The Electricity Journal 34 (2021) 106881

Contents lists available at ScienceDirect

The Electricity Journal

journal homepage: www.elsevier.com/locate/tej

Machine learning for power system protection and control

Hanyu Yang a, Xubin Liu b, Di Zhang a, Tao Chen a, Canbing Li c, *, Wentao Huang c
a
College of Electrical and Information Engineering, Hunan University, Changsha, 410082, China
b
School of Electrical and Electronic Engineering, Huazhong University of Science and Technology, Wuhan, 430073, China
c
School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China

A R T I C L E I N F O A B S T R A C T

Keywords: Since the power system is undergoing a transition into a more flexible and complex system, it urges improve­
Fault diagnosis ments in fault diagnosis techniques for the power system protection to avoid cascading damages at the occur­
Machine learning rence of faults. Facing with challenges of massive data, several machine-learning based methods for identifying
Deep learning
faults were proposed over the past years. In this paper, an overview of conventional and trending machine
Power system
learning applications for the fault diagnosis are summarized.

1. Introduction caused by material properties, structural characteristics and so forth,

which cannot be easily captured in the modern complex system.
With the increasing penetration of renewable energy resources and In contrast to the conventional power system, the modern power
electronic devices, the scale and complexity of the power system is system is currently challenged by the growing volumes of data of various
growing in all sectors, e.g., generation, transmission, distribution sys­ nature, the urging need of data storage, the ever-increasing penetration
tems. Power system protection, as a branch of power engineering, is of distributed generations and technological advancements. There is no
mainly about the protective scheme of power system after being suffered doubt that the modern power system calls for a more reliable and flex­
from various faults. Types of faults can be divided into open-circuit ible power system protection and control. Therefore, a huge number of
faults and short-circuit faults, e.g., symmetric and unsymmetric faults, monitoring devices (e.g., advanced metering infrastructure, AMI, su­
caused by multiple reasons, such as, equipment failures, abnormal pervisor control and data acquisition networks, SCADA, wide area
conditions, human errors, and environmental conditions, resulting in measurement system, WAMS), capable of exchanging information are
severe economic losses and blackouts (Lim and Dorr, 2000). Fig. 1 il­ connected to their respective networks for guaranteeing the system
lustrates the potential faults in the all sectors in the power system. Once operation (Ibrahim et al., 2020). These data tend to be heterogeneous,
the fault occurs, it is required to timely identify and perform a wide multi-modal, decentralized, and highly nonlinear, imposing significant
range of protective measures to clear or insulate it. Hence, an early and challenges to conventional model-driven methods. Inevitably, those
accurate detection, location and classification of faults can increase conventional techniques used in power systems are no longer adequate
safety and reliability and reduce downtime and interruption time. in dealing with the aforementioned issues because of their limitations in
Fault diagnosis, including location, detection and classification, is analyzing the generalizability of conventional models, storing the large
the primary step of the power system protection and control. It is volume of data, the adequacy and efficiency of real-time processing.
generally conducted based on the massive data connected from the Data-driven methods designed for acquiring hidden knowledge from
monitoring devices. Over the past years, power utilities have explored empirical data gradually attract the researchers’ interest, where ma­
various conventional techniques for fault diagnosis. One of the most chine learning techniques provides a promising way to meet the chal­
common-used technique is visual inspection which can be considered as lenges brought by big data. Machine learning is first proposed since
model driven method as well. Although the advanced signal processing 1956 and gradually considered as a suitable tool for fault diagnosis, due
methods can help figure out faults types or location in the power system, to its ability to scale to large systems with low computational cost
these methods highly rely on the specialized knowledge and the accu­ (Zhang et al., 2019a). Machine learning is the study of computer engi­
racy of the measurement data. Also, they require knowing the basic neering that use computational algorithms to “learn” information from
principles of the faults with the investigation of failure mechanisms data to formulate a model, known as “training data”, and when they

* Corresponding author.
E-mail address: [email protected] (C. Li).

https://doi.org/10.1016/j.tej.2020.106881

Available online 2 December 2020

1040-6190/© 2020 Elsevier Inc. All rights reserved.
H. Yang et al. The Electricity Journal 34 (2021) 106881

carry out tasks like humans or animals they can adaptively performance redundant information and may heavy the computational burden, thus
as the data availability increases. Machine learning techniques can be the feature selection methods have to adopted to select sensitive features
mainly classified into three types: supervised, unsupervised and rein­ reflecting the status of power system from the collected data. Based on
forcement learnings. the sensitive features obtained, the machine learning-based diagnosis
According to the development history of the machine learning, the models establish the relationship between the selected sensitive features
conventional machine learning techniques are usually supervised and the outputs that indicate the health states of equipment, which can
learning, including the expert system, back propagation neural network, be defined as the “learning” process. In step three, the diagnosis models
Bayesian network, support vector machine and so forth (Lei et al., are further trained with unlabeled or labeled data depending on the
2020a). With the general recognition that the conventional techniques availability of the selected faulty data to realize the fault diagnosis.
can no longer efficiently and accurately deal with the vast amount of Finally, based on the expected results of the fault diagnosis, the corre­
data, trending machine learning techniques have attracted more and sponding protective scheme will perform to disconnect faulted parts to
more researchers’ interest over the past years. A variety of machine protect the rest of the electrical network.
learning techniques, especially deep learning, transfer learning, unsu­
pervised learning methods are proposed to address the challenges in 3. Conventional machine learning based fault diagnosis
various application domains of the modern power system.
In this paper, we focus on the applications of machine learning to the 3.1. Artificial neutral network-based approaches
fault diagnosis, which is the basic step of the power system protection.
we provide a brief summary of various conventional and trending ma­ Artificial neutral network-based approach (ANN), as a supervised
chine learning techniques in the fault diagnosis, including their advan­ machine learning method, imitates the information processing activities
tages and applications. The remainder of this paper is organized as of human brains, to identify the faults with multiple sets of data inputs.
follows. Section 2 provides a general framework of machine learning- As indicated in Fig. 3, the training dataset are the inputs, which will be
based fault diagnosis. Section 3 summarizes the categories of the con­ further processed by multi-hidden layers and finally mapped into the
ventional machine learning techniques as well as the challenges of the results at the output layer. ANN is widely applied in various fault
big data brings are concluded in Section 3, while Section 4 provides a diagnosis researches. In (Purushothama et al., 2001; Hagh et al., 2007;
brief introduction of the trending machine learning techniques and Javadian and Massaeli, 2011; Aslan, 2012), various ANN-based methods
Section 5 concludes. are applied in fault identification in the distribution system for esti­
mating fault distance, detecting high impedance fault and identifying
2. General framework for machine learning based fault the fault types. In (Ledesma et al., 2020), A two-level ANN-Based
diagnosis method for high-impedance fault location in unbalanced distribution
systems using synchronized measurements units is proposed. In terms of
Compared with the model-driven based models, data-driven models the equipment failures caused by the inter-turn short circuit fault, ANN
adopting machine learning techniques is more like a “black-box”, with can be used in the fault diagnosis for the turn to turn stator winding of
no explicit analytical description of defining relationship between input permanent magnet synchronous motor (Nyanteh et al., 2013), asyn­
and output. It is based on the fault feature extracted from large set of chronous machines (Toma et al., 2012), three-phase induction machines
experimental input data set and to form a function relating inputs and rotor faults (Barzegaran et al., 2013), unbalanced terminal voltage of an
outputs. induction traction motor (Moosavi et al., 2012).
A general framework illustrating the procedure of the machine ANN has a lot of merits, including the availability of multiple
learning based fault diagnosis is shown in Fig. 2, where it consists of four training algorithms, less needs of the formal statistical training, the
steps: data collection, feature extraction, model learning and diagnosis. scalability of complex non-linear relationships between dependent and
In step one of the data collection, the monitoring devices distributed independent variables, and the implementation simplicity (Moosavi
within the power system constantly collect data, such as vibration, et al., 2015). However, the disadvantage of the method is that a
acoustic emission, feeder status and current data. These unstructured well-trained ANN algorithm highly depends on the amount and quality
and semi-structured data types are collected in forms of image, text, of the input training data. In other words, the amount of input training
video and so forth. Accordingly, based on the types of collected data and data will otherwise slow convergence and lower the training efficiency.
application domains, researchers use different machine learning tech­ Besides, as the parameters such as learning rate, hidden layers, neurons
niques to realize the feature extraction and model learning, respectively. are black-boxed and cannot be easily obtained, the ANN algorithm has
In feature extraction, some commonly-used features (time-domain fea­ to be retrained whenever the system changes.
tures, frequency-domain features, and time–frequency-domain features)
of the data from multi-source monitoring devices is extracted (Lei et al.,
2020a). Worth mentioning that the extracted features usually contains

Fig. 1. All sectors requiring protection in the power system.

2
H. Yang et al. The Electricity Journal 34 (2021) 106881

Fig. 2. General framework of the procedure of using machine learning techniques for fault diagnosis.

margin between the two types and identifies the dividing hyperplane.
The margin separating hyperplane is calculated by the sum of the min­
imum distance between training data set, as indicated in Fig. 4. A simple
linear SVM classifier works by making a straight line (hyperplane) be­
tween two types of training data, indicating that the data points on one
side of the line will represent a category while those on the other side of
the line will be put into a different category. The detailed theory is
further given in (Gururajapathy et al., 2017).
Many researches haven been conducted on the fault diagnosis in the
power system. In (Ray and Mishra, 2016; Pradhan et al., 2004),
SVM-based methods are used to identify the fault types and the fault
distance of the transmission lines. In (Rudsari et al., 2019), it proposed a
SVM-based method to identify signals of contact travel waveforms and
Fig. 3. Architecture of back-propagation ANN with two layers.
coil current about the fault occurrence for high voltage circuit breakers.
The authors in (Jan et al., 2017) proposed a SVM-based method to
extract the statistical time-domain features and train a fault diagnosis
3.2. Support vector machine-based approaches
model based on the data collected from sensors considering erratic, drift,
hard-over, spike, and stuck faults. In (Livani and Evrenosoğlu, 2013), it
Support vector machine-based approach (SVM) is a supervised
proposed a traveling-wave-based method for fault classification and
learning method, which is widely used in classification and regression.
location for three-terminal power transmission systems using SVM.
The theory of linear SVM classification is shown in Fig. 4 where the
Compared with ANN-based methods, SVM-based methods have a
marked points are training data set(e.g., current, voltage, angle), as
better interpretability of the models with more rigorous theories. It is
input data for fault classification. Support vectors are the elements of the
effective for dealing with large-sized problems and otherwise even high
input training sets to separate all training data and maximizes the
dimensional problems requires less heuristics. These advantages make
SVM more attractive in large-sized statistical learning problems than
other supervised machine learning methods. However, the disadvan­
tages of SVM-based diagnosis models cannot be ignored and are briefly
summarized as follows. First, SVM-based methods are originally used to
solve binary classification tasks, and it always needs to use variant ar­
chitectures in terms of multi-class classification tasks. Besides, SVM-
based methods can work only well with a small number of various
data, such as unstructured and semi structured data, i.e., text, images
and trees, dealing with large amount of data may otherwise result in
computation burden. Least but not the last, the performance of SVM-
based diagnosis models is highly sensitive to kernel function and
hyper-parameters, but an appropriate kernel parameters is not easily
determined.

Fig. 4. Architecture of SVM.

3
H. Yang et al. The Electricity Journal 34 (2021) 106881

3.3. Other approaches 3.4.1. High variety

With the rise of big data in forms of unstructured and semi-structured
3.3.1. Expert system data types, such as text, audio, and video, additional preprocessing is
The expert system, a computer program that provides expert-level required to derive the useful meaning and supported metadata. Take the
diagnosis knowledge to automatically identify the health states of monitoring data of a wind turbine as an example, the collected data
equipment. It was first proposed by Edward Feigenbaum and Joshua include not only the speed and the vibration data from the monitoring
Lederberg in 1965, and was then widely applied in fault diagnosis (Lee system but also the control information from SCADA, resulting in the
et al., 2000). The expert system-based diagnosis models consist the data heterogeneity.
inference engine, the knowledge base, the user interface, the database,
and the explanation system (Lei et al., 2020b), (Minakawa et al., 1995). 3.4.2. High volume
An inference is an algorithm that interprets the knowledge in the The high volume of the data results from two aspects. One is because
knowledge base and provide an answer. A knowledge base is an orga­ that numerous monitoring devices are installed in variety forms of AMI,
nized collection of the knowledge about the corresponding domain. such as, distributed energy resources (DER) monitoring systems, SCADA
However, the interpretation of expert knowledge greatly influences the and WAMS, are generating multi-source data. The other reason is that
accuracy of the diagnosis results. Furthermore, limited by the low the data are generated continuously during the long-term operation of
self-learning capability of the expert system, the knowledge base is each AMI. During the long-term operation of these devices, the moni­
difficult to be expanded or updated. toring system constantly acquires data, finally resulting in high volume
of data.
3.3.2. K-Nearest Neighbor
The K-Nearest Neighbor (KNN)-based method is one of the most 3.4.3. High velocity
basic algorithms for classification problems, which works by seeking to Velocity is defined as the speed of data generation. For example, the
minimize the distance between the test and training observations (Lei high frequency synchronized WAMS are generating immense volume of
et al., 2020a). In this method, a distance metric, such as Euclidean, energy data with high revolution. Dealing with the high velocity data
Tanimoto, Jaccard, Mahalanobis and cosine distance, is used to train k requires distinct (distributed) processing techniques.
similar neighbors by searching the entire training dataset. However, the
difficulty in selecting the optimal neighborhood parameter and the 3.4.4. Low veracity
indistinguishable neighborhood boundary will greatly influence the When the percentage of low value and veracity of the collected big
performance of KNN-based algorithms. Therefore, some KNN-based data is high, they can be defined as noise of the big data. However, faults
variants are investigated and used for fault diagnosis of transformers seldom happen during the long-term operation of system equipment.
(Benmahamed et al., 2017), induction motors (Ali et al., 2019). Massive meaningless data are usually generated from the sensors on the
Worth mentioning that KNN works well with a small number of data equipment operating in healthy status that takes up the majority of the
inputs, but struggles when the number of inputs dramatically scales up long-term operation of the system devices. The other reason is because
or the distribution of the inputs are imbalance. Besides, the performance the quality of the collected data may suffer from the emergencies, such
of the diagnosis models is sensitive to the parameter k that is difficult to as equipment failures and transmission interruption.
be determined in high dimensional problems where training data may
be similar and have very large distances. 4. Trending machine learning based fault diagnosis

3.3.3. Decision tree Unfortunately, the fault diagnosis model based on conventional
Decision tree, as a tree-like graph, is another commonly-used su­ machine learning methods is not accurate and efficient for such
pervised learning method in classification, given their intelligibility and abovementioned big data scenarios. The applications of deep learning,
simplicity (Lei et al., 2020a). In a decision tree, each non-leaf node is reinforcement learning, unsupervised learning and so forth began to
labeled with a feature and each leaf is labeled with a probability dis­ attract more and more researchers’ interest. Here, we briefly introduce
tribution over the classes, indicating that the data inputs has been deep learning, reinforcement learning and transfer learning, who are the
classified into either a specific class. The arcs illustrate the classification most trending machine learning methods. Strictly speaking, the classi­
process down the tree from the root to the leaf node. The advantage of fication of these three types of learning methods are overlapping with
the decision tree-based classifications is that it can be naturally under­ each other, for example, some reinforcement learning algorithms are
stood and can be easily converted into diagnosis rules with little data also belonging to categories of the deep learning methods.
preprocessing. However, the disadvantages of this method is that it is
prone to be over fitting and requires to be careful with parameter tuning 4.1. Deep learning
the low generalization performance, which would weaken the diagnosis
performance. Deep learning is derived from the neutral network learning, where
“deep” contrasts itself with conventional shallow neural networks
3.4. Challenges of big data considering the depth of the network architecture (LeCun et al., 2015). It
has attracted intensive interest in fault diagnosis because of its enormous
In terms of the advances in Information and Communication Tech­ representing ability, automated feature learning capability with the
nologies (ICT) that have led to the era of “Industrial Big Data”, the extensible hidden layers and the best-in-class performance in solving
above-mentioned methods are not capable for tackling with the chal­ complex problems. In deep learning, it employs hierarchical architec­
lenges of big data that tend to be unstructured, multi-modal, heteroge­ tures to represent the learning of extracted features and to generate the
neous, decentralized, and highly nonlinear. In terms of the huge data expected outputs automatically without any human intervention. Three
steam being generated throughout the whole power system and the typical deep learning approaches are reviewed in this subsection.
consequent need of data storage and processing, dealing with the big
data in the power systems calls for customized methods and platforms. 4.1.1. Auto-encoder-based approaches
Here, we briefly summarize the main characteristics and challenges of Auto-encoder(AE) network methods have a feed-forward architec­
the big data in power system. ture which is able to learn feature representations of the input data
without supervision. The AE network could be divided into two com­
ponents, i.e., an encoder and a decoder and is specially used for

4
H. Yang et al. The Electricity Journal 34 (2021) 106881

dimensionality reduction applications, as shown in Fig. 5. As an unsu­

pervised learning method, the encoder compresses input data to hidden
layers with a smaller number of neurons, from which the decoder tries to
generate from the reduced encoded representation as close as possible to
its original input, hence its name. There is no doubt that AE-based
method requires minimizing the average reconstruction loss between
encoding and decoding process. Furthermore, several variants, such as
sparse (Zhao et al., 2020a; Sun et al., 2016), denoising (Zou et al., 2020),
Laplacian (Zhao et al., 2020b) and contractive (Zhang et al., 2020)
AE-based methods, have proven effective in learning representations for Fig. 6. Architecture of a RBM.
subsequent classification tasks.
the most significant features are extracted from the output of a con­
4.1.2. Restricted boltsmann machine-based approaches
volutional layer. With the dimensionality of an intermediate layer
A restricted Boltzmann machine (RBM) is a bipartite graph structure
reduced, overfitting could be avoided. Moreover, the network can ach­
that divides the model into two fully interconnected layers- visible
ieve the translation and rotation invariance property, which are highly
layers and hidden layers, indicated in Fig. 6. The first layer, correspond
desirable in many kinds of fault diagnosis tasks.
to the components of an observation (e.g., one visible unit for each pixel
According to the architectures of CNN, they can be divided into the
of a digital input image), and therefore called the “visible” units. The
1-dimensional (1D) and 2-dimensional (2D) models. The former is
hidden units work as non-linear feature detectors to model dependencies
usually successfully applied for the end-to-end fault diagnosis of rolling
between the components of observations (e.g., dependencies between
bearings (Eren et al., 2019), while the later serves for image identifi­
the pixels in the images). In the RBM’s network graph, every neuron in
cation, in which the input images are 2D data. For fault diagnosis of
the visible layer is connected to every neuron in the hidden layer, but
equipment, the 2D CNN is unable to handle the 1D signals, such as vi­
there are no connections between neurons in the same layer. This leads
bration data. In order to improve the performance of a CNN-based
to a useful feature representation of the inputs in the hidden layer and
diagnosis model, researchers adopted several effective approaches. In
the outputs can reconstruct the input approximately from the hidden
(Islam and Kim, 2019; Han et al., 2018; Guo et al., 2018; Cao et al.,
layer in the same way. As two variants of RBM methods, deep relief
2019), the signal processing methods, the continue wavelet transform,
network(DBN) and Deep Boltzmann Machine (DBM) have been suc­
the dual-tree complex wavelet transform are employed respectively to
cessfully applied to process high dimensional data, such as images and
preprocess the input 1D data, which are expected to convert the signals
text. Reference (Sun et al. (2014)) proposed a fault diagnosis model for
from the time domain to the time–frequency domain.
the electromotor based on DBN. In (Han et al., 2017), the DBN-based
diagnosis models are conducted with a stacked multiple RBMs, per­
forming well with a high diagnosis accuracy. 4.2. Reinforcement learning

4.1.3. Convolution neutral networks-based approaches Machine learning methods include three basic categories, supervised
Convolution neutral networks (CNN) has a feedforward architecture, learning, unsupervised learning and reinforcement learning. Different
consisting of convolutional layers, pooling layers, and fully-connected from other machine learning methods the training mechanism of rein­
layers, as shown in Fig. 7. CNNs are designed to process data with forcement learning does not conform to the human cognitive neurosci­
multidimensional data, e.g., color image, language, audio spectrogram, ence, leading to the challenge for exploring and solving the fault
and video. The convolution operator completes features extraction from diagnosis problem autonomously (Ding et al., 2019a). As a remarkable
the input using different filters (also known as kernels). In the con­ breakthrough in deep reinforcement learning, AlphaGo provides inspi­
volutional layers, the filters kernels are used to convolve the input ration and direction for combining the advantages of deep learning and
vectors from the previous layer. Correspondingly, in the pooling layer, reinforcement learning, to build an end-to-end fault diagnosis approach
which can directly map raw fault data to the corresponding fault types.
Reinforcement learning is a set of goal-oriented algorithms con­
cerned with how multiple software agents ought to take actions in an
environment in order to maximize the cumulative reward. With rein­
forcement learning approach, optimal strategies can be learned through
trial and error to process sequential decision-making problems in
different application domains, such as control theory, game theory,
simulation-based optimization, operations research, swarm intelligence,
information theory, statistics, and multi-agent systems. Here, we focus
on its applications in the fault diagnosis. An online search method based
on reinforcement learning and knowledge transfer for representative
risky fault chains is proposed in (Zhang et al., 2019b), which aims at
improving efficiency by exploiting similarities of adjacent power flow
snapshots in operations. In (Zhang et al., 2018), a novel fault detection
with minimizing the noise-signal ratio using reinforcement learning to
detect unexpected faults. In (Ding et al., 2019b), authors combined
autonomous learning ability of reinforcement learning and the percep­
tion ability of deep learning, to achieve an end-to-end method for
rotating machinery fault diagnosis.
However, preparing the simulation environment is the main chal­
lenge in reinforcement learning application, which is highly depending
on the task. When it comes to building a model capable of cascading
failure in power system, building a realistic simulator is crucial before
Fig. 5. Architecture of an auto-encoder. letting a long-term blackout in the realistic power system. The model has

5
H. Yang et al. The Electricity Journal 34 (2021) 106881

Fig. 7. Architecture of a 1D convolutional neural network.

to figure out how to insulate or disconnect a fault in a safe environment, machine learning-based approaches, such as ANN, SVM and other ap­
where sacrificing even thousands of prosumers comes at a minimal cost. proaches, which are usually supervised or semi-supervised learning
Although reinforcement learning, deep learning, and machine learning methods. However, with the era of big data comes, the fault diagnosis
are interconnected, no one of them is going to replace the others or used model faces more challenges, high variety, volume, velocity and low
in every case. Nevertheless, reinforcement learning still seems to be the veracity. Correspondingly, deep learning, reinforcement learning and
most likely way in making a diagnosis model creative – as seeking new, transfer learning, representative of trending machine learning ap­
innovative ways to perform its tasks. proaches, are introduced. Their applications, advantages and disad­
vantages are discussed in each respective subsection. This review is
expected to systematically present the development of machine learning
4.3. Transfer learning in applications in fault diagnosis of power system and provides a road­
map for future applications.
Transfer learning (TL) is a promising topic in machine learning,
which is applying storing knowledge gained from one problem solving
Declaration of Competing Interest
progress to other related scenarios. It is motivated by the concerns that
enormous resources required to train deep learning models, for example,
The authors report no declarations of interest.
the large datasets on which deep learning models are trained. Therefore,
by the transfer learning, a base network is first trained on a base dataset
or task with the learned features achieved, and then these features are Acknowledgements
repurposed and transferred for training a second target network on an
another target dataset or task. This process will perform well if the Funding: this work is supported by the National Natural Science
features are general to both base and target tasks. Obviously, transfer Foundation of China (Grant No. 51722701).
learning is related to multi-task learning problems and is not exclusively
an area of study for deep learning, which is expected to solve the References
problem without sufficient labeled data to train a reliable intelligent
Ali, M.Z., Shabbir, M.N.S.K., Liang, X., Zhang, Y., Hu, T., 2019. Machine learning-based
model. fault diagnosis for single-and multi-faults in induction motors using measured stator
Nowadays, transfer learning is gradually explored in the application currents and vibration signals. IEEE Trans. Ind. Appl. 55 (3), 2378–2391.
in the security issues in power system. In (Li et al., 2019), a novel Aslan, Y., 2012. An alternative approach to fault location on power distribution feeders
with embedded remote-end power generation using artificial neural networks.
transfer learning fault diagnosis method based on deep learning is Electr. Eng. 94 (3), 125–134.
conducted, in which the diagnostic knowledge firstly learned from suf­ Barzegaran, M., Mazloomzadeh, A., Mohammed, O.A., 2013. Fault diagnosis of the
ficient supervised data of multiple rotating machines and then trans­ asynchronous machines through magnetic signature analysis using finite-element
method and neural networks. IEEE Trans. Energy Convers. 28 (4), 1064–1071.
ferred to the target equipment with domain adversarial training. In (Ren Benmahamed, Y., Teguar, M., Boubakeur, A., 2017. Application of SVM and KNN to
and Xu, 2019), it proposes a novel data-driven method for pre-fault Duval Pentagon 1 for transformer oil diagnosis. IEEE Trans. Dielectr. Electr. Insul. 24
dynamic security assessment (DSA) of power systems. In (Sun et al., (6), 3443–3451.
Cao, X.C., Chen, B.Q., Yao, B., He, W.P., 2019. Combining translation-invariant wavelet
2018), a deep transfer learning network based on sparse auto encoder
frames and convolutional neural network for intelligent tool wear state
(SAE) is presented with three transfer strategies, that is, weight transfer, identification. Comput. Ind. 106, 71–84.
transfer learning of hidden feature, and weight update, used to transfer Ding, Y., Ma, L., Ma, J., Suo, M., Tao, L., Cheng, Y., Lu, C., 2019a. Intelligent fault
diagnosis for rotating machinery using deep Q-network based health state
an SAE trained by historical failure data to a new object.
classification: a deep reinforcement learning approach. Adv. Eng. Inform. 42,
The applications of transfer learning have great potential in using 100977.
available data more efficiently, improving the models’ ability to Ding, Y., Ma, L., Ma, J., Suo, M., Tao, L., Cheng, Y., Lu, C., 2019b. Intelligent fault
generalize and making the models more robust. As Andrew Ng, chief diagnosis for rotating machinery using deep Q-network based health state
classification: a deep reinforcement learning approach. Adv. Eng. Inform. 42,
scientist at Baidu and professor at Stanford, said during his widely 100977.
popular keynote in 2016 Conference and Workshop on Neural Infor­ Eren, L., Ince, T., Kiranyaz, S., 2019. A generic intelligent bearing fault diagnosis system
mation Processing Systems (NIPS) tutorial that transfer learning will be using compact adaptive 1D CNN classifier. J. Signal Process. Syst. 91 (2), 179–189.
Guo, S., Yang, T., Gao, W., Zhang, C., 2018. A novel fault diagnosis method for rotating
– after supervised learning – the next driver of machine learning com­ machinery based on a convolutional neural network. Sensors 18 (5), 1429.
mercial success. Gururajapathy, S.S., Mokhlis, H., Illias, H.A., 2017. Fault location and detection
techniques in power distribution systems with distributed generation: a review.
Renew. Sustain. Energy Rev. 74, 949–958.
5. Conclusion Hagh, M.T., Razi, K., Taghizadeh, H., 2007. Fault classification and location of power
transmission lines using artificial neural network. In: 2007 International Power
In this paper, we review the applications of machine learning to fault Engineering Conference (IPEC 2007, 1109-1114. IEEE.
Han, D., Zhao, N., Shi, P., 2017. A new fault diagnosis method based on deep belief
diagnosis in the power system, which is the primary step of the power network and support vector machine with Teager–Kaiser energy operator for
system protection and control. We first summarize the conventional bearings. Adv. Mech. Eng. 9 (12), 1687814017743113.

6
H. Yang et al. The Electricity Journal 34 (2021) 106881

Han, Y., Tang, B., Deng, L., 2018. Multi-level wavelet packet fusion in dynamic ensemble Toma, S., Capocchi, L., Capolino, G.A., 2012. Wound-rotor induction generator inter-turn
convolutional neural network for fault diagnosis. Measurement 127, 246–255. short-circuits diagnosis using a new digital neural network. IEEE Trans. Ind.
Ibrahim, M.S., Dong, W., Yang, Q., 2020. Machine learning driven smart electric power Electron. 60 (9), 4043–4052.
systems: current trends and new perspectives. Appl. Energy 272, 115237. Zhang, D., Lin, Z., Gao, Z., 2018. A novel fault detection with minimizing the noise-signal
Islam, M.M., Kim, J.M., 2019. Automated bearing fault diagnosis scheme using 2D ratio using reinforcement learning. Sensors 18 (9), 3087.
representation of wavelet packet transform and deep convolutional neural network. Zhang, L., Lin, J., Liu, B., Zhang, Z., Yan, X., Wei, M., 2019a. A review on deep learning
Comput. Ind. 106, 142–153. applications in prognostics and health management. IEEE Access 7, 162415–162438.
Jan, S.U., Lee, Y.D., Shin, J., Koo, I., 2017. Sensor fault classification based on support Zhang, Z., Yao, R., Huang, S., Chen, Y., Mei, S., Sun, K., 2019b. An online search method
vector machine and statistical time-domain features. IEEE Access 5, 8682–8690. for representative risky fault chains based on reinforcement learning and knowledge
Javadian, S.A.M., Massaeli, M., 2011. A fault location method in distribution networks transfer. IEEE Trans. Power Syst. 35 (3), 1856–1867.
including DG. Indian J. Sci. Technol. 4 (11), 1446–1451. Zhang, Y., Li, X., Gao, L., Chen, W., Li, P., 2020. Ensemble deep contractive auto-
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436–444. encoders for intelligent fault diagnosis of machines under noisy environment.
Ledesma, J.J.G., do Nascimento, K.B., de Araujo, L.R., Penido, D.R.R., 2020. A two-level Knowledge Based Syst. 105764.
ANN-based method using synchronized measurements to locate high-impedance Zhao, X., Jia, M., Liu, Z., 2020a. Semi-supervised deep sparse auto-encoder with local
fault in distribution systems. Electr. Power Syst. Res. 188, 106576. and non-local information for intelligent fault diagnosis of rotating machinery. IEEE
Lee, H.J., Park, D.Y., Ahn, B.S., Park, Y.M., Park, J.K., Venkata, S.S., 2000. A fuzzy expert Trans. Instrum. Meas. https://doi.org/10.1109/TIM.2020.3016045 [online]. (Early
system for the integrated fault diagnosis. IEEE Trans. Power Deliv. 15 (2), 833–838. access).
Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., Nandi, A.K., 2020a. Applications of machine Zhao, X., Jia, M., Lin, M., 2020b. Deep Laplacian Auto-encoder and its application into
learning to machine fault diagnosis: a review and roadmap. Mech. Syst. Signal imbalanced fault diagnosis of rotating machinery. Measurement 152, 107320.
Process. 138, 106587. Zou, L., Li, Y., Xu, F., 2020. An adversarial denoising convolutional neural network for
Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., Nandi, A.K., 2020b. Applications of machine fault diagnosis of rotating machinery under noisy environment and limited sample
learning to machine fault diagnosis: a review and roadmap. Mech. Syst. Signal size case. Neurocomputing. 407, 105–120.
Process. 138, 106587.
Li, X., Zhang, W., Ding, Q., Li, X., 2019. Diagnosing rotating machines with weakly
Hanyu Yang received the B.E. degree in electrical engineering from the College of Elec­
supervised data using deep transfer learning. IEEE Trans. Industr. Inform. 16 (3),
trical Engineering in Hunan University, Changsha, China, in 2015. She was a visiting
1688–1697.
student at Robert W. Galvin Center for Electricity Innovation at Illinois Institute of
Lim, P.K., Dorr, D.S., 2000. Understanding and resolving voltage sag related problems for
Technology. She is currently pursuing the Ph.D. degree in electrical engineering from
sensitive industrial customers. In: 2000 IEEE Power Engineering Society Winter
Hunan University, Changsha, China. Her research interests include power system plan­
Meeting. Conference Proceedings, 4, pp. 2886–2890.
ning, integrated energy system, machine learning.
Livani, H., Evrenosoğlu, C.Y., 2013. A fault classification and localization method for
three-terminal circuits using machine learning. IEEE Trans. Power Deliv. 28 (4),
2282–2290. Xubin Liu received the B.E. degree in automation from the College of Electrical Engi­
Minakawa, T., Ichikawa, Y., Kunugi, M., Shimada, K., Wada, N., Utsunomiya, M., 1995. neering in Northwest University for Nationalities, Lanzhou, China in 2013, and the Ph.D.
Development and implementation of a power system fault diagnosis expert system. degree in electrical engineering from Hunan University, Changsha, China in 2019. He is
IEEE Trans. Power Syst. 10 (2), 932–940. currently a Postdoctoral Researcher with the School of Electrical and Electronic Engi­
Moosavi, S.S., Djerdir, A., Aït-Amirat, Y., Khaburi, D.A., 2012. Fault detection in 3-phase neering, Huazhong University of Science and Technology (HUST), Wuhan 430074, China.
traction motor using artificial neural networks. June. In: 2012 IEEE Transportation His research interests include power system dispatch, self-healing of smart grid, and power
Electrification Conference and Expo. IEEE, pp. 1–6. electronics control.
Moosavi, S.S., Djerdir, A., Ait-Amirat, Y., Khaburi, D.A., 2015. ANN based fault diagnosis
of permanent magnet synchronous motor under stator winding shorted turn. Electr.
Di Zhang received her B.E. degree in electric engineering and automation from Henan
Power Syst. Res. 125, 67–82.
Normal University, Xinxiang, China, in 2014. She is currently pursuing her Ph.D. degree in
Nyanteh, Y.D., Srivastava, S.K., Edrington, C.S., Cartes, D.A., 2013. Application of
electrical engineering at Hunan University, Changsha, China. She was a visiting PhD
artificial intelligence to stator winding fault diagnosis in permanent magnet
student at the Robert W. Galvin Center for Electricity Innovation at Illinois institute of
synchronous machines. Electr. Power Syst. Res. 103, 201–213.
technology. Her major research interests include power system operation and control,
Pradhan, A.K., Routray, A., Biswal, B., 2004. Higher order statistics-fuzzy integrated
energy policy and cyber security for smart grid.
scheme for fault classification of a series-compensated transmission line. IEEE Trans.
Power Deliv. 19 (2), 891–893.
Purushothama, G.K., Narendranath, A.U., Thukaram, D., Parthasarathy, K., 2001. ANN Tao Chen received the B.E. degree in Electrical engineering and automation from Anhui
applications in fault locators. Int. J. Electr. Power Energy Syst. 23 (6), 491–506. Agricultural University, Hefei, China in 2019. He is currently pursuing the M.S. degree at
Ray, P., Mishra, D.P., 2016. Support vector machine based fault classification and the College of Electrical and Information Engineering, Hunan University, Changsha,
location of a long transmission line. Eng. Sci. Technol. Int. J. 19 (3), 1368–1380. China. His main research interests are power system planning, equipment automation and
Ren, C., Xu, Y., 2019. Transfer learning-based power system online dynamic security machine learning.
assessment: using one model to assess many unlearned faults. IEEE Trans. Power
Syst. 35 (1), 821–824. Canbing Li received the B.E. and Ph.D. degree from Tsinghua University, Beijing, China,
Rudsari, F.N., Razi-Kazemi, A.A., Shoorehdeli, M.A., 2019. Fault analysis of high-voltage in 2001 and 2006, respectively, both in electrical engineering. He is currently with the
circuit breakers based on coil current and contact travel waveforms through School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong Univer­
modified SVM classifier. IEEE Trans. Power Deliv. 34 (4), 1608–1618. sity. His research interests include power systems, smart grid, renewable energy, with an
Sun, J., Wyss, R., Steinecker, A., Glocker, P., 2014. Automated fault detection using deep emphasis on large-scale power system dispatch, economic and secure operation of power
belief networks for the quality inspection of electromotors. tm-Technisches Messen systems, energy efficiency and energy saving in smart grid, electric demand management
81 (5), 255–263. of data centers, vehicle-to-grid technologies.
Sun, W., Shao, S., Zhao, R., Yan, R., Zhang, X., Chen, X., 2016. A sparse auto-encoder-
based deep neural network approach for induction motor faults classification.
Measurement 89, 171–178. Wentao Huang received the Ph.D. degree in electrical engineering from Shanghai Jiao
Sun, C., Ma, M., Zhao, Z., Tian, S., Yan, R., Chen, X., 2018. Deep transfer learning based Tong University, Shanghai, China, in 2015. He is currently a Lecturer with the Department
on sparse autoencoder for remaining useful life prediction of tool in manufacturing. of Power Electrical Engineering, Shanghai Jiao Tong University. His research interests
IEEE Trans. Industr. Inform. 15 (4), 2416–2425. include protection and control of active distribution systems, microgrids, smart grid, and
renewable energy.