A Fault Diagnosis and Visualization Method for High-Speed Train Based on Edge and Cloud Collaboration

Abstract

Safety is the most important aspect of railway transportation. To ensure the safety of high-speed trains, various train components are equipped with sensor devices for real-time monitoring. Sensor monitoring data can be used for fast intelligent diagnosis and accurate positioning of train faults. However, existing train fault diagnosis technology based on cloud computing has disadvantages of long processing times and high consumption of computing resources, which conflict with the real-time response requirements of fault diagnosis. Aiming at the problems of train fault diagnosis in the cloud environment, this paper proposes a train fault diagnosis model based on edge and cloud collaboration. The model first utilizes a SAES-DNN (stacked auto-encoders deep neural network) fault recognition method, which can integrate automatic feature extraction and type recognition and complete fault classification over deep hidden features in high-dimensional data, so as to quickly locate faults. Next, to adapt to the characteristics of edge computing, the model applies a SAES-DNN model trained in the cloud and deployed in the edge via the transfer learning strategy and carries out real-time fault diagnosis on the vehicle sensor monitoring data. Using a motor fault as an example, when compared with a similar intelligent learning model, the proposed intelligent fault diagnosis model can greatly improve diagnosis accuracy and significantly reduce training time. Through the transfer learning approach, adaptability of the fault diagnosis algorithm for personalized applications and real-time performance of the fault diagnosis is enhanced. This paper also proposes a visual analysis method of train fault data based on knowledge graphs, which can effectively analyze fault causes and fault correlation.

1. Introduction

The fault diagnosis of high-speed trains is crucial for ensuring train safety, diagnosing faults on site, and reducing maintenance costs [1]. Research on intelligent fault diagnosis technology has become an important part of high-speed railway technology development. As a kind of advanced technology and complex-structure technical equipment, high-speed trains utilize numerous electronic components and equipment for complex information processing. Thanks to the rapid development of sensor technology, the railway industry has collected a large amount of train monitoring data. The core of high-speed train intelligent maintenance technology is a fault diagnosis method based on train monitoring data [2]. With the rapid development of artificial intelligence, the theory of fault diagnosis based on artificial intelligence has been of great interest to many scholars [3].

To perform high-speed train fault diagnosis, fault characteristics are studied by detecting the parameters of on-board component sensors [4]. Based on large-scale historical monitoring data, the use of deep learning for fault feature extraction is an effective fault diagnosis method for high-speed rails [5]. However, the current use of machine learning for fault diagnosis possesses certain problems, such as long processing times and high demand for computing resources. Therefore, it is necessary to optimize and learn the diagnosis model according to big data obtained from train monitoring. For trains with a single car body, it is crucial to deploy the fault diagnosis model on the edge, which will help predict and identify faults, diagnose current train operation faults in real time and provide fault location [6]. With the continuous development of Internet of things technology, edge computing, cloud computing and monitoring technology, we can combine these to develop a train running condition monitoring system based on edge cloud collaboration computing. This system will send on-site train running status data to the edge computing and cloud service center in real time for analysis. The data preprocessing will be done via edge computing, and in-depth analysis and additional tasks will be completed by the cloud service center [7]. Transfer learning will then be applied to distribute the trained model to the edge end, and the edge end will be used for fault identification, which will ensure the timeliness of fault diagnosis.

At the same time, the offline monitoring data access and data analysis platform can be used for fault correlation analysis and fault category frequency analysis. Considering the visual view effect of fault analysis, the potential application value is mined from various data, which can build a more intelligent, efficient and accurate fault analysis system, and provide a basis for high-speed train health assessment.

This paper mainly explores the intelligent fault diagnosis method and high-speed train applications based on intelligent edge computing [8]. Aiming at the problems of slow calculations, poor accuracy and weak anti-interference in fault diagnosis, this paper proposes a novel real-time fault diagnosis method for high-speed trains, which is suitable for the edge deployment of single car bodies and real-time diagnosis of current running fault conditions.

2. Related Work

High-speed trains are a type of equipment utilizing advanced technology and complex structures. Safety is the top priority in high-speed railway operation [1], which can be achieved through train fault diagnosis and prediction [9]. High-speed train fault diagnosis methods can be divided into manual diagnosis, automatic test equipment (ATE) diagnosis and built-in test equipment (BITE) diagnosis. Most existing mechanical fault diagnosis technologies need extensive professional knowledge to judge and select the appropriate features in the feature extraction process [10], which requires highly skilled technicians. Additionally, many unpredictable factors exist in the mechanical operation process, making fault diagnosis even more complex. Furthermore, current technology in manual intervention fault diagnosis has great limitations, which poses another problem [3]. The ATE and BITE fault diagnosis methods are both based on equipment status. With the development of artificial intelligence and deep learning [11], machine learning has become a popular method for learning effective internal features from train operation data for fault recognition, realizing the entire intelligence process from fault feature learning to classification, reducing human intervention as much as possible, and completing train fault diagnosis in an intelligent and effective manner.

Much research has been done on high-speed train fault diagnosis methods. Fault diagnosis research based on evidence theory applies the improved DS(Dempster Shafer Theory) [12] method to trains for component fault feature extraction, which can effectively solve the fault diagnosis problems of multi-type sensors or data source fusion [13,14]. Empirical mode decomposition (EMD) train fault diagnosis [15,16,17,18] and feature dimension reduction [19,20] can greatly improve the reliability of train equipment and reduce manual maintenance costs. However, due to the strong reliability of high-speed train systems, there is limited fault data on train equipment. With the development of artificial intelligence, numerous intelligent diagnosis methods have emerged, of which intelligent algorithms for small samples have been widely utilized in high-speed train fault diagnosis. HMM (Hidden Markov Model) has also been utilized [21,22], in which a Markov chain is used to describe data changes and differences between Markov chains to achieve fault diagnosis. However, small data discrepancies cause the Markov chain to vary, so the generalization and robustness of an HMM-based train fault diagnosis method is very low. At present, a large number of scholars have introduced machine learning algorithms into high-speed train fault diagnosis. Fault diagnosis research based on support vector machines [4,23,24] uses SVM (Support Vector Machine) to train fault classifiers, which can effectively improve diagnosis accuracy. In neural-network–based fault diagnosis, the model is trained [5,25] by using train operation historical data, and the trained model is applied to fault diagnosis of unknown samples, which can provide good diagnosis and prediction results.

Every fault diagnosis method has its own advantages and limitations. In recent years, many researchers have proposed and developed multi-class fusion fault diagnosis algorithms, demonstrating that the technical advantages of these methods complement and combine with each other. Some studies [26,27,28,29,30,31] also describe the fusion of different fault diagnosis technologies, which have achieved good results in certain application fields. At the same time, with regard to fault feature extraction, there have been many improved algorithms that eliminate the dependence on domain knowledge and expert experience [32,33]; and as result, great breakthroughs have been made in the automatic extraction of mechanical fault features. However, due to increasingly prominent complex nonlinear and strong interference problems, there is no immutable general model for fault diagnosis. Although the traditional neural network is effective for fault diagnosis, certain unsolved problems remain, and the deep learning model needs to be continuously optimized and utilized in fault diagnosis research and applications [34,35]. Based on the deep neural network model, this paper proposes an improved fault diagnosis training model for high-speed train monitoring data, to further improve the accuracy and speed of fault identification.

The deep-learning–based fault diagnosis model requires a large number of labeled training data for full training before accurate fault identification can be completed [31,36]. However, the collection of labeled samples and accurate training require a great deal of computing time and resources. Based on high-speed train monitoring data, similar deep learning tasks have been conducted mostly in the cloud, with abundant computing and storage resources. However, if we only train through the cloud, we must consider the adaptability of various personalized diagnosis cases and train different application scenarios to achieve the appropriate model parameters. This, in turn, will lead to a significant increase in training parameters and heavy training tasks, which conflicts with the response time requirement of fault diagnosis. With the rapid development of Internet of things technology, an increasing amount of data are generated at the edge of the network; however, it is more efficient to process only certain data there [37]. Similarly, if we only rely on personalized application data to train through the edge end, the process will be limited by the computing power and storage capacity of the edge end, resulting in increased training time. At the same time, it is difficult to collect sufficient data to meet the needs of full model training. In mechanical fault diagnosis, experimental methods with excellent performance are often not suitable for practical applications. However, this kind of cross domain learning and knowledge transfer is still an important aspect of current research in this field [38,39]. Using prior research on cross domain learning, this study incorporates the deep learning concept of combining cloud resources with high-performance computing and storage and edge devices with strong personalized adaptability and good time limit control [40]. In this paper, a real-time high-speed train fault diagnosis method based on edge and cloud collaboration is proposed, which aims to utilize the advantages of both cloud and edge computing, and conduct real-time fault diagnosis through the coordination of computing resources and real-time requirements.

Fault data analysis is also a process of knowledge mining. Knowledge mining methods can be divided into four categories: linguistics, expert knowledge, subject and cognition [41]. When analyzing large-scale fault data, traditional diagnosis methods often face difficulties such as deep fault mechanisms, a large number of fault modes, and multi-fault information fusion. To solve these problems, it is necessary to establish fault diagnosis knowledge graphs based on information technology. Fault knowledge graphs [42] transform expert experience and diagnosis knowledge into machine-readable and permanently-stored data, which is crucial for fault data visualization.

Based on high-speed train monitoring operation data, this paper presents a deep neural-network–based optimized fault diagnosis model and high-speed train health monitoring model based on edge cloud collaboration computing. In addition, advantages of the proposed model in terms of fault identification accuracy and timeliness are compared through experiments, to provide support for the fault visual analysis and health management of high-speed trains.

3. Methods

We designed a high-speed train fault diagnosis and analysis method based on intelligent edge and edge cloud collaboration, which can conduct real-time diagnosis and fault cloud analysis based on knowledge mapping. Our method consists of three parts:

• The integration of cloud computing advantages of abundant computing resources and storage capacity and the edge computing advantage of good real-time response, which allowed us to design the overall fault diagnosis framework and model training.
• An improved fault diagnosis learning model for edge nodes (based on the deep learning model) to improve the timeliness and accuracy of fault diagnosis.
• A knowledge acquisition method for structured fault data (expressed in graph form) to solve the problem of generating the initial knowledge graph on high-speed train fault data and to improve the utilization rate of train history fault knowledge and fault diagnosis efficiency, which has a very high application value in fault reasoning and knowledge sharing.

3.1. Fault Identification System Model Based on Intelligent Edge Computing

Currently, high-speed train control systems [43] collect operation parameter data from train components in real time. Then, the operation monitoring data that have accumulated in a fixed period of time are downloaded, stored and analyzed. Train fault diagnosis and analysis are extremely time-consuming processes, and completion time cannot be guaranteed. To solve the problem of real-time train operation fault diagnosis, we analyzed the advantages and disadvantages of cloud/edge end, and designed a train fault diagnosis model based on edge intelligent computing. The overall framework is shown in Figure 1. In the model, the cloud mainly stores the training sample data and trains the universal model. The training sample data is composed of personalized train diagnosis sample data sets from train operation condition monitoring. First, based on the abundance of training samples and computing resources in the cloud, the pervasive train operation fault diagnosis model is continuously trained and updated before finally obtaining the pervasive training results. The training results can then serve as different diagnosis scenarios, which can be used as intermediate results.

The edge device collects the real-time train state data under the personalized working condition through the data sensing device, and then transmits it to the cloud for storage. At the same time, training samples of personalized working conditions are formed at the edge end to modify the cloud-to-edge end universal diagnosis model, so as to improve the applicability of specific diagnosis tasks and form a personalized model for real-time diagnosis of bearing faults. The data interaction between the cloud and edge in the framework consists of two main processes: (1) the edge will upload the locally stored personalized samples to the cloud at a specified frequency to enrich the training and verification sets of the cloud; (2) with the updating of data sets stored in the cloud, the universal diagnosis model parameters will be optimized through training, and the optimized model parameters will then be transferred to the edge to modify the model.

The task of the edge end is to diagnose the train fault in real time and transfer the universal model parameters of the cloud to the edge end through transfer learning. In the edge controller, the local samples are used for personalized training of the diagnosis model, and the personalized diagnosis model is then generated, which can be applied to the edge. At this time, the real-time signal acquisition is loaded on the personalized diagnosis model, diagnosis results can be generated in real time, and the corresponding response can be made according to the diagnosis results. The fault diagnosis method in this paper can be generalized for the fault diagnosis of all key train components, which provides support for fault prediction of the entire train. In addition, the proposed fault visual analysis method can effectively analyze fault reasons and relationships between faults.

At the beginning of the task, we used the edge data and open data set stored in the cloud for the universal training of the diagnosis model. During training, the samples were divided into training set and verification set according to proportion. After training a single SAE (stacked auto-encoder) model, the global SAES-DNN (stacked auto-encoders deep neural network) was trained greedily layer by layer, and parameters of the whole network were fine-tuned. The whole diagnosis process is shown in Figure 2. For the fault data identified in the cloud, fault analysis based on knowledge mapping was carried out to provide support for the fault cause and correlation analyses.

3.2. Introduction of the SAES-DNN Model

Compared with shallow machine learning, deep multi-layer models are better at reflecting essential data characteristics. The characteristics of a deep belief network [44], for example, include easy training and learning optimization as well as advantages of a deep network structure. Based on the DBN (Deep Belief Nets) network structure, we developed a deep learning neural network model for train fault diagnosis by using a stacked auto-encoder [45] composed of a sparse auto-encoder and DNN network. We define the network model as SAES-DNN, where SAE refers to the sparse auto-encoder, and network refers to the deep network composed of multiple sparse auto-encoders. The network structure is shown in Figure 3. The SAES-DNN model contains multiple hidden layers and is composed of a series of SAEs (sparse auto-encoders) stacked in a certain way. For each SAE unit, considering that the input data is the sampling sequence of the train sensor in a certain period of time, we designed the middle layer of the network in the form of a full connection.

Compared with the traditional signal feature processing method, the encoder has better feature expression technology, which can greatly reduce the data integrity labeling requirement in the process of neural network training. A single sparse auto-encoder can minimize the reconstruction error by adding a penalty factor. Our SAES-DNN model uses the feature of a sparse auto-encoder to learn hidden information in the process of data reconstruction and keep extracted features in the hidden layer. The automatic encoder was then trained greedily layer by layer, and the data was reconstructed through the decoder decoding process to minimize reconstruction error. A sparse auto-encoder is a kind of derived automatic encoder with specific functions due to added constraints. The mechanism of decreasing the number of neurons can retain the original function in a quicker and more efficient manner. Since original train fault data information is usually miscellaneous and high-dimensional, a sparse auto-encoder is used as the AE portion in our model. For the SAEs, we define the sample data set as $\left\{x_1,x_2,\dots ,x_n\right\}$ and define the average activity of the j neuron in its hidden layer as follows:

$${\widehat{\rho }}_j=\frac{1}{N}{\displaystyle \sum }_{i=1}^N[a_j(x^{(i)})]$$

(1)

where $a_j\left(x^{\left(i\right)}\right)$ is the activation degree of a sample on the number j hidden neuron. To ensure that the hidden layer meets the sparsity constraint, we tried to make ${\widehat{\rho }}_j$ = ${\rho }_j$, with ${\widehat{\rho }}_j$ representing the sparsity parameter, which was set to 0.06 in this study. To achieve constraint adjustment, an additional penalty factor was added to the optimization objective function based on the relative entropy method, which is defined as follows:

$${\displaystyle \sum }_{j=1}^{m}KL\left(\rho \parallel {\widehat{\rho }}_j\right)={\displaystyle \sum }_{j=1}^m\left(\rho \log \frac{\rho }{{\widehat{\rho }}_j}+\left(1-\rho \right)\log \frac{1-\rho }{1-{\widehat{\rho }}_j}\right)$$

(2)

where $\mathrm{m}$ is the number of neurons in the hidden layer, and index j represents the number $j$ neuron in the hidden layer. When ${\widehat{\rho }}_j={\rho }_j$, the penalty factor $KL$ is the minimum. The overall loss function of the sparse auto-encode is defined as follows, where $\beta$ is the weight coefficient of the sparsity penalty factor. The definition of the loss function of a single sparse auto-encoder is shown in Equation (4), where $\beta$ is the weight coefficient of the sparsity penalty factor, and $J\left(W,b\right)$ is the definition of the loss function without sparsity constraint, which is composed of the mean square error term and the weight attenuation rule term.

$$J\left(W,b\right)=\left[\frac{1}{N}{\displaystyle \sum }_{i=1}^N\left(\frac{1}{2}{\parallel h_{W,b}\left(x^{\left(i\right)}\right)-y^{\left(i\right)}\parallel }^2\right)\right]+\frac{\lambda }{2}{\displaystyle \sum }_{l=1}^{n_l-1}{\displaystyle \sum }_{i=1}^{s_i}{\displaystyle \sum }_{j=1}^{s_{i+1}}{\left(W_{ji}^{\left(l\right)}\right)}^2$$

(3)

$$J_{\mathrm{sparse}}\left(W,b\right)=J\left(W,b\right)+\beta {\displaystyle \sum }_{j=1}^{m}KL\left(\rho \parallel {\widehat{\rho }}_j\right)$$

(4)

We added a softmax supervised classifier based on SAE unsupervised training to form the deep learning SAES-DNN method in this paper. For each class of vectors, the softmax function output formula and loss function of the whole network model are defined as follows:

$$p_i=p\left(Y|Q\right)=\frac{e^{w_{i}Q+b_i}}{{{\displaystyle \sum }}_{k=1}^N{e}^{w_{k}Q+b_k}}$$

(5)

$$J\left(w,b\right)=-\frac{1}{m}\left[{\displaystyle \sum }_{j=1}^m{\displaystyle \sum }_{k=1}^{N}l\left\{Y_j=i\right\}\log \frac{e^{w_i{Q}_j+b_i}}{{{\displaystyle \sum }}_{k=1}^N{e}^{w_k{Q}_j+b_k}}\right]$$

(6)

where {$Y_j=i$} is the indicative function, and the true value is 1 for the output with consistent fault and high probability, otherwise it is set to 0. $Q$ is the softmax output vector, and $w_i$ and $b_i$ are the weights and offsets that softmax needs to learn.

The softmax classifier is a generalization of multi-classification problem solving based on logistic regression. The output probability value of each class of feature vectors can be evaluated by Equation (5), and the SAES-DNN loss function is shown in Equation (6). The basic training process of the SAES-DNN model is as follows:

• Original input data was preprocessed, and network structure was initialized.
• Pre training. The encoder was trained greedily layer by layer, and data was reconstructed by a decoder decoding process to minimize reconstruction error.
• The output layer of the trained stacked automatic encoders was discarded and the middle-hidden coding feature layer was reserved.
• Output data of the last layer of the automatic encoder (the extracted hidden layer of the feature representation) was taken as the input data of the classifier.
• Fine tuning. According to the output and expected error, the parameters were fine-tuned by a back propagation algorithm to optimize the performance of the whole deep neural network. We used a set of samples to train the overall SAES-DNN network parameters ($w_i$,$b_i$), calculate the activation value of each layer in the network, and obtain the final output through the softmax model of the last layer. Using the sample label and network output, the cost function was formed, which is shown in Equation (6). The partial derivatives of each parameter in the cost function were then obtained, and the gradient descent method was used to iterate to the convergence of the cost function. In each iteration, the derivative of the cost function with respect to the network parameters of the first n−1 layer was calculated according to the forward backward propagation algorithm, while the derivative of the cost function with respect to the parameters of the last layer was obtained by Equation (6).

3.3. Visual Analysis Method Based on the Fault Diagnosis Model

Knowledge extraction consisting of entity extraction, relation extraction and attribute extraction was carried out for the processed train operation fault data. The entity triples were extracted from the existing relational database and transformed into a graph database, so as to construct the equipment entity atlas. Through knowledge extraction technology, entities, relations, attributes and other knowledge elements can be extracted from the semi-structured and unstructured data related to fault disposal. Knowledge fusion eliminates the ambiguity between reference items such as entity, relationship, attribute and fact object, and forms a high-quality knowledge base. Knowledge processing aims to integrate, refine and evaluate extracted knowledge. Event extraction extracts and expresses knowledge from fault history records. The structure of the fault graph generation is shown in Figure 4.

In our study, we used a feature-based method for the entity extraction. A pre-labeled entity corpus was used to train the model, so that the model could learn the probability of a word as an entity component, and then calculate the degree of a candidate field as an entity. The goal of relation extraction is to solve the problem of semantic connection between entities, so as to weave massive entities into a graph. The distance supervision method is used to label the training samples automatically. The entity and topology in the structured database are regarded as prior knowledge to pre-label the corpus. The relations that need to be extracted from the operation fault data text of the train operation fault disposal knowledge map include causality, compliance relationship and so forth.

With regard to the train operation data fault disposal, a train fault map was constructed based on ontology, which mainly included entity ontology, operation ontology, state ontology (regarding device concept) and connection topology. Operation ontology mainly refers to the collection of all actual actions of entity ontology, such as running, maintenance, parking, breaking equipment, etc. State ontology is the summary of the current operation status of all train equipment, such as running, speed, etc. Knowledge discovery can start from existing entity and relation data in the knowledge base, the mining of new entities, or the establishment of new associations between entities through machine learning, so as to expand and enrich the knowledge network. In different dimensions, entities are linked to represent the complex semantic relationship between entities. The model defines the evaluation functions for each triplet (h, r, t) in the knowledge base, and the evaluation function is defined as follows:

$$f_r\left(h,t\right)={\mu }_r^{T}g\left(l_h{M}_r{l}_t+M_{r.1}{l}_h+M_{r.2}{l}_t+b_r\right)$$

(7)

where $\mathsf{\mu }\mathrm{rT}\in \mathrm{Rk}$ is the vector representation of the relation R, $g\left(R\right)$ is a $\tan h$ function, $M_r\in$ Rd*d*K is a third-order tensor, and $M_{r.1}$, $M_{r.2}$, $\in$ Rd*K are two projection matrices defined by relation R.

4. Discussion

This paper mainly discusses the following aspects based on the cloud-edge collaboration framework:

• Whether the algorithm had an obvious effect when compared with CNN [46], WDCNN [13], MLNN [47], SAE-DBN [14] and so forth.
• Whether transfer learning can improve the accuracy of fault diagnosis and effectively save training time and resources during the bearing fault diagnostic task.
• Effect verification of the visual effect analysis of high-speed train operation state data based on knowledge extraction.

The experimental platform designed in this study consisted of cloud and edge computing. The hardware system in the cloud was mainly composed of six gtx1080 GPUs and an expanded 1T hard disk. The edge controller was a NVIDIA Jetson TX2. Ubuntu was selected as the cloud and edge operating system. The deep learning framework was built based on Python, which was also used to complete the implementation of the whole model training. SIMPACK was used on the computational dynamics of the multi-body system, including a number of professional modules and virtual prototype development system software used in professional fields. In this study, a motor running fault was taken as the research object, and the model of a 300 kW permanent magnet synchronous motor was established by SIMPACK.

4.1. Implementation and Performance Research of the Cloud Fault Diagnosis Algorithm

In this study, the machine learning fault feature representation was established through the simulation of motor vibration signals under different fault types, and the acquisition and processing of vibration signals [48]. The experimental analysis object mainly used a differing load output voltage signal after the rectifier circuit failed. To effectively test the experimental results of the subsequent fault diagnosis method, it was necessary to code the motor fault types, as shown in

Table 1. Calculation of the training accuracy of different models in the same data set.

Diagnostic Model	Network Depth	Learning Rate	Minimum Accuracy	Highest Accuracy	Average Accuracy	Average Loss Rate
MLNN	4	0.1	61.5%	92.2%	85.7%	11.3%
CNN	4	0.01	82.9%	94.6%	90.7%	7.62%
SAE-DBN [14]	6	0.1	92.9%	97.3%	95.6%	5.56%
WDCNN [13]	4	0.1	93.3%	97.2%	95.3%	6.41%
SAES-DNN	6	0.1	93.8%	98.1%	96.5%	5.43%
BP-MLNN	4	0.1	78.6%	92.5%	89.5%	9.69%

. According to fault location, the traction motor fault types were divided into a—stator fault, b—rotor fault, c—air gap eccentric fault and d—bearing fault. The fault types were then further divided into 15 types: a, b, c, d, a-b, a-c, a-d, b-c, b-d, c-d, a-b-c, a-b-d, a-c-d, b-c-d, and a-b-c-d. Among them, the last 11 types were combined fault models, such as a-b, which indicated that both stator and rotor faults occurred at the same time.

SIMPACK software was used to simulate different fault states, run the motor, obtain the operation data, and synthesize the data set with 15 fault types and one normal state type. The structure of the SAES-DNN model was determined by repeated optimization, and a six-layer network was designed. The input layer was 256 neurons, the two hidden layers were 65 neurons, and the output layer was 16 neurons. The learning rate was 0.1, and the number of iterations was 60 times. At the same time, we compared the accuracy and running time of fault diagnosis under CNN, WDCNN, MLNN, SAE-DBN, BP-MLNN and SAES-DNN models based on different numbers of training samples. The initial parameters of the six models are shown in Table 1. It can be seen from Figure 5a,b that the SAES-DNN method achieved the minimum calculation time and the highest accuracy under the same computing resources and data, which shows that the SAES-DNN method has better feature learning and diagnostic ability than traditional deep learning methods.

The above results also demonstrate that our method has slight advantages in classification accuracy and time loss over 10 iterations. Compared with SAE-DBN and WDCNN, the accuracy of the proposed method improved by 0.8% and 0.9%, respectively. Through the cloud training samples, our method shows advantages in accuracy and training time for the fault classification model. Higher accuracy and time efficiency can improve fault diagnosis and maintenance, and more efficient model construction time can allow the method to be applied to other mechanical parts in a shorter time [49]. To study the effects of increased sample data or expanded fault types on the proposed method (which would be useful for practical applications), the text training model must first be updated. Furthermore, further research would be needed to determine the degree of influence.

4.2. Transfer Learning of the Fault Diagnosis Algorithm

In this study, the GPU server in the cloud based on the SAES-DNN algorithm was used to train the train simulation motor fault data set and obtain the universal motor fault diagnosis model. In the motor bearing data set of Case Western Reserve University [50], some fault data are similar to the fault types of the transfer learning source domain data set. Therefore, we used this data set to simulate the personalized motor bearing data collected at the edge end.

Through the migration of universal model parameters in the cloud and the personalized adjustment and training at the edge, the model at the edge end completed the diagnosis and recognition of the motor bearing fault. Due to the high frequency of data acquisition used at the edge end, part of the data was intercepted as the input of the model to diagnose the bearing fault; however, by identifying the edge end model, the state of the motor bearing was quickly fed back. The appropriate transfer strategy was chosen by comparing fine tuning, freezing and training. The gradient design of the training samples in the target domain is shown in

Table 2. Sample size for each target domain training set in the experiment.

Training Set	Number of Samples for Each Type of Fault	Total Number of Fault Samples of Each Type	The Proportion
Train set 1	40	2480	1:62
Train set 2	60	2480	1:41
Train set 3	90	2480	1:28
Train set 4	140	2480	1:18
Train set 5	210	2480	1:12
Train set 6	320	2480	1:8

. In each experiment, 160 groups of untrained labeled data were selected as the verification set for verifying fault diagnosis accuracy.

Fine tuning refers to applying the weight parameters of cloud training directly to the edge end data, using the edge end training data for model training at the edge end, and then updating the parameters to all layers. Freezing and training kept all layers frozen except for the parameters of the last SAE and the softmax fully connected layer. The training sample gradient set in Table 2 was used to train the model, and only the parameters of the last SAE and softmax layers were updated in the training.

When using the transfer learning method, we first used the source domain samples (simulation samples) to train the improved SAES-DNN algorithm in the cloud, before finally obtaining a model with a fault prediction accuracy of >95%. All model parameters were transmitted to the edge controller. We used fine tuning [51] and freezing and training [52] for the migration learning; that is, the model parameters transmitted from the cloud were directly loaded at the edge. The initial model at the edge end was trained with the gradient of training samples set in Table 2, and parameters of all layers were updated in the training. Each experiment was repeated 10 times, and the final result was the average of the 10 experimental results, as shown in

Table 3. Effect comparison of the transfer learning strategies.

Transfer Strategy	Training Set	Train Times
		20	30	40	50	60	70	80	90	100
No-Transfer	Train set 1	21.3%	29.6%	38.5%	47.3%	48.6%	48.9%	50.1%	50.2%	49.7%
No-Transfer	Train set 2	32.5%	39.8%	49.2%	56.3%	64.0%	66.1%	66.4%	67.2%	63.1%
No-Transfer	Train set 3	42.6%	60.3%	65.8%	69.7%	71.3%	74.5%	75.4%	74.1%	73.2%
No-Transfer	Train set 4	51.9%	63.3%	70.5%	74.0%	77.7%	78.2%	79.0%	79.2%	76.3%
No-Transfer	Train set 5	63.0%	72.8%	79.1%	84.3%	86.9%	87.4%	87.1%	86.7%	84.1%
No-Transfer	Train set 6	72.3%	79.1%	85.6%	87.2%	90.1%	89.9%	90.3%	89.4%	89.7%
Fine-tuning	Train set 1	63.5%	65.7%	66.4%	67.9%	67.8%	67.9%	68.0%	68.1%	67.9%
Fine-tuning	Train set 2	70.3%	76.6%	77.1%	78.0%	78.6%	79.2%	78.8%	78.2%	77.7%
Fine-tuning	Train set 3	76.7%	77.7%	78.6%	79.4%	80.2%	80.1%	80.0%	79.7%	79.2%
Fine-tuning	Train set 4	78.9%	81.8%	85.6%	86.0%	86.9%	87.3%	87.4%	87.2%	86.4%
Fine-tuning	Train set 5	81.4%	84.7%	88.6%	89.1%	89.2%	90.0%	90.3%	89.9%	89.0%
Fine-tuning	Train set 6	84.8%	88.6%	89.7%	90.1%	90.3%	91.4%	91.0%	90.9%	90.7%
Freeze	Train set 1	50.4%	55.3%	56.1%	56.5%	56.1%	56.8%	56.9%	56.7%	56.4%
Freeze	Train set 2	54.3%	57.1%	58.6%	59.9%	59.9%	60.8%	61.0%	60.9%	59.6%
Freeze	Train set3	58.9%	61.6%	61.8%	62.4%	62.8%	62.8%	62.9%	63.3%	63.1%
Freeze	Train set 4	59.4%	62.3%	63.1%	63.5%	64.1%	64.5%	64.9%	64.7%	64.4%
Freeze	Train set 5	60.7%	63.4%	64.3%	64.8%	65.4%	65.6%	65.9%	65.8%	65.7%
Freeze	Train set 6	61.9%	64.1%	65.2%	65.7%	66.1%	66.2%	66.7%	66.1%	66.0%

.

By comparing the experimental verification and analysis, we can draw the following conclusions:

• When there is a small number of training samples, the application of transfer learning can significantly improve the diagnostic accuracy, especially when the number of samples is less than 120.
• With a small number of training samples, the effect of using transfer learning to shorten the training time is significant.
• In this task, the accuracy of fine tuning was higher than that of freezing and training.

To achieve good transfer learning effect, while selecting transfer learning samples, we should strive to obtain as many different fault performance samples as possible in the source data domain. Our experiment results showed that the motor fault data sets of different working conditions stored in the cloud can enhance the transfer learning effect.

4.3. Effect Verification of Fault Visual Analysis

In the long-term process of train operation, trains accumulate a vast amount of state monitoring data, which are large in quantity and high in dimension. Traditional relational databases of train operation data are transformed into graph databases based on triple representation via knowledge mapping technology. Triple refers to a general representation of a knowledge graph, that is G = (E, R, S), where E is the entity set in the knowledge base, R is the relation set, and S is the triple set. In this study, a train operation fault knowledge map was generated for the train operation data set, as shown in Figure 6. The structured data in our operation data set mainly included temporary repair and wheel change records, extended three-level repair records, retrofit records, advanced repair records, loading records, factory resumption information, vehicle operation fault history records, vehicle running records, advanced repair history records, vehicle operation conditions, and so forth. In the train fault data visualization method based on knowledge mapping, fault nodes were drawn as points, and fault relationship links were rendered as lines. This graphical representation can directly represent the train operation data, which is helpful for obtaining a better understanding of the entity relationship and structure information in the fault data.

Through Figure 6, we can see that the associated maintenance information and operation information were presented intuitively based on different train entities. With the help of the generated train fault graph, the collaborative filtering analysis model was used to calculate the similarity of different train fault causes, sort the fault causes, and determine the most similar train fault causes, which can help people quickly obtain intelligent recommendations with regard to the fault causes. At the same time, the fault knowledge graph can also be used to calculate the similarity of different train faults, determine which set of train faults are caused by the same reason, and analyze correlations between faults.

4.4. Limitations and Future Research

In this study, we introduced an efficient train fault diagnosis model and an extended train fault visual analysis method. To examine the real-time effects and accuracy of the method, applications in fault prediction and diagnosis of key components such as gearboxes, bogies and pantographs will be explored in the follow-up research, to provide a highly feasible implementation scheme for fault diagnosis and predictive maintenance. As a novel fault diagnosis and visualization method, there were some limitations in our study, which we plan to address in our future work.

Regarding the fault diagnosis model, we are currently using the DNN network composed of stacked auto-encoders. In the single SAE model, since the network input is the sensor sampling data, we used a full connection layer in the model. When prolonging the sampling time or increasing the sampling frequency, the dimension of input data parameters will inevitably increase, which will lead to difficulties to the network training. It is possible to transform the SAE internal network structure and reduce the dimension of high-dimensional data and network parameters (such as using a combination of CNN network and SAE), which is one of our future research directions. For fault visual analysis, as an extension of fault diagnosis, analysis of the whole fault type dimension can assist people in quickly analyzing fault causes. In the case of small-scale data sets, our current method can effectively analyze train faults, fault causes and correlations between faults; however, for large-scale data, it is a great challenge to display global data in a limited view space. Therefore, we need to conduct further study of multi-view layout algorithms and graph data visual analysis methods.

5. Conclusions

In this paper, a train fault diagnosis model based on deep neural networks and intelligent edge computing was proposed to realize the real-time diagnosis of high-speed train faults. The model took a train motor fault as the research object, used a large amount of cloud-stored sample data to train and fine tune the SAES-DNN diagnosis model layer by layer, and generated a universal diagnosis model suitable for similar train fault diagnosis tasks. Through transfer learning, the trained universal model was transferred to the edge end, and a small number of samples at the edge were used to fine tune the model to realize real-time train motor fault diagnosis. The experimental verification and comparative analysis showed that the bearing fault diagnosis method can be used to train the diagnosis model in the cloud, and the edge only needs a small amount of personalized adjustment training, which can save a lot of training time. At the same time, based on the cloud-computed train operation fault data, we designed a visualization method of train fault analysis data based on knowledge mapping, which can effectively present and analyze train operation faults. Due to the good outcomes of this method with regard to real time, accuracy and sample limitations, our follow-up study will further explore applications in the fault diagnosis of major train parts, such as bogies and pantographs, so as to provide a highly feasible implementation scheme for the fault diagnosis and predictive maintenance of high-speed trains.