A Review on Fault Diagnosis Technology of Key Components in Cold Ironing System

Abstract

Nowadays, cold ironing technology has been demonstrated to be an effective solution to deal with the environmental and social problems brought by port ship emissions and relevant effects. The working states of cold ironing equipment, especially the key components such as circuit breakers, transformers and frequency converters, have a significant effect on the safety and reliability of the whole system. However, due to the harsh working environment of cold ironing equipment, they are prone to a high risk of failure. In this respect, fault diagnosis methods can play a significant role in detecting potential failure in time and guarantee the safe and reliable operation of the cold ironing system. In recent years, research on the fault diagnosis of a cold ironing system has been rapidly growing, and this paper aims to present a comprehensive review of this literature, with an emphasis on the fault diagnosis technology applied to the key components in a cold ironing system. This review classifies the literature according to the type of key component, and, for each special type of component, the fault diagnosis methods are further categorized and analyzed. This paper provides useful references for professionals and researchers working on the fault diagnosis of a cold ironing system and points out valuable research directions in the future.

1. Introduction

In the last few years, as one of the crucial technologies of port energy conservation and emission reduction, cold ironing technology has received more and more attention from governments and academies all over the world. Port ship emissions have a significant impact on climate change and local air pollution [1]. Cold ironing has been demonstrated to be an effective technology, which can provide a great deal of social and environmental advantages by decreasing the air pollution caused by ship berthing in harbor areas [2,3,4,5,6,7]. From a technical point of view, cold ironing is a method of providing power to ships by connecting to a fixed grid [8]. As shown in Figure 1, a cold ironing system (also named a shore-to-ship power system) mainly includes three parts: a cold ironing supply system, ship-to-shore connection system and cold ironing receiving system.

As one of the most significant environmental protection innovations of ports all over the world, cold ironing technology has been widely used. Well-known ports, such as Port of Los Angeles [9], Port of Shanghai, Port of Dalian [10] and Port of Barcelona [3], have adopted cold ironing technology, leading to a noticeable reduction in sulfide and nitride emissions. However, due to the harsh working environment, including high temperature and humidity, as well as high corrosion in the port, cold ironing equipment is prone to a high risk of failure. If not discovered and solved in time, it may cause serious pecuniary losses and even injuries and deaths. Therefore, in order to ensure the safety and reliability of a cold ironing system, we must attach great importance to the fault diagnosis of the system. The research value on the fault diagnosis of a cold ironing system includes, but is not limited to, the following aspects: (1) ensuring the reliability of the cold ironing system; (2) avoiding economic losses caused by system failure; (3) improving the satisfaction rate of cold ironing services; (4) promoting further applications of cold ironing technology and increasing the utilization rate of cold ironing, thereby contributing to the green port construction.

This paper mainly summarizes related literature on the fault diagnosis of cold ironing systems, aiming at providing a comprehensive view of the application of fault diagnosis technology to a cold ironing system in the past few years, as well as exploring the potential research direction in this area. Approximately 180 articles were screened and sifted, and, finally, 118 articles were analyzed and researched. The research method utilized in this paper is described in Figure 2.

Step 1: Identify keywords. Firstly, by reading articles related to cold ironing, we find that “cold ironing” is also named “shore-to-ship power (SSP)” thus, the first keyword can be determined. Then, considering that this paper studies the fault diagnosis technology of cold ironing systems, the second key word “fault diagnosis” or “fault prediction” is determined.

Step 2: Identify relevant databases: ScienceDirect-Elsevier, EBSCO database, IEEE Xplore.

Step 3: Search for the related articles in databases.

Selection criteria: 1. Related to key equipment in a cold ironing system. 2. Related to the fault diagnosis/prediction technology. 3. Journal articles with certain impact factors.

Exclusion criteria: 1. Not related to key equipment in a cold ironing system. 2. Related to cold ironing but irrelevant to fault diagnosis. 3. Conference papers or book chapters.

Step 4: Summarize and classify the results: on obtaining the searching results in accordance with our criteria, the fault diagnosis technology of cold ironing systems involved in these articles are summarized and classified.

It is observed that the three subsystems (namely cold ironing supply system, ship-to-shore connection system and cold ironing receiving system) in a cold ironing system consist of many cold ironing components, among which the circuit breaker, transformer and frequency converter are the most important components, which have a significant impact on the safety and reliability of the cold ironing system. Therefore, this paper focuses on the fault diagnosis technology of the three key components of cold ironing systems.

The researched articles are first classified by fault diagnosis object, and further categorized by the specific fault diagnosis methods. According to Ref. [11], the key and fault-prone components in cold ironing systems include the circuit breaker, transformer and frequency converter. Through investigating the other literature, it can be found that most of the cold ironing fault diagnosis research focuses on these key components.

Generally, the fault diagnosis methods can be divided primarily into model-based methods, data-driven methods and knowledge-based or rule-based methods. The flowchart of the fault diagnosis method is described in Figure 3.

However, the operating environment of equipment in a cold ironing system is quite complex and varies from port to port. It is greatly affected by solar radiation, wind level, air humidity and other factors, making it difficult to construct accurate fault mechanism models of the cold ironing components for different ports. Additionally, inaccurate physical models will seriously affect the accuracy of subsequent fault diagnosis. Due to the above reasons, the application of physical-model-based fault diagnosis methods is quite limited in the cold ironing scenario. In contrast, data-driven and knowledge-based or rule-based methods are more suitable for fault diagnosis in cold ironing systems and, thus, become our focus. The classification method used in this paper is presented in Figure 4.

The remainder of this paper is organized as follows. Section 2 provides an introduction of the key equipment in a cold ironing system as well as fault diagnosis technology. Section 3 classifies the literature according to the type of key component, and, for each special type of component, the fault diagnosis methods are further categorized and analyzed. Finally, Section 4 concludes the paper and points out some future research directions on the fault diagnosis technology of cold ironing systems.

2. Introduction of Key Components in Cold Ironing System and Fault Diagnosis Technology

2.1. Key Components in Cold Ironing System

The working states of cold ironing equipment, especially the key components such as circuit breakers, transformers and frequency converters, directly affect the safety and reliability of the whole system. Operating in a harsh environment makes them more prone to a high risk of failure. In this section, the key components in a cold ironing system and their common fault types are introduced.

2.1.1. Circuit Breaker

The circuit breaker is the key component for the safety protection of the cold ironing system. In the case of a local fault occurring in the system, the circuit breaker can cut off the faulty part to prevent further spread of the fault and ensure the safety of staff and the other normal cold ironing devices. The typical structure of a circuit breaker consists of a pedestal, support insulation part, transmission part, breaking element and operating mechanism, etc., as shown in Figure 5. Its common failure modes include an iron core fault, base screw looseness, poor lubrication, coils fault, spring fault, damper fault, etc.

2.1.2. Transformer

A complete cold ironing system involves many components, among which the transformer always plays a vital role in system operation and is regarded as a key component. A typical structure of the transformer is shown in Figure 6, and its function is to change the voltage amplitude on demand. When the voltage provided by the shore power does not match the working voltage of the ship, the transformer is used to change the voltage amplitude to meet the normal working requirements of the ship. The common fault modes of a transformer include partial discharge, low energy discharge, arc discharge, overheating, etc.

2.1.3. Frequency Converter

So far, the power frequency of equipment is inconsistent for different countries and regions, and the power frequency used by ships is also inconsistent with that of shore power; moreover, a similar situation also occurs with seaport ships and inland ships. Therefore, a frequency converter, whose function is to convert the power frequency used on shore to make it consistent with the ship demand, is a critical component in the cold ironing system. As shown in Figure 7, the cold ironing frequency conversion system is composed of a step-down transformer, rectifier device, inverter device, filter structure and other devices. The common failure modes of frequency converters include overvoltage, overcurrent, overload failure, etc.

2.2. Fault Diagnosis Technology

As introduced in Section 1, due to the adaptability to the fault diagnosis problem in the cold ironing scenario, data-driven methods and knowledge-based or rule-based methods are the focus of this paper, which are introduced in the following.

2.2.1. Data-Driven Methods

(1)

Clustering

As a typical unsupervised learning method in machine learning, clustering algorithm has attracted a great deal of research interest from academies, resulting in plenty of corresponding research work and derivation of various different methods based on it. Among the clustering algorithms, K-means algorithm is the most common one due to its simple structure and fast calculation speed. Compared with the hard clustering of K-means algorithm, fuzzy C-means (FCM) algorithm can provide more flexible clustering results. However, in most cases, it is impractical to divide the objects in a dataset into distinct clusters, and, thus, simply assigning an object to a particular cluster can be error-prone. To address this problem, a weight can be assigned to each object corresponding to each cluster, indicating how likely it is that the object belongs to that cluster.

Suppose that $X=\left\{x_1,x_2,\dots ,x_i,\dots ,x_n\}\right.$ is a sample set containing n samples, and $x_i=\left\{d_{i1},d_{i2},\dots ,d_{im}\right\}$ denotes a sample with $m$ eigenvectors. The sample set is divided into $c$ clusters; we need to find the cluster centers denoted by $V=\left\{v_1, v_2,\dots ,v_j,\dots , v_c\}\right.$ and the degree of membership $u_{ij}$ (the membership degree of sample $x_i$ to the $j$-th cluster), which minimizes the value function of the dissimilar indicators, as shown in Equation (1).

$$J={\displaystyle \sum }_{i=1}^n{\displaystyle \sum }_{j=1}^c{u}_{ij}^b||x_i-v_j||{}^2,1\le m<\infty$$

(1)

where $v_j$ denotes the cluster center of the $j$-th cluster, and $b$ is the smoothing index.

The FCM clustering algorithm works as follows:

Step 1: Set cluster number $c$, threshold $\epsilon$ for terminating the iteration, weighting index $b$ and initialize the degree matrix of membership denoted by $U=\left\{u_{ij}\right\}$ ($i$ = 1, 2, 3, …, n; $j$ = 1, 2, 3, …, c). Note that $b$ is a parameter that controls the flexibility of the algorithm. Specifically, the larger $b$ is, the more fuzzy the classification is; otherwise, it becomes closer to the hard clustering algorithm.

Step 2: Calculate the centers of all clusters $V=\left\{v_1, v_2,\dots ,v_j,\dots , v_c\}\right.$ as follows:

$$v_j=\frac{{{\displaystyle \sum }}_{i=1}^n{u}_{ij}{}^b{x}_i}{{{\displaystyle \sum }}_{i=1}^n{u}_{ij}{}^b}$$

(2)

Step 3: Calculate and update the degree matrix of membership as follows:

$$v_j=1/{\displaystyle \sum }_{k=1}^C{(\frac{||x_i-v_j||}{||x_i-v_k||})}^{\frac{2}{b-1}}$$

(3)

According to $u_{ij}$ and $v_j$ obtained from the above steps, the objective function can be calculated by Equation (1), and the iteration will be stopped until |J^(k+1) − J^(k)| < ε is satisfied ($k$ denotes the number of iterations). Here, we can obtain the cluster centers, fuzzy membership degree as well as the number of iterations, and fault diagnosis can be realized based on the membership degree matrix. Otherwise, go back to Step 2 and execute the following steps.

(2)

Support vector machine (SVM)

Since being formally proposed by Vapnik in 1995, SVM has received a great deal of attention from researchers [12]. SVM is a kind of machine learning method that is applicable to small sample classification [13]. It tries to use a nonlinear function to map the input vector to a high-dimensional feature space, and then constructs a linear classification hyperplane with the largest classification interval in this space so that the two types of sample sets can be separated with a satisfactory accuracy.

The primary idea of SVM can be illustrated by the binary classification problem shown in Figure 8, where the circle and triangle represent the two categories of samples, respectively; $H$ denotes the optimal dividing line; while $H_1$ and $H_2$denote the closest parallel lines to $H$ in the two types of samples, and the distance between them is named as classification margin. The samples on lines $H_1$ and $H_2$ are called support vectors.

If we extend $H$ from two-dimensional to high-dimensional space, the optimal classification line will become the optimal classification surface, that is, the optimal hyperplane. The purpose of classification is to find the optimal hyperplane and completely separate the two types of samples.

(3)

Shallow neural network

MP model is the first neuron mathematical model proposed by psychologist McCulloch and mathematical logician Pitts in 1943 [14], which is of groundbreaking significance and provides a theoretical foundation for subsequent research work. In the late 1950s and early 1960s, Rosenblatt added a learning function to the traditional MP model, and, based on this, he proposed a single-layer perceptron model. This is the first time to bring neural network research into practice [15,16]. However, the proposed model is not able to handle the problem of linear inseparability. Rumelhart et al. [17] proposed a multi-layer feedforward network model, which is trained by the algorithm of error backpropagation. The backpropagation network (BP network) was proposed in 1986, which can solve some problems where single-layer perceptron is incapable. Since the 1990s, various kinds of shallow neural network models have emerged, among which BP neural network and extreme learning machine (ELM) are the most widely used methods.

• BP neural network

BP neural network is proposed as a multi-layer feedforward neural network. The basic idea of BP neural network is to apply the gradient descent method to train the network according to the error backpropagation algorithm so as to minimize the mean square error between the expected output value of the neural network and the actual output value. A general structure of the BP neural network is shown in Figure 9.

2.

Extreme learning machine (ELM)

ELM was proposed by Guang-Bin Huang et al. in 2004 and published on the IEEE International Joint Conference that year. The purpose of ELM is to improve the backward propagation algorithm by accelerating its learning efficiency and simplifying the setting of learning parameters [18]. In 2006, after further evaluation of the proposed algorithm, the original authors of ELM published the research paper on neurocomputing, and much attention has been received [19].

ELM is an improved algorithm for single-hidden-layer feedforward neural network, which is composed of input layer, hidden layer and output layer. The learning process of ELM is as follows:

(a)

Determine the number of neurons in the hidden layer, the activation function and the connection weight $w$ between the input layer and the hidden layer as well as the threshold $b$ of the neurons in the hidden layer.

(b)

Calculate the output matrix $H$ of the hidden layer.

(c)

Calculate the weight $\beta =H^{+}T$ of the output layer.

In the above description, $H^{+}$ denotes the Moore–Penrose generalized inverse matrix of $H$, and $T$ represents the expected output.

(4)

Ensemble learning (EL)

EL is also a popular method in the field of machine learning, the key idea of which is to establish multiple learning models and then fuse their decision results to obtain better results than single learning model. Commonly used methods of ensemble learning include bagging [20], boosting [21], random forest (RF) [22], XGboost [23], GBDT [24], etc. The ensemble learning process consists of two main steps: base model construction and fusion method selection. A well-chosen fusion method can effectively improve the performance of ensemble learning. We take RF as an example to introduce how it works, and the concrete steps of RF are as follows:

• Given a training set $X=\left\{x_1,x_2,\dots ,x_n\}\right.$, together with the category label $Y=\{y_1$,$y_2$,…,$y_n\}$, where $\mathrm{x}\in R^m$, $n$ denotes the number of samples in the training set. Then, the original training set is resampled to generate $N$ new training sets {$X_1$,$X_2$,…,$X_n$}, and the corresponding label sets are {$Y_1$,$Y_2$,…,$Y_n$}.
• Based on {$X_1$,$X_2$,…,$X_n$} and {$Y_1$,$Y_2$,…,$Y_n$}, $N$ complete decision trees {$t_1$,$t_2$,…,$t_n$} are generated. During the construction of each decision tree, each node performs a random selection of the feature subsets.
• Finally, the decision results of the test samples are obtained by the majority voting method.

(5)

Deep Learning (DL)

In 2006, Hinton et al. published a paper of significant value in science [25]. The main points include: (1) artificial neural networks (ANN) with multiple hidden layers have excellent capabilities of feature learning; (2) they can effectively overcome the difficulties of deep neural network training through layer-wise pre-training, which has since led to the future study of DL and ANN [26].

In the layer-by-layer pre-training algorithm of DL, it is the first time to apply unsupervised learning to the pre-training procedure of each layer, and, for each time, only one layer of unsupervised learning is performed. The training result of one layer then acts as input to the next layer. Afterwards, the pretrained network is fine-tuned by utilizing supervised learning method, such as BP algorithm [27]. In Ref. [28], Bengio systematically introduces the network structures and learning methods in deep learning. At present, well-known deep learning models include convolutional neural networks (CNN), recurrent neural networks (RNN), etc.

(1)

CNN

The earliest CNN include time-delay network and LeNet-5 [29]. It has been developing rapidly with the fast improvement and upgrade of the numerical computing equipment. The basic structure of CNN consists of an input layer, a pooling layer and a fully connected layer. Figure 10 shows a simplified example of CNN, where the convolution kernel size is 2 × 2 and the stride is 2, the pooling size is 2 × 2 and the stride is 2.

Convolution layer: feature extraction is performed on the input data through the convolution kernel. The size of the convolution kernel determines the size of the connection area between the convolution layer and the input layer, while the step size determines the number of steps the convolution kernel moves in the horizontal and vertical directions. The elements in the convolution kernel represent the weights. By calculating the weighted summation of local input and convolution kernel, adding the bias, and, further, performing nonlinear transformation through the activation function, the output $h_n$ of the convolution layer can be calculated by:

$$h_n=f\left(w{x}_n+b\right) , n=1, 2,\dots , N$$

(4)

where $w$ represents the weight matrix of the convolution kernel; $b$ is the bias; $x_n$ is the local input of the $n^{th}$ convolution kernel size and $f$ denotes the activation function.

Pooling layer: after extracting the features through the convolution layer in the previous step, the pooling layer performs feature selection and information filtering on the obtained features. Commonly used methods include maximum pooling, minimum pooling and average pooling. For instance, in maximum pooling, the maximum value in the pooling area is selected.

Fully connected layer: in this layer, multi-dimensional features are expanded into one-dimensional feature through multiple hidden layers to the output layer. The network structure and principle are the same as those of traditional neural network.

(2)

RNN

RNN is a novel network structure developed by adding recurrent structure to the traditional ANN. The recurrent structure of RNN can take advantage of the processed information at the previous moment and apply it to the data processing process at the current moment, as shown in Figure 11. Based on this characteristic, RNN becomes a powerful tool to deal with time series data.

The output of the hidden layer in RNN can be obtained by:

$$h_t=f\left(W_x{x}_t+W_h{h}_{t-1}+b\right)$$

(5)

where $W_x$ is the weight matrix of the input data, and $W_h$ denotes the weight matrix of the hidden layer output at the previous moment, which enters the hidden layer recurrently at the current moment. $x_t$ represents the input data matrix at the current moment; $h_t$ and $h_{t-1}$ denote the output matrix of the hidden layer at the current moment and the previous moment, respectively.

2.2.2. Knowledge-Based or Rule-Based Methods

(1)

Fuzzy logic (FL)

Fuzzy logic is a kind of method that can imitate human brain’s judgment on concept of uncertainty, as well as the way of thinking and reasoning. For system whose model is unknown or indescribable, fuzzy sets and fuzzy rules can be applied for reasoning. With the help of the concept of membership function, fuzzy logic does well in distinguishing fuzzy sets, dealing with fuzzy relationships and simulating human brain to perform rule-based reasoning, which endows it with the ability to express qualitative knowledge and experience with unclear boundaries.

Suppose $U$ is an object space, $x$ is an element in space $U$ and set $A$ ($A\in U$) is a set of $x$ that satisfies a certain condition, i.e., A = {$x|x$ satisfies a certain condition}. We can also define set $A$ through the characteristic function $A\left(x\right)$:

$$A\left(x\right)=\left\{\begin{array}{c}1 x\in A\\ 0 x\notin A\end{array}\right.$$

(6)

Based on the characteristic function, the mapping relationship between set $A$ and element $x$ can be established, i.e., $A\left(x\right):U\to \left\{0,1\right\}$. This kind of mapping reflects the deterministic relationship between set $A$ and element $x$, and the elements of A can only take two values (i.e., 0 or 1); therefore, it is called a clear set.

The problem becomes complicated when the belonging relationship is not that clear. In order to describe the uncertain belonging relationship, membership function $u_A\left(x\right)$ is used to characterize the degree to which element $x$ belongs to set $A$, where 0 ≤ $u_A\left(x\right)$ ≤ 1. When $u_A\left(x\right)$ = 0, it means that element $x$ does not belong to set $A$ at all, while $u_A\left(x\right)$ = 1 means that element $x$ completely belongs to set $A$. A larger value of $u_A\left(x\right)$ implies a higher possibility that element $x$ belongs to set $A$. However, clear sets cannot describe such boundaries or affiliations with unclear relationships. To deal with this problem, fuzzy set $A$ on the object space $U$ is introduced, which can be defined as an ordered pair set:

$$A=\left\{\left(x, u_A\left(x\right)\right) | x\in U\right\}$$

(7)

When adopting fuzzy sets to deal with practical problems, commonly used membership functions $u_A\left(x\right)$ take the form of Gaussian function, trapezoid, triangle, bell or other functions.

(2)

Rough Set Theory (RST)

As an effective tool for dealing with imprecise, inconsistent or incomplete information, rough set theory does not need any a priori knowledge besides the data set to be processed. Moreover, it is highly complementary to other theories dealing with uncertainty (especially fuzzy theory).

A knowledge system $S$ in rough set theory can be expressed as:

$$S=\left(U, A, V, f\right)$$

(8)

where $U$ is the object set; $A=C\cup D$ denotes the attribute set, $C\cap D\ne \varnothing$, $C$ denotes the conditional attribute set of $A$ and $D$ is the decision attribute set of $A$; $V$ is the value range set of attributes;$f:U\times S\to V$ is the information function.

For some special type of equipment fault, using eigenvector $U=\{x_1$,$x_2$,…,$x_n\}$ as an expression of the fault mode can evidently facilitate the fault decision by establishing a decision table according to the knowledge expression system of condition attributes and decision attributes. By utilizing the decision table, we can explore the correlation and dependence between attributes and obtain the ranking of attribute importance to obtain the optimal attribute set so as to make accurate decisions.

(3)

D-S evidence theory

D-S evidence theory is often used as a method to fuse the results of multiple classifiers in ensemble learning. The relevant basic knowledge and concepts are as follows:

Definition 1.

Let $U$ be a set representing all values of $x$, and all elements in $U$ are incompatible with each other, then $U$ is called the recognition frame on $x$.

Definition 2.

Let $U$ be a recognition frame. When function m: $2^U\to \left[0, 1\right]$ satisfies the conditions of $m\left(\varnothing \right)=0$ and ${\displaystyle \sum }_{A\subset U}m\left(A\right)=1$, $m\left(A\right)$ is the basic believability of $A$, and function $m$ is the basic belief distribution (function) on $U$. All $A$ elements satisfying $m\left(A\right)$ > 0 are called focal elements of an evidence.

Definition 3.

Let $U$ be a recognition frame. Function $Bel:2^U\to \left[0, 1\right]$ can be called a belief function on this recognition framework when the following two conditions are met:

$$\left\{\begin{array}{c}Bel\left(\varnothing \right)=0\\ Bel\left(A\right)=\sum m\left(B\right) \left(\forall A\subset U\right)\end{array}\right.$$

(9)

where $Bel\left(A\right)$ represents the sum of the basic believability of all subsets of $A$.

Definition 4.

Let $U$ be a recognition frame. Function $Pl: 2^U\to \left[0,1\right]$ can be called a likelihood function on $U$ when the following two conditions are satisfied:

$$\left\{\begin{array}{c}Pl\left(\varnothing \right)=0\\ Pl\left(A\right)={\displaystyle {\displaystyle \sum }_{B\cap A\ne \varnothing }}m\left(B\right) \left(\forall A\subset U\right)\end{array}\right.$$

(10)

Besides, [0, $Bel\left(A\right)$] represents the belief interval of $A$, [0, $Pl\left(A\right)$] represents the likelihood interval of $A$ and thus the interval [$Bel\left(A\right)$, $Pl\left(A\right)$] represents the uncertainty interval of $A$. Dempster’s combination rule of evidence theory is expressed as follows:

Assuming that there are two independent evidences on the same recognition frame $U$, with basic belief assignments $m_1$ and $m_2$, respectively. Then, for any assumptions $\theta$, we have the following combination rules:

$$m\left(\theta \right)=\left[m_1⨁m_2\right]\left(\theta \right)=\left\{\begin{array}{c}0, \theta =\varnothing \\ \frac{{{\displaystyle \sum }}_{B\cap A\ne \theta }{m}_1\left(A\right)m_2\left(B\right)}{1-K},\theta \ne \varnothing \end{array}\right.$$

(11)

where $K$ is the conflict factor reflecting the conflict degree of evidences, and it can be defined as:

$$K={\displaystyle \sum }_{B\cap A=\varnothing }{m}_1\left(A\right)m_2\left(B\right)$$

(12)

2.2.3. Comparison of Fault Diagnosis Methods

The comparison of fault diagnosis methods is shown in

Table 1. Comparison of fault diagnosis methods.

Method			Advantages	Disadvantages
Data-driven	Clustering	K-means	(1) Simple, efficient and fast convergence; (2) Can achieve better effect when the cluster is close to Gaussian distribution.	(1) The average distance must be defined; (2) K should be given in advance; (3) The value of K affects the clustering effect and has a great impact on outliers.
		DBSCAN	(1) Can deal with clusters of any shape without knowing K in advance; (2) Can identify noise points, and has good robustness to outliers.	(1) If the cluster density varies greatly and is uneven, the clustering effect is not good; (2) The sample is large and the convergence time is long.
		Hierarchical clustering	(1) The similarity between distance and rule is easy to define; (2) Can deal with clusters of any shape without knowing K in advance; (3) The hierarchical relationship of classes can be found.	(1) The calculation complexity is too high; (2) Singular value can also have a great influence; (3) The algorithm is likely to cluster into chains.
	SVM	Construction of multi-classification SVM by direct method	(1) Only need to train k classifiers; (2) The number is small; (3) The classification speed is relatively fast.	(1) The training speed will slow down sharply, and the situation of sample asymmetry tends to be serious with the increase in training data; (2) When a new category is added, all models need to be retrained.
		Construction of multi-classification SVM by indirect method	(1) In case of increasing samples, there is no need to retrain all SVMs, can only retrain and increase sample-related classifiers; (2) When training a single model, the relative speed is faster.	(1) The number of binary classifiers that need to be constructed and tested increases as a quadratic function with respect to k; (2) The total training time and testing time are relatively slow.
	BP Neural Network		Does not require much calculation, simple to implement;Powerful parallel capability;Automatic adjustment capability.	(1) Sensitive to the initial weight, which may make the algorithm fall into local extremum, resulting in the failure of network training; (2) The network structure determined by experience may not guarantee the accuracy of fault classification; (3) The convergence speed is slow.
	ELM		The weight $\omega$ and threshold $\beta$ of neural network do not need to be adjusted.	(1) The parameters are determined once and no longer adjusted; (2) Difficult to guarantee that the parameters are the most suitable.
	EL	Random Forest	(1) Strong anti-noise ability, not sensitive to outliers; (2) Can process high-dimensional data without feature selection; (3) Easy to realize parallelization.	For the data of attributes with different values, the attributes with more values will have a greater impact on the random forest, so the attribute weight produced by the random forest on this data is not credible.
		Adaboost	(1) Various regression classification models can be used to build weak learners; (2) The generalization has low error rate and high precision; (3) Can be applied to most classifiers without adjusting parameters.	(1) The number of weak classifiers, is not easy to set; (2) Data imbalance leads to the decline in classification accuracy; (3) Time-consuming and sensitive to outliers.
		GBDT	(1) Can flexibly handle various types of data, including continuous value and discrete value; (2) High classification accuracy and strong robustness to outliers.	(1) Due to the dependency between weak learners, it is difficult to train data in parallel; (2) Not suitable for high-dimensional sparse features.
	CNN		(1) Shared convolution kernel; (2) Can easily process high-dimensional data; (3) Feature extraction can be performed automatically.	(1) Due to the encapsulation of feature extraction, a black box is covered for the improvement of network performance; (2) The pooling layer will lose a lot of valuable information; (3) Ignore the correlation between the part and the whole.
	RNN		(1) Suitable for processing sequence data; (2) Can be used with CNN to obtain better mission effect.	(1) Gradient disappearance and gradient explosion; (2) Compared with other CNN and full connection, RNN needs more video memory space and is more difficult to train; (3) If tanh and ReLu are used as activation functions, too-long sequences cannot be processed.
Knowledge-based	FL		(1) Strong ability to express uncertain and incomplete information; (2) High degree of combination with other methods.	Difficult to obtain fuzzy rules and membership functions.
	RST			Difficult to deal with continuous data.
	D-S evidence theory			(1) Improper composition rules will cause the problem of “composition paradox” (2) There is a problem of exponential explosion in calculation.

, in which advantages and disadvantages of each method are presented.

$\mathrm{The}$$$

3. Fault Diagnosis of Key Components in Cold Ironing System

3.1. Distribution of Articles by Key Component in Cold Ironing System and Fault Diagnosis Technology

The article distribution classified by the proposed classification model is described by Figure 12, and, in

Table 2. Reference classification.

Key Component in Cold Ironing System	Fault Diagnosis Method		References
Circuit Breaker	Shallow Neural Network	ELM	[30,31]
		Probabilistic Neural network (PNN)	[32]
		BP	[33]
	Deep Learning	CNN	[34]
		SVM	[35,36,37,38,39,40,41]
		Clustering	[42,43,44,45,46,47]
		FL/RST	[48,49]
		EL	[50]
		Hybrid Model	[51,52,53,54]
Transformer	Shallow Neural Network	ELM	[55,56]
		PNN	[57]
		BP	[58,59,60,61,62]
	Deep Learning	RNN	[63,64,65,66,67]
		Deep Belief Network (DBN)	[68,69]
		Auto-encoder	[70]
		CNN	[63,71,72,73,74,75,76,77,78,79,80,81,82]
		SVM	[83,84,85,86,87,88,89,90,91,92,93,94,95]
		Clustering	[96,97,98,99,100]
		FL/RST	[101,102,103,104,105,106,107,108,109]
		EL	[110,111,112]
		Hybrid Model	[113,114,115,116,117,118,119,120,121,122]
Frequency Converter	Shallow Neural Network	ELM	[123,124]
		BP	[125,126,127,128,129]
	Deep Learning	RNN	[130,131]
		CNN	[132,133,134,135,136,137]
		SVM	[138,139,140,141,142]
		FL/RST	[143,144]
		EL	[145]

, the detailed reference classifications are displayed.

It can be found that almost 60% of the research papers on the fault diagnosis in cold ironing systems are about the transformer, indicating that research on transformer fault diagnosis has been relatively mature. Further, 25 papers on the fault diagnosis of the circuit breaker have been researched, accounting for 21.6%, and 23 papers on the fault diagnosis of the frequency converter have been studied, accounting for 19.8%.

3.2. Application of Fault Diagnosis Methods to Key Components in Cold Ironing System

3.2.1. Fault Diagnosis of Circuit Breaker

A high-voltage circuit breaker (HVCB) plays a significant role in the control and protection of the power grid. Therefore, its reliability is of great significance in the power system. As a controller for current switching, the control voltage of a circuit breaker will inevitably fluctuate in the actual working process, which will affect the data acquisition, resulting in the “sampling asynchrony” phenomenon and unstable diagnostic results. Chen & Wan [30] designed the kernel function of K-ELM based on dynamic time warping (DTW) distance and proposed an improved K-ELM method named triangular global alignment kernel extreme learning machine (TGAK-ELM), which can solve this problem and improve the diagnosis performance. However, the weights and offsets between the input layer and the hidden layer of ELM are generated randomly, making the training process unstable. Gao et al. [31] adopted integrated extreme learning machine (IELM) to classify the extracted mixed features. A group of weak ELMs generated by small samples constitute the IELM method, which makes up for the training instability in traditional ELM and, meanwhile, acquires rapidity and globality.

As a representative of shallow neural network, BP network is also widely used in the fault diagnosis of a circuit breaker [33]. Probabilistic neural network (PNN) is a radial basis function feedforward neural network based on Bayesian decision theory, which has a strong fault tolerance in pattern classification. Yao & Wang [32] implemented the accurate classification of the faults of miniature circuit breakers by using PNN. Jianzhong et al. [53] proposed a hybrid fault diagnosis method of HVCB in which BP neural network is built to automatically train the weight of sensors, and D-S evidence theory is utilized for fault identification.

In addition to the vibration signal, the current signal can also reflect the hidden health state of a circuit breaker. Zhao & Wang [37] constructed the current-vibration entropy weight characteristics and introduced grey wolf optimization into the hyperparameter selection problem of SVM, aiming at improving the fault diagnosis accuracy of circuit breakers.

Due to the advantage of solving problems with regard to small sample pattern recognition, SVM is considered as a powerful tool for fault diagnosis of power equipment. Miao Di [38] and Dou [39] implemented the fault classification by SVM after obtaining the fault characteristics of HVCB. Generally, the distinction between early failure (or minor failure) and normal state of equipment such as a circuit breaker is not obvious enough. To address this problem, a feature-based additional algorithm for fault patterns on the basis of SVM as a classifier was proposed by Rudsari et al. [35]. With the development of manufacturing technology, the equipment becomes more and more complicated, and there can be many different fault modes, making the fault diagnosis difficult. The traditional SVM cannot distinguish between unknown faults, which seriously restricts the performance of SVM in practical fault diagnosis. To solve the problem, Huang et al. [36] designed a novel classifier constructed by two one-class support vector machines (OCSVM) and an SVM. Lin et al. [40] constructed an optimal subset of hierarchical hybrid classifiers based on OCSVM and RF for state recognition.

Different from classical fuzzy theory, which only considers the degree of membership [49], intuitive fuzzy set (IFS) also takes the degree of non-membership and hesitation (representing opposition and neutrality to information, respectively) into consideration, which makes it more practical and powerful in dealing with uncertainty of faults. Combined with the reasoning method of Petri net, Zhang et al. [48] described the uncertainty of circuit breaker fault events based on IFS theory.

Clustering algorithm is also widely used in the diagnosis of circuit breakers, but simple clustering may lead to unstable fault diagnosis results [43]. In most cases, it is difficult for the objects in the dataset to specify which cluster they belong to, and there is a certain fuzziness. Combined with fuzzy theory, FCM clustering can provide more flexible results and achieve a more stable classification [47]. However, FCM is sensitive to isolated points and noise (Problem 1). Moreover, its performance relies on the selection of the initial cluster center, resulting in uncertainty of converging to the global optimal (Problem 2). KFCM algorithm can improve the clustering performance by introducing the kernel function, making the algorithm more robust to noise and isolated points, and can solve Problem 1 well [46]. Zhu et al. [42] overcame the second difficulty by using particle swarm optimization (PSO) to optimize the initial value of KFCM globally. According to the clustering logic, the methods above are all partition-based methods. There are also some other clustering methods for the fault diagnosis of the critical components in cold ironing systems. For example, He et al. [45] proposed a double clustering algorithm DPCA-KFCM in fault diagnosis tests of circuit breakers, in which density peaks clustering algorithm (DPCA) is adopted to solve the problem of selecting the initial values of KFCM. This idea provides a new way to solve Problem 2. Lu & Li [44] employed a graph-based clustering algorithm, named affinity propagation (AP), to classify faults of circuit breakers. In addition, it is worth mentioning that AP is particularly suitable for small sample cases, which is of great importance for equipment such as a circuit breaker, transformer and frequency converter due to the high cost of data acquisition.

In addition to the above methods, DL and EL also have certain applications in the fault diagnosis of circuit breakers. CNN, as a kind of DL, simplifies the complex prework of fault diagnosis, which is endowed with the ability to directly identify the signal and automatically extract features. Yang et al. [34] took the time-frequency diagram of the signal as the input of CNN and achieved an accurate condition evaluation of the HVCB’s opening damper. Compared with the DL, EL usually performs better in fault classification by constructing and combining multiple weak classifiers in fault diagnosis [50]. RF is a kind of EL method that can facilitate parallel training processes and perform well in a large-sample environment. Lin et al. [40] obtained the Gini importance of features with the help of RF. Similar to the idea of EL, we can use different classifiers to complement each other in a hybrid way to obtain better results. Wan & Chen [51] trained a hybrid classifier combining both the advantages of ELM (fast learning speed and generalization ability) and SVDD (identification ability of unknown faults). Huang et al. [52] proposed a hybrid model from SVDD to FCM, and then to SVDD. The first SVDD is designed for the dichotomy between normal and failure, then FCM is used to classify different faults and the last SVDD is utilized to judge the unknown faults.

3.2.2. Fault Diagnosis of Transformer

Dissolved gas analysis (DGA) is considered to be one of the most powerful fault diagnosis methods for power transformers [74]. Using the chromatographic analysis technique [59], we can obtain the gas concentration or production rate of different components in the transformer oil. Traditional DGA diagnosis methods, such as Duval triangle, IEC ratio, C, mainly use the ratio of gas in oil as criteria to explain the different fault types of transformers. However, the traditional method has some disadvantages, such as missing coding and improper judgment criteria [69].

In recent years, artificial intelligence technology and machine learning algorithms have been widely applied to the fault diagnosis of mechanical equipment due to their powerful nonlinear fitting ability, and so is the case for DGA analysis of transformers. For instance, clustering [97], SVM [91], BP neural network [61], K-ELM [55], DL [65,71,77] and RST/ FL [101,106] are commonly applied. Considering that each type of machine learning algorithm has its own advantages and disadvantages, the idea of EL [111] and hybrid models [113,114,115,116,117,118] has been proposed by scholars in order to achieve better diagnostic results as much as possible.

The diagnosis method based on neural network has strong nonlinear fitting and adaptive ability, which can effectively capture complex DGA features. Yan et al. [60] connected BP-Adaboost with PNN in series, which not only improves the diagnosis ability of the BP AdaBoost algorithm but also improves the diagnosis accuracy of the PNN model. However, the classification performance of PNN is greatly affected by the smooth factor of the hidden layer. Yang et al. [57] proposed an improved PNN method with a BA algorithm, which solves this challenge and enhances the fault diagnosis accuracy of transformers by optimizing the smooth factor of PNN.

The fault of transformers will bring high losses, and it rarely occurs in practice, resulting in a problem that the fault data are insufficient. Recalling that SVM performs well in processing small-sample cases, it is appropriate to take advantage of SVM in transformer diagnosis [91]. The performance of SVM mainly depends on the selection of kernel function and parameters; therefore, it motivates many scholars to use other intelligent algorithms to optimize the hyperparameters of SVM for the sake of achieving better classification results, such as particle swarm optimization [84], Krill Herd algorithm [92] and other methods [90]. Jifang et al. [93] enhanced the SVM by applying the AdaBoost algorithm to deeply explore the transformer fault data. Moreover, improved particle swarm optimization algorithm (IPSO) is used to optimize the parameters of the SVM. However, SVM is essentially a binary classifier, and, although there are several ways to implement multi-classification tasks by SVM, its limitations in dealing with this kind of problem are nonnegligible.

FL and RST can solve the multi-classification problem quickly and effectively, and they can also express multiple faults that may exist simultaneously. Peng et al. [103] described a diagnosis expert system to detect the fault of ship transformers based on RST. In general, the division of fuzzy boundaries requires a great deal of prior knowledge; thus, it is difficult to be defined. Hoballah et al. [106] proposed a novel method to solve this problem using grey wolf optimizer (GWO). Sherif & Ghoneim [107] improved the diagnosis accuracy of DGA techniques with the help of fuzzy logic theory. Combining D-S evidence theory and RST, Xu et al. [102] proposed a transformer fault diagnosis scheme based on multi-source information fusion.

Insufficient fault data will aggravate the impact of sample imbalance on the diagnosis results. The lack of recognition ability for rare faults (e.g., the mixture of dielectric and thermal faults [97]) may have a limited effect on the diagnosis accuracy, but it may have a significant impact on practical application. Dai et al. [68] tried to distinguish multiple fault types from discharge or thermal faults by DBN and obtained excellent results. Nonetheless, it is worth noting that DBN’s ability to extract feature information decreases with the increase in information. Chen et al. [69] applied fast relevance vector machine (FRVM) to separate the discharge and overheating faults of transformers before using deep belief network (DBN) to realize further fault diagnosis, which achieves higher diagnosis accuracy. In response to the dilemma of machine learning dealing with the “imbalanced classification” problem, Zhang [71] proposed a novel one-dimensional convolution neural network (1D-CNN) model based on cost-sensitive learning and verified its classification accuracy by transformer diagnostic experiments. Yuan et al. [88] used twin support vector machines (TWSVMs) as classifiers to solve problems with unbalanced and insufficient samples. However, the performance of SVM will not be improved even if the samples on common faults are sufficient, indicating that it cannot benefit much from large-scale samples. Wang & Zhang [85] proposed a combined diagnosis model by integrating a variety of fault diagnosis models to implement a preliminary diagnosis and SVM for the second step of diagnosis, which can deal with this defect to some extent.

The overall operation state of the transformer can be explained by DGA, but it has some limitations. For example, it is not sensitive to the fault of transformer winding. It can be seen from Figure 4 that winding is an indispensable component in a transformer; this is where frequency response analysis (FRA) comes in handy as another commonly used transformer diagnostic method. Abbasi et al. [96] proposed a hierarchical-based clustering method to classify FRA features of transformers. SVM is also applicable to FRA-based transformer fault analysis after optimizing the penalty factor and kernel function [83,91].

It can be found that either DGA or FRA have their own advantages, but there also exists a common disadvantage. They are not able to determine the fault location, which is crucial for the subsequent maintenance of transformers. Dey et al. [75] proposed a novel method based on CNN to identify and localize the faults of transformer winding under an impulse test.

The analysis method based on the vibration signal is one of the most popular methods in the field of mechanical fault diagnosis. It is also applicable to transformers and is capable of describing the fault location [87]. In addition, the analysis method based on the vibration signal is also applicable for condition monitoring of transformer winding [108]. However, the collected vibration signals are always complicated and even chaotic due to various external factors (such as environmental vibrations, magnetostriction and oil pumps operations), making it difficult to fully and effectively utilize the information contained therein. Zollanvari et al. [66] established GRU, LSTM models to capture the time-series hidden patterns of transformer vibration signals for early fault prediction and published the open source code. Zhang et al. [110] proposed an EL method composed of DBN, stacked denoising autoencoder (SDA) and relevance vector machines (RVM), which remedies the inadequate information of features extracted by a single type of method. Hong et al. [78] proposed a novel feature extraction scheme based on vibration analysis in which the vibration monitoring data with load information are converted into a vibration image, and then a deep learning approach based on CNN is applied to the problem of image classification.

In addition, partial discharge and arc fault are also some of the common fault forms of transformer windings. Specifically for these two problems, Do et al. [74] proposed a CNN model to classify partial discharge defects in power transformers and utilize the phase–amplitude response to reduce the input size of CNN. Zhou et al. [118] constructed the PD pattern recognition model CNN-LSTM, combining the advantages of both CNN and RNN. Yang et al. [76] converted the collected current signal into a gray image in time sequence and send it to CNN. The proposed method skips the dependence on the prior knowledge of features, realizes the arc fault diagnosis of transformers and obtains excellent feedback.

Differential relay is a basic device for power transformer protection, and inrush current is a normal transient current surge phenomenon (not a fault) caused by its working process [72]. However, the differential relay always fails to distinguish the inrush current from internal faults. Based on CNN, Afrasiabi et al. [73] developed a method to predict and classify the early faults of transformers, which improves the speed and performance of differential protection. The transformer rectifier unit (TRU) is also a crucial component in a transformer. Considering the characteristics of TRU fault modes, Chen et al. [63] developed a hierarchical deep convolutional neural network (HDCNN) for TRU fault diagnosis and proposed a method to quickly build different TRU models based on transfer learning, through which the difficulty of training models for each TRU category can be overcome.

Note that RNN was originally applied to natural language processing (NLP), dealing with unstructured data, such as text narrative type. However, it is worth attention that there is also a large amount of unstructured data that can be used in equipment fault diagnosis (i.e., malfunction inspection report), which makes it possible to extend the application of RNN to fault diagnosis. Wei et al. [64] proposed a method to apply the principles of RNN-LSTM for NLP to the fault diagnosis of a transformer in a power system.

3.2.3. Fault Diagnosis of Frequency Converter

It can be seen from Figure 5 that the frequency converter in a cold ironing system consists of a rectifying device, inverter device, filter device, etc. Among them, the inverter is used for the AC/DC conversion of port-supply current and ship-use current, and it further consists of some complex electronic components, making it more possible to suffer from different types of fault [127]. Talha [125] et al. tested nineteen different faults of three-phase inverters based on neural network, and the fault location can be obtained [126]. Huang et al. [128] proposedd a new algorithm for multiple open-circuit fault diagnosis of photovoltaic inverters, which can provide valuable reference under the background of increasing use of clean energy in ports. IGBT is another key component in inverter, and the damage of a single IGBT may lead to the failure of the whole inverter. In view of this, Hu et al. [129] proposed a novel method to detect the location of the failed IGBT.

In addition to shallow neural networks, deep neural networks are also used to diagnose open-circuit faults of inverters, such as RNN [130] and CNN [135]. Ye et al. [131] proposed a deep learning method combining wavelet packet transform and LSTM for inverter fault diagnosis, through which both open-circuit and short-circuit switch faults can be accurately detected and located. Sun et al. [133] proposed a hybrid CNN model that combines 1D-CNN and 2D-CNN, which enables CNN to own better feature extraction capabilities. Kim et al. [134] proposed an improved diagnostic method based on CNN, which greatly improves the diagnostic speed. Considering the demand of cold ironing system, DC-DC inverter is most commonly used in ship power system. Gong et al. [132] proposed a modified CNN model by using a global average pooling layer to replace part of the fully connected layer in order to increase the diagnosis efficiency of the DC-DC inverter.

Due to the suitability for small sample pattern recognition, SVM is of great significance for the fault diagnosis of frequency converters and their components. Yuan et al. [138] proposed a fault diagnosis strategy based on PCA-SVM to address the difficulty in distinguishing the two types of similar faults. Yu et al. [140] proposed a method combining sparse representation and SVM for fault feature extraction and classification, through which the diagnosis effect can be improved. However, it is difficult to implement SVM on a large scale of training samples, and we can also make use of the advantages of large samples via processing the original data [139].

Finally, rough set is another inexact reasoning method, Oliveira et al. [143,144] employed rough set theory to generate diagnostic rules for inverters. RF is a kind of ensemble learning, which can facilitate parallel training process and perform well in large sample environment. Kou et al. [145] attempt to improve the robustness of fault diagnosis classifier based on the fusion of both knowledge-driven (Concordia transform) and data-driven methods (RF).

3.3. Distribution of Articles by Fault Diagnosis Technology

Figure 13 shows the distribution of the articles from the perspective of fault diagnosis technology. It can be seen that deep learning, SVM and shallow neural network are the most commonly used methods for the fault diagnosis of key components in cold ironing systems, together accounting for about 62.9%. Note that, in deep learning methods, deep neural networks such as CNN and RNN are most widely used. Through the analysis, we find that the application of neural networks is very popular in the fault diagnosis of circuit breakers, transformers and frequency converters in cold ironing systems, which is also the key direction of the related research in the future.

3.4. Distribution of Articles by Component Type and Publication Year

Figure 14 shows the distribution of the articles from the perspective of component type and publication year. Overall, it can be seen that research work on the fault diagnosis of the key components in cold ironing systems has been growing in the last few years. Among the research papers, the fault diagnosis of transformers has attracted the most attention from the scholars, while the interest on the fault diagnosis of the frequency converter is also growing.

4. Conclusions

With the background of carbon peak and carbon neutralization, where energy conservation and emission reduction are widely advocated, the application of cold ironing technology is bound to become more and more extensive. However, due to the harsh working environment of the ports, including high temperature and humidity as well as high risk of corrosion, a cold ironing system is prone to suffer from various kinds of failures. In order to guarantee the safety and reliability of the cold ironing system, it is of great significance to focus sufficient attention on the fault diagnosis of the cold ironing system, especially those components that play key roles in the system operation. This paper conducted a comprehensive review of the fault diagnosis technology regarding the key components of the cold ironing system. Firstly, the key components, including circuit breakers, transformers and frequency converters, were introduced as well as their common failure modes. Then, we summarized the commonly used fault diagnosis technology, including both data-driven methods and knowledge-based or rule-based methods. By investigating the related literature on the fault diagnosis of cold ironing systems, this paper makes a clear classification according to the type of key component, and, for each special type of component, the fault diagnosis methods were further categorized and analyzed. Based on the study of the recent research works and analysis of the existing problems as well as the future trend, we have the following findings that may provide useful references for researchers working in this direction:

(1)

The fault diagnosis technology based on artificial intelligence and the real-time diagnosis of cold ironing equipment from the perspective of condition monitoring will still be the focus of future work. Although this idea can make full use of the collected data and dispense with the analysis of the complex working environment of ports to some extent, it should be noted that the complex port environment may give rise to some special types of failures, which are difficult to reveal under normal conditions. If the failure datasets cannot take these kinds of failures into account, it is highly possible that the research work may risk falling into a "false boom" of high precision results while ignoring the catastrophic consequences caused by rare failures. Therefore, the capability of distinguishing unknown faults is necessary for the intelligent fault diagnosis algorithm design with an emphasis on the key components in a cold ironing system. On this basis, improving the recognition rate of the unknown faults is also a key direction in the future.

(2)

As an effective way to solve the problems of imbalanced samples and label missing, deep transfer learning is worthy to be explored for the fault diagnosis of cold ironing systems. From our study, there are many research works studying the fault diagnosis of circuit breakers, transformers and frequency converters in general scenarios, while the research targeting the components used in a cold ironing system is relatively sparse. Focusing on this special application scenario, the problem of insufficient samples is nonnegligible. Transfer learning can help with taking advantage of various types of data to discover the "invariants" so as to achieve satisfactory diagnosis accuracy in the case of small samples. The combination of transfer learning and deep learning can make a more ideal fault recognizer.

(3)

A distributed fault diagnosis algorithm design based on deep learning is another challenging direction in the future. In the research of equipment fault diagnosis, there are various basic methods (such as DGA, FRA and vibration signal in transformer), while different methods often have their own advantages and limitations. Based on a clear understanding of the fault diagnosis equipment, we can construct distributed diagnosis models for the main components of the equipment, which can be integrated into a fault diagnosis model for the whole equipment by applying EL or hybrid methods. The distributed fault diagnosis models can make full use of the advantages of different methods so as to further improve the diagnosis effect of the equipment and give full play to the advantages of multi-source data.