2734–
[135] J. F. Martins, V. F. Pires, C. Lima, and A. J. Pires, “Fault detection 2752, Nov. 2010.
and diagnosis of grid-connected power inverters using PCA and current [158] H. Oh, B. Han, P. McCluskey, C. Han, and B. D. Youn, “Physics-of-
mean value,” in 38th Annu. Conf. IEEE Ind. Electron. Soc., 2012, pp. failure, condition monitoring, and prognostics of insulated gate bipolar
5185–5190. transistor modules: A review,” IEEE Trans. Power Electron., vol. 30,
[136] W. Chen, L. Zhang, K. Pattipati, A. M. Bazzi, S. Joshi, and E. M. no. 5, pp. 2413–2426, May 2015.
Dede, “Data-driven approach for fault prognosis of SiC MOSFETs,” [159] M. Al-Greer, M. Armstrong, M. Ahmeid, and D. Giaouris, “Advances
IEEE Trans. Power Electron., vol. 35, no. 4, pp. 4048–4062, Apr. 2020. on system identification techniques for DC-DC switch mode power
[137] J. Han, M. Kamber, and J. Pei, Data Mining: concepts and techniques, converter applications,” IEEE Trans. Power Electron., vol. 34, no. 7,
3rd ed. The Morgan Kaufmann, 2012. pp. 6973–6990, Jul. 2019.
[138] M. Glavic, R. Fonteneau, and D. Ernst, “Reinforcement learning for [160] S. Zhao, S. Chen, F. Yang, E. Ugur, B. Akin, and H. Wang, “A
electric power system decision and control: Past considerations and composite failure precursor for condition monitoring and remaining
perspectives,” IFAC Papersonline, vol. 50, no. 1, pp. 6918–6927, Jul. useful life prediction of discrete power devices,” IEEE Trans. Ind.
2017. Informat., 2020, Early access.
[139] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. [161] Matlab, “Diagnostic Feature Designer,” https://www.mathworks.com/
MIT press, 2018. help/predmaint/gs/explore-features-in-diagnostic-feature-
[140] P. Kofinas, S. Doltsinis, A. I. Dounis, and G. A. Vouros, “A rein- designer.html, 2020.
forcement learning approach for MPPT control method of photovoltaic [162] M. A. F. Pimentel, D. A. Clifton, L. Clifton, and L. Tarassenko, “A
sources,” Renewable Energy, vol. 108, pp. 461–473, Aug. 2017. review of novelty detection,” Signal Process., vol. 99, pp. 215–249,
[141] A. S. Elwer, S. A. Wahsh, M. O. Khalil, and A. M. Nur-Eldeen, Jun. 2014.
“Intelligent fuzzy controller using particle swarm optimization for [163] R. Liu, B. Yang, E. Zio, and X. Chen, “Artificial intelligence for
control of permanent magnet synchronous motor for electric vehicle,” fault diagnosis of rotating machinery: A review,” Mech. Syst. Signal
in Proc. 29th Annu. Conf. IEEE Ind. Electron. Soc., Vols 1 - 3,, 2003, Process., vol. 108, pp. 33 – 47, Aug. 2018.
pp. 1762–1766. [164] Matlab, “Predictive Maintenance Toolbox,” https://
[142] F. Harashima, Y. Demizu, S. Kondo, and H. Hashimoto, “Application www.mathworks.com/products/predictive-maintenance.html, 2020.
of neutral networks to power converter control,” in Conf. Record IEEE [165] A. Hanif, Y. Yu, D. DeVoto, and F. Khan, “A comprehensive review
Industry Appl. Soc. Annu. Meeting,, 1989, pp. 1086–1091 vol.1. toward the state-of-the-art in failure and lifetime predictions of power
[143] S. Khomfoi and L. M. Tolbert, “Fault diagnosis and reconfiguration electronic devices,” IEEE Trans. Power Electron., vol. 34, no. 5, pp.
for multilevel inverter drive using AI-based techniques,” IEEE Trans. 4729–4746, May 2019.
Ind. Electron., vol. 54, no. 6, pp. 2954–2968, Dec. 2007. [166] M. A. Eleffendi and C. M. Johnson, “In-service diagnostics for wire-
[144] H. Zhang, J. Zhao, R. Wang, and T. Ma, “Multi-objective reinforcement bond lift-off and solder fatigue of power semiconductor packages,”
learning algorithm and its application in drive system,” in Proc. 34th IEEE Trans. Power Electron., vol. 32, no. 9, pp. 7187–7198, Sep. 2017.
Annu. Conf. IEEE Ind. Electron. Soc., Vols 1-5,, 2008, pp. 225–230. [167] M.-F. Ng, J. Zhao, Q. Yan, G. J. Conduit, and Z. W. Seh, “Predicting
[145] G. Bramerdorfer, J. A. Tapia, J. J. Pyrhonen, and A. Cavagnino, the state of charge and health of batteries using data-driven machine
“Modern electrical machine design optimization: Techniques, trends, learning,” Nature Machine Intelligence, vol. 2, no. 3, pp. 161–170, Mar.
and best practices,” IEEE Trans. Ind. Electron., vol. 65, no. 10, pp. 2020.
7672–7684, Oct. 2018. [168] X.-S. Si, W. Wang, C.-H. Hu, and D.-H. Zhou, “Remaining useful life
[146] C. Versele, O. Deblecker, and J. Lobry, “Multiobjective optimal choice estimation - a review on the statistical data driven approaches,” Eur. J.
and design of isolated dc-dc power converters,” in Proc. 14th Eur. Conf. Oper. Res., vol. 213, no. 1, pp. 1–14, Aug. 2011.
Power Electron. Appl., 2011, pp. 1–10. [169] Alpha and Omega Semiconductor, “Power semiconductor reliabil-
[147] S. Vighetti, J. Ferrieux, and Y. Lembeye, “Optimization and design of ity handbook,” http://www.aosmd.com/media/reliability-handbook.pdf,
a cascaded DC/DC converter devoted to grid-connected photovoltaic Sunnyvale, CA, USA, pp. 8–10, May 2010.
systems,” IEEE Trans. Power Electron., vol. 27, no. 4, pp. 2018–2027, [170] R. Zhao, R. Q. Yan, Z. H. Chen, K. Z. Mao, P. Wang, and R. X. Gao,
2012. “Deep learning and its applications to machine health monitoring,”
[148] H. Helali, D. Bergogne, H. Morel, and J. B. H. Slama, “Power Mech. Syst. Signal Process., vol. 115, pp. 213–237, Jan. 2019.
converter design methodology: Uses of multiple objective techniques [171] Q. Wang, G. Michau, and O. Fink, “Domain adaptive transfer learning
for optimization of a (42/14V) buck converter,” in 4th Int. Conf. for fault diagnosis,” in Prognostics Syst. Health Manage. Conf., 2019,
Integrated Power Syst., 2006, pp. 1–5. pp. 1–7.
[149] H. Zhang, F. Mollet, C. Saudemont, and B. Robyns, “Experimental [172] L. Liao, “Discovering prognostic features using genetic programming
validation of energy storage system management strategies for a local in remaining useful life prediction,” IEEE Trans. Ind. Electron., vol. 61,
DC distribution system of more electric aircraft,” IEEE Trans. Ind. no. 5, pp. 2464–2472, May 2014.
Electron., vol. 57, no. 12, pp. 3905–3916, Dec. 2010. [173] European Union, “General Data Protection Regulation (GDPR),” https:
[150] Y. F. Yin, J. X. Liu, J. A. Sanchez, L. G. Wu, S. Vazquez, J. I. Leon, //gdpr.eu/, 2018.
and L. G. Franquelo, “Observer-based adaptive sliding mode control [174] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning:
of NPC converters: An RBF neural network approach,” IEEE Trans. Concept and applications,” ACM Trans. Intell. Syst. Technol., vol. 10,
Power Electron., vol. 34, no. 4, pp. 3831–3841, Apr. 2019. no. 2, Jan. 2019.
[151] F. J. T. Filho, L. M. Tolbert, and B. Ozpineci, “Real time selective
harmonic minimization for multilevel inverters using genetic algorithm
and artificial neural network angle generation,” in Proc. 7th Int. Power
Electron. Motion Control Conf., vol. 2, 2012, pp. 895–899.
[152] K. H. Tan, “Squirrel-cage induction generator system using wavelet
petri fuzzy neural network control for wind power applications,” IEEE
Trans. Power Electron., vol. 31, no. 7, pp. 5242–5254, Jul. 2016.
[153] J. S. R. Jang, “ANFIS: adaptive-network-based fuzzy inference sys-
tem,” IEEE Trans. Syst., Man, and Cybernetics, vol. 23, no. 3, pp.
665–685, May/Jun. 1993.
[154] S. Zhao, V. Makis, S. Chen, and Y. Li, “Evaluation of reliability
function and mean residual life for degrading systems subject to
condition monitoring and random failure,” IEEE Trans. Rel., vol. 67,
no. 1, pp. 13–25, Mar. 2018.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.3024914, IEEE
Transactions on Power Electronics
27

Shuai Zhao (S’14-M’18) received the B.E. (Hons),

M.E., and Ph.D. degrees in information and com-
munication engineering from Northwestern Poly-
technical University, Xi’an, China, in 2011, 2014,
and 2018, respectively. He is currently a postdoc-
toral researcher with the Center of Reliable Power
Electronics (CORPE), Department of Energy Tech-
nology, Aalborg University, Denmark. From Sep.
2014 to Sep. 2016, he was a visiting Ph.D. Student
with the Department of Mechanical and Industrial
Engineering at the University of Toronto, Toronto,
ON, Canada, with the scholarship from China Scholarship Council (CSC).
In August 2018, he was a visiting scholar with the Power Electronics and
Drives Laboratory, Department of Electrical and Computer Science at the
University of Texas at Dallas, Richardson, TX, USA. His research interests
include system informatics, intelligent condition monitoring, diagnostics &
prognostics, and tailored AI tools for power electronic systems.

Frede Blaabjerg (S’86-M’88-SM’97-F’03) was

with ABB-Scandia, Randers, Denmark, from 1987
to 1988. From 1988 to 1992, he got the PhD degree
in Electrical Engineering at Aalborg University in
1995. He became an Assistant Professor in 1992, an
Associate Professor in 1996, and a Full Professor of
power electronics and drives in 1998. From 2017
he became a Villum Investigator. He is honoris
causa at University Politehnica Timisoara (UPT),
Romania and Tallinn Technical University (TTU) in
Estonia.
His current research interests include power electronics and its applications
such as in wind turbines, PV systems, reliability, harmonics and adjustable
speed drives. He has published more than 600 journal papers in the fields of
power electronics and its applications. He is the co-author of four monographs
and editor of ten books in power electronics and its applications.
He has received 32 IEEE Prize Paper Awards, the IEEE PELS Distinguished
Service Award in 2009, the EPE-PEMC Council Award in 2010, the IEEE
William E. Newell Power Electronics Award 2014, the Villum Kann Ras-
mussen Research Award 2014, the Global Energy Prize in 2019, and the 2020
IEEE Edison Medal. He was the Editor-in-Chief of the IEEE Transactions on
Power Electronics from 2006 to 2012. He has been Distinguished Lecturer
for the IEEE Power Electronics Society from 2005 to 2007 and for the IEEE
Industry Applications Society from 2010 to 2011 as well as 2017 to 2018.
In 2019-2020 he serves a President of IEEE Power Electronics Society. He
is Vice-President of the Danish Academy of Technical Sciences too. He is
nominated in 2014-2019 by Thomson Reuters to be between the most 250
cited researchers in Engineering in the world.

Huai Wang (M’12-SM’17) received the B.E. degree

in electrical engineering, from Huazhong University
of Science and Technology, Wuhan, China, in 2007
and the Ph.D. degree in power electronics, from the
City University of Hong Kong, Hong Kong, in 2012.
He is currently Professor with the Center of Re-
liable Power Electronics (CORPE), Department of
Energy Technology at Aalborg University, Denmark.
He was a Visiting Scientist with the ETH Zurich,
Switzerland, from Aug. to Sep. 2014, and with the
Massachusetts Institute of Technology (MIT), USA,
from Sep. to Nov. 2013. He was with the ABB Corporate Research Center,
Switzerland, in 2009. His research addresses the fundamental challenges in
modelling and validation of power electronic component failure mechanisms,
and application issues in system-level predictability, condition monitoring,
circuit architecture, and robustness design. He leads a project on Light-AI for
Cognitive Power Electronics.
Dr. Wang received the Richard M. Bass Outstanding Young Power Elec-
tronics Engineer Award from the IEEE Power Electronics Society in 2016, and
the Green Talents Award from the German Federal Ministry of Education and
Research in 2014. He is currently the Chair of IEEE PELS/IAS/IES Chapter
in Denmark. He serves as an Associate Editor of IET Electronics Letters,
IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER
ELECTRONICS, and IEEE TRANSACTIONS ON POWER ELECTRONICS.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.