A high-speed train bogie fault diagnosis method based on residual neural network

Abstract

The application discloses a high-speed train bogie fault diagnosis method based on a residual neural network, and belongs to the technical field of high-speed train fault diagnosis. The high-speed train bogie system is divided into a power bogie and a non-power bogie, and the complex bogie system may induce various faults. Firstly, a plurality of different types of sensors are installed in a train bogie state monitoring system, state monitoring and fault identification are performed on important components, an RS-GAT fault diagnosis model is established, fault diagnosis is performed on each part of the bogie, and state information of the train bogie is transmitted to a ground data center in real time through a communication network, off-line model training is performed by the data center, and high-speed train bogie fault diagnosis is performed, so that the safety of high-speed railway train operation in a journey is ensured.

Description

Technical Field

[0001] This invention belongs to the field of high-speed train fault diagnosis technology, and specifically relates to a high-speed train bogie fault diagnosis method based on residual neural networks. Background Technology

[0002] With the continuous expansion of my country's high-speed railway network, the number of high-speed trains in my country has also gradually increased, making the safety of high-speed train operation a crucial research issue. The train bogie, a key component of high-speed trains, is the only structure connecting the car body and the track. It not only bears the weight of the train body and controls the relative distance between the car body and the track to ensure the stability of the car body, but also plays a role in traction, braking, and cushioning. In other words, the performance of the bogie directly affects the safety and comfort of high-speed train operation. With the rapid development of high-speed railways, the operating speed of high-speed trains is increasing, and the single-trip time is getting longer, which leads to an increasingly longer service life for each train. During long-term service, since the bogies mainly house mechanical and electrical equipment, prolonged operation will cause wear and tear and aging problems over time. If a high-speed train's bogie malfunctions, it will seriously affect the safety of train operation. The health condition of the train bogie is directly related to the safety of train operation.

[0003] The biggest challenge currently is how to utilize data from high-speed train condition monitoring systems to determine the operating status of train bogies and to diagnose and warn of bogie faults. The most feasible method at present is to monitor the entire operation of the train bogies, analyze and process the monitoring data, understand the safety status of each component on the bogie, and take appropriate measures to address bogie faults. To implement this monitoring, the China Academy of Railway Sciences has installed a train bogie condition monitoring system on high-speed trains. This system consists of multiple sensors of different types, which transmit the bogie status in real time to a ground data center via a communication network. The data center then performs offline model training to diagnose bogie faults. Summary of the Invention

[0004] The purpose of this invention is to propose a fault diagnosis method for high-speed train bogies based on residual neural networks, characterized by the following specific steps:

[0005] Step 1, Bogie Fault Analysis and Classification

[0006] High-speed train bogie systems are divided into powered bogies and non-powered bogies. Powered bogies include power units such as motors and gearboxes. Due to the complex structure of the bogie system, there are numerous causes and diverse types of failures that can lead to malfunctions in its various components. Furthermore, the bearings in powered bogies bear relatively heavy loads, making them more prone to failure than non-powered bogies. To ensure train safety, it is necessary to conduct timely condition monitoring and fault identification of critical components of the train bogie system, classifying fault states and analyzing their potential causes. These fault states are categorized into seven types: healthy state, frame failure, bearing failure, wheelset failure, gearbox failure, traction motor failure, and air spring failure, encompassing 24 sensor measurement points with the most significant impact.

[0007] The seven types of faults include 24 sensor measurement points with the most significant impact. The specific fault names are as follows:

[0008] Structural failures: cracks, wear, bending deformation;

[0009] Bearing failures: fatigue spalling, wear, plastic deformation, corrosion, and galling;

[0010] Wheelset failures: surface wear, surface peeling, axle cracks, axle overheating;

[0011] Gearbox malfunctions: broken teeth, pitting, wear, scratches.

[0012] Traction motor malfunctions: short circuit, excessive voltage, rotor breakage;

[0013] Air spring malfunctions: cracking, weakening, corrosion;

[0014] Step 2, Select bogie status data

[0015] During actual train operation, there are many factors that may affect bogie failure. The degree of failure varies under different operating conditions such as starting, braking, acceleration, and deceleration. To improve the accuracy of bogie failure diagnosis, it is necessary to selectively select the data most relevant to the occurrence of bogie failure from various bogie detection data and perform feature fusion. Therefore, Pearson coefficient is used for feature selection, and the threshold of Pearson coefficient is 0.8 when selecting relevant features. Through screening, 24 sensor measurement points with the most significant impact on Table 1 are obtained.

[0016] Step 3: Establish the bogie data network architecture.

[0017] Since the train is structurally integrated, there is spatial correlation between the various measuring points on the train. Using 24-dimensional sensor status data from 24 sensor measuring points as input, a bogie fault diagnosis model is established, outputting seven types of fault states; this model is a typical multi-sensor fusion fault diagnosis model.

[0018] Step 4: Use grey theory to determine the framework of the graph structure, and use the Granger causal analysis method (GGC method) to determine the direction of the graph edges; the specific steps are as follows:

[0019] Step 1: Construct the sensor information matrix X = (x ij ) n×t Calculate the value x for each comparison sensor. j The correlation coefficient ζ between the corresponding elements and the reference sensor ij (k); the specific formula is as follows:

[0020]

[0021]

[0022] Where n is the number of sensors, t is time, and the sensor values ​​in the 1st and 1st columns at time j are x, respectively. 1j and x ij ρ is the correlation constant.

[0023] Step 2: Calculate the correlation degree r′ ij As shown in the following formula:

[0024] Where σ j Let be the correlation coefficient between all corresponding elements of the sensor values ​​at time j.

[0025] Step 3: Determine the weight matrix W = (w ij ) n×n If the weights of the graph structure are defined by the correlation coefficients between sensors, then the formula for W is as follows:

[0026] W = (w ij ) n×n =(r ij ) n×n

[0027] The fourth step is to define the constraints of the high-speed train diagram structure. Three constraints are selected: data relationship R1, system to which it belongs R2, and carriage to which it belongs R3. These constraints work together on the sensor nodes, and a weighted method is used to determine the final association matrix R of the high-speed train diagram structure. The specific formula is shown below:

[0028]

[0029]

[0030]

[0031] i = 1, 2, ..., n

[0032] j = 1, 2, ..., n

[0033] R = (r ij ) n×n = w1·R1 + w2·R2 + w3·R3

[0034] w1 + w2 + w3 = 1

[0035] Where w1 represents the weight of data relationship R1, w2 represents the weight of system R2, and w3 represents the weight of carriage R3.

[0036] Step 5: Determine the direction of the edge. In the time series of the high-speed train sensor measurement points, if the vehicle information measurement point v... i The value X i Within 1 second, the measuring point v can be measured. j The value X j To have an effect, that is, to observe X i The information change can explain the X that appears after 1 second. j The change in information indicates that X i With X j There is a causal relationship between the two in the data, namely X i →X j ; or by extension, v i and v j There is a causal relationship between them, i.e., v i →v j e ij =1 Given that the time series of sensor measurement points on high-speed trains is a stable series, this invention employs Granger causality testing based on a VAR model to establish (X i X j The VAR model is shown below:

[0037]

[0038]

[0039] Among them, X i (t) and X j (t) represents the value of the sensor measuring point on the high-speed train at time t; X i (tk) and X j(tk) represents the value of the sensor measurement point on the high-speed train at time tk, and the value space of k is [0, l]; a k and c k It is the Granger causality coefficient; b k and d k It is the autoregressive coefficient; ε t and μ t This represents the prediction error, which is independent of the time point and is assumed to be white noise.

[0040] If X i (t) is not X j If (t) is a Granger cause, then a1 = a2 = ... = a l =0. The Granger causality test is performed using a constrained F-statistic, which specifies that X = 0. i (t) The sum of squared residuals is RSS U X j The sum of squared residuals of (t) is RSS R The F-statistic test formula is as follows:

[0041]

[0042] If the F-test conclusion rejects the null hypothesis H0: a1=a2=...=a l =0, then sensor i is the Granger cause of sensor j, that is, from vertex v i To vertex v j It is a directed edge; otherwise, sensor i is not a Granger cause of sensor j. Similarly, if the F-test conclusion rejects the null hypothesis H0: c1=c2=...=c l =0, then sensor i is the Granger cause of sensor j, that is, from vertex v j To vertex v i It is a directed edge. Furthermore, the stronger the Granger causality, the smaller the value of F. Therefore, the edge E = (e...) in the high-speed train graph structure... ij ) n×n The following conditions must be met:

[0043]

[0044] Step 5, RS-GAT fault diagnosis model

[0045] Based on the RSNet model framework, the directed graph for bogie fault diagnosis is input, and the original convolutional layer is replaced with GAT, which has stronger aggregation capabilities, to construct the RS-GAT fault diagnosis model. Since GAT already has aggregation capabilities in graph neural networks, the RS-GAT model does not use the Squeeze operation in the RSGAT unit, but uses the GAP operation in the final output to ensure that the average value of all input feature maps can be taken as the output.

[0046] Among them, the RSNet model is a CNN model designed based on the residual learning framework of ResNet, inspired by the DCNN model and the Squeeze operation of SENet, and its input is a one-dimensional time-series signal.

[0047] The RSNet model uses the ResNet network as its framework, replacing the original residual learning units with convolutional layers. Each convolutional layer includes a BN layer, a ReLU layer, and a convolution operation. After passing through three convolutional layers with 16 kernels and three convolutional layers with 32 kernels, the RSNet model performs global average pooling (GAP) on all features. Finally, it passes through a fully connected layer (FC) and a softmax operation to obtain the final result.

[0048] The RS-GAT model can be divided into three parts:

[0049] The first part consists of the input cells for the RS-GAT model. The model input consists of the directed graph data X for bogie fault diagnosis and the graph structure A. After passing through a GAT, the normalized attention coefficient a of node j with respect to node i is obtained. ij The specific calculation formula is as follows:

[0050] e ij =Attention(Wx i Wx j =LeakyReLU(w T [Wx i ||Wx j ])

[0051]

[0052] Among them, LeakyReLU(w T [Wx i PWx j ]) is the activation function, exp(LeakyReLU(w T [Wx i PWx j])) represents an exponential operation, and the data at time i and time j are x. i and x j The first-order neighborhoods of nodes i and j are and The learning parameter is W, and the training parameter is w. The attention coefficient a is used. ij Compute the reconstructed vector x of node i i The calculation formula is as follows:

[0053]

[0054] Where f(·) is a non-linear activation function. Multiple attention units are concatenated to improve the feature capture capability of node i. Specifically, the following equation applies:

[0055]

[0056] Where || represents the concatenation operation, and K represents the number of attention units.

[0057] The second part consists of l RSGAT units. The RSGAT unit is designed with six layers based on the residual network framework: BN layer, ReLU layer, GAT layer, BN layer, ReLU layer, and GAT layer. Taking the first layer as an example (l=1), the residual can be calculated as F1, and the output is H1=F1+X0. After every three RSGAT units, the output H matrix is ​​pooled, and then the graph attention units of the next three RSGAT units are doubled. After processing through l layers (l must be a multiple of 3) of RSGAT units, the final output matrix H is obtained. l .

[0058] The third part is the output unit of the RS-GAT model. The H output in the second part... l It contains a large amount of fault characteristic information, H l Pooling operations are performed, followed by GAP, FC, and Softmax operations to finally obtain the fault classification results.

[0059] The beneficial effects of this invention are that it installs multiple different types of sensors in the train bogie condition monitoring system to diagnose faults in various parts of the bogie, obtain fault characteristic information, and transmit the train bogie condition information to the ground data center in real time through a communication network. The data center then performs offline model training to diagnose faults in the high-speed train bogie, thus ensuring the safety of high-speed trains during operation. Attached Figure Description

[0060] Figure 1 This describes the distribution of sensors on the bogie.

[0061] Figure 2 For GGC method flow;

[0062] Figure 3 The process of constructing the bogie fault diagram structure, wherein the left side of the diagram is an undirected bogie fault diagram that is transformed into a directed bogie fault diagram on the right side of the diagram;

[0063] Figure 4 RSNet Model Structure

[0064] Figure 5 RS-GAT Fault Diagnosis Model Structure

[0065] Figure 6 Dataset distribution under different speed conditions, where (a)v train Distribution (b)v test Distribution;

[0066] Figure 7 Classification results under different working conditions;

[0067] Figure 8 Accuracy results of different models in 10 trials on small datasets; where a) accuracy on 50% dataset, (b) accuracy on 75% dataset, and (c) accuracy on 100% dataset. Detailed Implementation

[0068] This invention proposes a fault diagnosis method for high-speed train bogies based on residual neural networks; the invention will be further described below with reference to the accompanying drawings and embodiments. The specific steps are as follows:

[0069] Step 1, Bogie Fault Analysis and Classification

[0070] High-speed train bogie systems are divided into powered bogies and unpowered bogies. Powered bogies include the power unit, including the motor and gearbox. Due to the complex structure of the bogie system, numerous factors can cause failures in its components, resulting in a wide variety of fault types. Furthermore, the bearings in powered bogies bear relatively heavy loads, making them more prone to failure than unpowered bogies. To ensure train safety, it is necessary to conduct timely condition monitoring and fault identification of critical components of the train bogie system, classifying fault states and analyzing their potential causes. These fault states are categorized into seven types: healthy condition, frame fault, bearing fault, wheelset fault, gearbox fault, traction motor fault, and air spring fault. This includes 24 sensor measurement points with the most significant impact (such as...). Figure 1 (as shown);

[0071] The seven types of faults include 24 sensor measurement points with the most significant impact. The specific fault names are as follows:

[0072] Structural failures: cracks, wear, bending deformation;

[0073] Bearing failures: fatigue spalling, wear, plastic deformation, corrosion, and galling;

[0074] Wheelset failures: surface wear, surface peeling, axle cracks, axle overheating;

[0075] Gearbox malfunctions: broken teeth, pitting, wear, scratches.

[0076] Traction motor malfunctions: short circuit, excessive voltage, rotor breakage;

[0077] Air spring malfunctions: cracking, weakening, corrosion;

[0078] Step 2, Select bogie status data

[0079] During actual train operation, many factors can affect bogie failures. The severity of failures varies depending on the train's starting, braking, acceleration, and deceleration conditions. To improve the accuracy of bogie fault diagnosis, it is necessary to selectively extract the most relevant data from various bogie detection data for feature fusion. Therefore, Pearson coefficients are used for feature selection, with a threshold of 0.8 for the Pearson coefficient used when selecting relevant features. Through screening, the 24 sensor measurement points that have the most significant impact on Table 1 are obtained (the distribution of measurement points on the bogie is as shown in Table 1). Figure 1 (as shown);

[0080] Step 3: Establish the bogie data network architecture.

[0081] Because the train is structurally integrated, there is spatial correlation between the various measuring points on the train. Using 24-dimensional sensor status data obtained from 24 sensor measuring points as input, a bogie fault diagnosis model is established (e.g., ...). Figure 2 As shown in the figure, it outputs 7 types of fault states; this model is a typical multi-sensor fusion fault diagnosis model.

[0082] Step 4: Use grey theory to determine the framework of the graph structure, and use the Granger causal analysis method (GGC method) to determine the direction of the graph structure edges (e.g., ...). Figure 3 (As shown); the specific steps are as follows:

[0083] Step 1: Construct the sensor information matrix X = (x ij ) n×t Calculate the value x for each comparison sensor. j The correlation coefficient ζ between the corresponding elements and the reference sensor ij (k); the specific formula is as follows:

[0084]

[0085]

[0086] Where n is the number of sensors, t is time, and the sensor values ​​in the 1st and 1st columns at time j are x, respectively. 1j and x ij ρ is the correlation constant.

[0087] Step 2: Calculate the correlation degree r′ ij As shown in the following formula:

[0088] Where σ j Let be the correlation coefficient between all corresponding elements of the sensor values ​​at time j.

[0089] Step 3: Determine the weight matrix W = (w ij ) n×n If the weights of the graph structure are defined by the correlation coefficients between sensors, then the formula for W is as follows:

[0090] W = (w ij ) n×n =(r ij ) n×n

[0091] The fourth step is to define the constraints of the high-speed train diagram structure. Three constraints are selected: data relationship R1, system to which it belongs R2, and carriage to which it belongs R3. These constraints work together on the sensor nodes, and a weighted method is used to determine the final association matrix R of the high-speed train diagram structure. The specific formula is shown below:

[0092]

[0093]

[0094]

[0095] i = 1, 2, ..., n

[0096] j = 1, 2, ..., n

[0097] R = (r ij ) n×n = w1·R1 + w2·R2 + w3·R3

[0098] w1 + w2 + w3 = 1

[0099] Where w1 represents the weight of data relationship R1, w2 represents the weight of system R2, and w3 represents the weight of carriage R3.

[0100] Step 5: Determine the direction of the edge. In the time series of the high-speed train sensor measurement points, if the vehicle information measurement point v... i The value X i Within 1 second, the measuring point v can be measured. j The value X j To have an effect, that is, to observe X i The information change can explain the X that appears after 1 second. j The change in information indicates that X i With X j There is a causal relationship between the two in the data, namely X i →X j ; or by extension, v i and v j There is a causal relationship between them, i.e., v i →v j e ij =1 Given that the time series of sensor measurement points on high-speed trains is a stable series, this invention employs Granger causality testing based on a VAR model to establish (X i X j The VAR model is shown below:

[0101]

[0102]

[0103] Among them, X i (t) and X j (t) represents the value of the sensor measuring point on the high-speed train at time t; X i (tk) and X j (tk) represents the value of the sensor measurement point on the high-speed train at time tk, and the value space of k is [0, l]; a k and c k It is the Granger causality coefficient; b k and d k It is the autoregressive coefficient; ε t and μ t This represents the prediction error, which is independent of the time point and is assumed to be white noise.

[0104] If X i (t) is not X j If (t) is a Granger cause, then a1 = a2 = ... = a l =0. The Granger causality test is performed using a constrained F-statistic, which specifies that X = 0. i (t) The sum of squared residuals is RSS U X j The sum of squared residuals of (t) is RSS RThe F-statistic test formula is as follows:

[0105]

[0106] If the F-test conclusion rejects the null hypothesis H0: a1=a2=...=a l =0, then sensor i is the Granger cause of sensor j, that is, from vertex v i To vertex v j It is a directed edge; otherwise, sensor i is not a Granger cause of sensor j. Similarly, if the F-test conclusion rejects the null hypothesis H0: c1=c2=...=c l =0, then sensor i is the Granger cause of sensor j, that is, from vertex v j To vertex v i It is a directed edge. Furthermore, the stronger the Granger causality, the smaller the value of F. Therefore, the edge E = (e...) in the high-speed train graph structure... ij ) n×n The following conditions must be met:

[0107]

[0108] This invention uses the above method to construct a fault diagram of a high-speed train bogie. The specific construction process (as shown in Figure 2) and the resulting diagram structure (as shown in Figure 2) are described below. Figure 3 (As shown).

[0109] Step 5, RS-GAT fault diagnosis model

[0110] Based on the RSNet model framework, the directed graph for bogie fault diagnosis is input, and the original convolutional layers are replaced with GAT, which has stronger aggregation capabilities, to construct the RS-GAT fault diagnosis model. Since GAT already possesses aggregation capabilities in graph neural networks, the RS-GAT model does not use the Squeeze operation in the RSGAT units, but uses the GAP operation in the final output to ensure that the average value of all input feature maps is used as the output. (The structure of the RS-GAT fault diagnosis model is as follows...) Figure 5 As shown.

[0111] The RSNet model is a CNN model inspired by the DCNN model and the Squeeze operation of SENet, designed based on the residual learning framework of ResNet, with a one-dimensional time-series signal as its input. (The structure of the RSNet model is as follows...) Figure 4 As shown.

[0112] The RSNet model uses the ResNet network as its framework, replacing the original residual learning units with convolutional layers. Each convolutional layer includes a BN layer, a ReLU layer, and a convolution operation. After passing through three convolutional layers with 16 kernels and three convolutional layers with 32 kernels, the RSNet model performs global average pooling (GAP) on all features. Finally, it passes through a fully connected layer (FC) and a softmax operation to obtain the final result.

[0113] The RS-GAT model can be divided into three parts:

[0114] The first part consists of the input cells for the RS-GAT model. The model input consists of the directed graph data X for bogie fault diagnosis and the graph structure A. After passing through a GAT, the normalized attention coefficient a of node j with respect to node i is obtained. ij The specific calculation formula is as follows:

[0115] e ij =Attention(Wx i Wx j =LeakyReLU(w T [Wx i ||Wx j ])

[0116]

[0117] Among them, LeakyReLU(w T [Wx i PWx j ]) is the activation function, exp(LeakyReLU(w T [Wx i PWx j ])) represents an exponential operation, and the data at time i and time j are x i and x j The first-order neighborhoods of nodes i and j are and The learning parameter is W, and the training parameter is w. The attention coefficient a is used. ij Compute the reconstructed vector x′ of node i i The calculation formula is as follows:

[0118]

[0119] Where f(·) is a non-linear activation function. Multiple attention units are concatenated to improve the feature capture capability of node i. Specifically, the following equation applies:

[0120]

[0121] Where || represents the concatenation operation, and K represents the number of attention units.

[0122] The second part consists of l RSGAT units. The RSGAT unit is designed with six layers based on the residual network framework: BN layer, ReLU layer, GAT layer, BN layer, ReLU layer, and GAT layer. Taking the first layer as an example (l=1), the residual can be calculated as F1, and the output is H1=F1+X0. After every three RSGAT units, the output H matrix is ​​pooled, and then the graph attention units of the next three RSGAT units are doubled. After processing through l layers (l must be a multiple of 3) of RSGAT units, the final output matrix H is obtained. l .

[0123] The third part is the output unit of the RS-GAT model. The H output in the second part... l It contains a large amount of fault characteristic information, H l Pooling operations are performed, followed by GAP, FC, and Softmax operations to finally obtain the fault classification results.

[0124] In summary, this invention proposes a novel fault diagnosis framework, RS-GAT. The RS-GAT model replaces the convolutional operations in the original RS-Net model with the node aggregation operations of the GAT model, thereby reducing the redundancy of information in bogie data and improving the accuracy of fault feature extraction. This invention uses a 6-layer RSGAT framework to extract the spatial features of the bogie and uses Global Average Pooling (GAP) to obtain the bogie fault category.

[0125] Example

[0126] All data used in this invention are sensor data collected during the actual operation of high-speed trains, with a sampling interval of 30 seconds. An example of the data obtained after resampling the data from 24 sensors is shown in Table 1.

[0127] Table 1 Data Example

[0128]

[0129] To facilitate comparison and robustness analysis, this invention divides the 12 hours from 8:30 AM to 8:30 PM when the train is in operation into six parts. Since time features need to be extracted later, each part contains time-series data. After processing, six datasets are obtained, coded as C1, C2, C3, C4, C5, and C6. Each dataset includes different train operating conditions: acceleration, deceleration, and constant speed. Furthermore, each dataset contains seven fault features. Each dataset includes a training set, a validation set, and a test set; the dataset sizes are shown in Table 2.

[0130] Table 2. Dataset descriptions under different conditions

[0131]

[0132] This invention uses C1 as an example to divide speed ranges and statistically analyze sensor data under various working conditions. Figure 6 (a) shows the distribution of the training set data under different speed conditions; (b) shows the distribution of the test set data under different speed conditions. Figure 6 As shown in (a) and (b), the data for various fault detections appear in multiple speed ranges, indicating that the detection data closely matches the actual working scenario, which undoubtedly greatly improves the robustness of the model.

[0133] (1) Evaluation indicators

[0134] The main evaluation metrics for the RS-GAT model are cost time and accuracy. Cost time primarily refers to the time spent on fault diagnosis after model training; accuracy is used to calculate the probability of correct classification. The formula for calculating accuracy is shown below:

[0135]

[0136] Where TP represents the amount of data that was originally classified as fault but is now classified as fault, TN represents the amount of data that was originally classified as non-fault but is now classified as non-fault, and Total represents the amount of data in all categories.

[0137] (2) Experimental Environment

[0138] The experimental environment for high-speed train bogie fault diagnosis using the RS-GAT model and its comparative methods is shown in Table 3. Both the RS-GAT model and its comparative methods are implemented using Python, and the deep learning architecture used in this invention is the PyTorch library. Training of the RS-GAT model is primarily performed on a GPU.

[0139] (3) Analysis of experimental results

[0140] This invention uses data from six different scenarios for experiments. Taking dataset C1 as an example, the accuracy of the RS-GAT model for different labels and the overall accuracy are shown in Table 4.

[0141] Table 3. Experimental Environment Description

[0142]

[0143]

[0144] (3) Analysis of experimental results

[0145] This invention uses data from six different scenarios for experiments. Taking dataset C1 as an example, the accuracy of the RS-GAT model for different labels and the overall accuracy are shown in Table 4.

[0146] As shown in Table 4, the overall accuracy rate for the C1 dataset is 95.83%, indicating that the RS-GAT model can effectively identify the vast majority of faults in train bogies. Except for Label 2 and Label 6, the classification accuracy of the other labels is above 95%, demonstrating the effectiveness of the RS-GAT model in fusing data from multiple sensors. Among the seven labels, Label 0 has the highest classification accuracy, meaning it is classified as a normal state. This is because the high-speed train operation data in the C1 dataset is mostly normal data, thus the weight for learning the normal state as the correct normal state is the highest in feature learning. The label with the lowest classification accuracy is Label 6, which represents an air spring fault. Since air spring fault features are mainly extracted from data collected by pressure sensors CR_R_ZD_KHYL and CR_L_ZD_KHYL, the amount of data collected for air spring features in the C1 dataset is relatively small compared to other faults, resulting in a relatively lower accuracy.

[0147] Table 4. C1 Fault Classification Results

[0148]

[0149]

[0150] To further analyze the classification performance of the RS-GAT model on different datasets, the classification results of the RS-GAT model on six datasets are presented as follows: Figure 7As shown in the figure, the classification accuracy of each label is higher than 92% under all operating conditions, with a calculated average accuracy of 91.31%, indicating that the RS-GAT fault diagnosis model has strong robustness and can achieve high classification accuracy under various complex operating conditions and measured data. Analysis of the accuracy under different operating conditions shows that scenario C4 has the best classification performance, with an average accuracy of 96.34%. Analysis of the C4 dataset reveals that because it contains fewer stopping and starting scenarios, and the train is mostly traveling at a constant speed, the measured data is more accurate and can better extract the fault characteristics of each fault.

[0151] (4) Multi-method comparative analysis

[0152] To further investigate the performance of the RS-GAT model in train bogie fault diagnosis, this invention conducts comparative experiments using the RS-GCN model against the RSNet-14 model, GAT model, GCN model, and GRU model. All models used the same training, validation, and test sets in the experiments. Specific classification results are shown in Table 5. Only the average classification accuracy is displayed for each dataset.

[0153] Table 5. Classification results of different models on 6 datasets.

[0154]

[0155] Table 5 shows that the RS-GAT model achieves the highest accuracy across all datasets. The following section will compare the RS-GAT model with other models:

[0156] 1) Compared to traditional methods for diagnosing faults in high-speed train bogies, ResNet-based models demonstrate better diagnostic performance. This is because ResNet models can integrate many shallow networks, making the overall model more networked and three-dimensional. Simultaneously, the ResNet framework ensures that feature information flows more easily between layers, better capturing the characteristic information of the data.

[0157] 2) Comparing the RS-GAT model and RS-GCN model that use graph structures with the RSNet-14 model that does not use graph structures, it can be seen that the model using graph structures performs better than the model that does not use graph structures. This is because the bogies of high-speed trains are spatially correlated. For a node, using the information of its neighboring nodes to supplement the information of that node is more helpful in extracting spatial features, thereby achieving better classification results.

[0158] 3) Comparing GAT-type models and GCN-type models reveals that the accuracy of GAT-type models is higher than that of GCN-type models. This is because the input of GAT-type models is a directed graph of the train bogie diagnostic network, which adds directional information compared to the input of GCN. GAT-type models use the importance of adjacent nodes to the center node instead of the Laplace transform in GCN-type models, which not only improves the interpretability of the model but also enhances its spatial expressiveness.

[0159] (5) Performance comparison on small datasets

[0160] Although all the data used in this invention comes from data collected from actual train operations, training the model with massive datasets is an important way to improve the accuracy of fault identification for deep learning models. Furthermore, since unexpected situations may occur during actual train operation, it is necessary to study the model's classification performance on small datasets to verify its robustness.

[0161] This invention extracts 523,200 (24 × 21,800) data points with frequent fault occurrences from six datasets to create a new dataset. This dataset includes 355,200 (24 × 14,400) training data points, 84,000 (24 × 3,500) validation data points, and 84,000 (24 × 3,500) test data points. Then, a training set is randomly sampled from Ct, with the training set sizes being 75% and 50% of the original training set, respectively. Simultaneously, comparative analyses are performed using the RS-GAT model, RS-GCN model, RSNet-14 model, GAT model, and GCN model. To avoid the random influence of incomplete selection, each model is tested 10 times, and the results are as follows. Figure 8 As shown, the accuracy rates are (a) for the 50% dataset, (b) for the 75% dataset, and (c) for the 100% dataset. The RS-GAT model performs best in fault diagnosis on the 50% and 75% datasets. Furthermore, the ResNet framework models (RS-GAT, RS-GCN, and RCNet-14) significantly outperform the single-layer graph structure models (GAT and GCN), indicating that the ResNet framework models are more robust, with the RS-GAT model exhibiting the best robustness. Table 6 shows the accuracy and classification time for the 50%, 75%, and 100% datasets.

[0162] Table 6 compares the accuracy and processing time of different models on a small dataset.

[0163]

[0164] As shown in Table 6, the RS-GAT model also achieves high accuracy on small datasets (e.g., Figure 7 , Figure 8 (As shown), however, the classification process takes longer than that of a single-layer classification model. Considering both data accuracy and classification time, the RS-GAT model can achieve higher accuracy bogie fault diagnosis in a relatively short time. This further verifies that the generalization ability of the RS-GAT model proposed in this invention is higher than that of the RS-GCN model, RSNet-14 model, GAT model, and GCN model.

Claims

1. A method for fault diagnosis of high-speed train bogies based on residual neural networks, characterized in that, The specific steps are as follows: Step 1, Bogie Fault Analysis and Classification High-speed train bogie systems are divided into powered bogies and non-powered bogies. Powered bogies include power units such as motors and gearboxes. Due to the complex structure of the bogie system, there are many causes and types of failures that can lead to various components of the bogie system. At the same time, the bearing load in powered bogies is relatively large, making them more prone to failure than non-powered bogies. In order to ensure the safety of train operation, it is necessary to monitor the condition and identify faults of important components of the train bogie system in a timely manner and classify the fault conditions. Analyze the possible causes of the malfunction; The fault states are divided into 7 categories, including: healthy state, frame fault, bearing fault, wheelset fault, gearbox fault, traction motor fault, and air spring fault; and include 24 sensor measurement points with the most obvious impact. Step 2, Select bogie status data During actual train operation, many factors can affect bogie failures. The severity of failures varies depending on the train's starting, braking, acceleration, and deceleration conditions. To improve the accuracy of bogie fault diagnosis, it is necessary to selectively filter the data most relevant to bogie failures from various bogie detection data and perform feature fusion. Therefore, Pearson coefficients are used for feature selection, with a threshold of 0.8 for the Pearson coefficient used when filtering relevant features. Through screening, 24 sensor measurement points with the most significant impact are obtained. Step 3: Establish the bogie data network architecture. Since the same train is structurally integrated, there is spatial correlation between the various measuring points on the train. Using 24-dimensional sensor status data obtained from 24 sensor measuring points as input, a bogie fault diagnosis model is established, and 7 types of fault states are output. This model is a typical multi-sensor fusion fault diagnosis model. Step 4: Use grey theory to determine the framework of the graph structure, and use the Granger causal analysis method (GGC method) to determine the direction of the graph structure edges; specifically including: Step 1: Construct the sensor information matrix X = (x ij ) n×t Calculate the value x for each comparison sensor. j The correlation coefficient ζ between the corresponding elements and the reference sensor ij (k); the specific formula is as follows: Where n is the number of sensors, t is time, and the sensor values ​​in the 1st and 1st columns at time j are x, respectively. 1j and x ij ρ is the correlation constant; Step 2: Calculate the correlation degree r′ ij As shown in the following formula: Where σ j Let be the correlation coefficient between all corresponding elements of the sensor values ​​at time j; Step 3: Determine the weight matrix W = (w ij ) n×n If the weights of the graph structure are defined by the correlation coefficients between sensors, then the formula for W is as follows: W=(w ij ) n×n =(r ij ) n×n The fourth step is to define the constraints of the high-speed train diagram structure. Three constraints are selected: data relationship R1, system to which it belongs R2, and carriage to which it belongs R3. These constraints work together on the sensor nodes, and a weighted method is used to determine the final association matrix R of the high-speed train diagram structure. The specific formula is shown below: i = 1, 2, ..., n j＝1，2，...， R=(r ij ) n×n =w1·R1+w2·R2+w3·R3 w1 + w2 + w3 = 1 Where w1 represents the weight of data relationship R1, w2 represents the weight of system R2, and w3 represents the weight of carriage R3. Step 5: Determine the direction of the edge. In the time series of the high-speed train sensor measurement points, if the vehicle information measurement point v... i The value X i Within l seconds, measure point v j The value X j To have an effect, that is, to observe X i Explanation of the information change: X appears after 1 second. j The change in information indicates that X i With X j There is a causal relationship between the two in the data, namely X i →X j ; or by extension, v i and v j There is a causal relationship between them, i.e., v i →v j e ij =1 Given that the time series of high-speed train sensor measurement points is a stable series, a Granger causality test based on a VAR model is adopted to establish (X i ,X j The VAR model is shown in the following formula: Among them, X i (t) and X j (t) represents the value of the sensor measuring point on the high-speed train at time t; X i (tk) and X j (tk) represents the value of the sensor measurement point on the high-speed train at time tk, and the value space of k is [0, l]; a k and c k It is the Granger causality coefficient; b k and d k It is the autoregressive coefficient; ε i and μ i This is the prediction error, independent of the time point, and is assumed to be white noise. If X i (t) is not X j If (t) is a Granger cause, then a1 = a2 = ... = a i =0; The Granger causality test is performed using a constrained F-statistic, which specifies that X = 0; i (t) The sum of squared residuals is RSS U X j The sum of squared residuals of (t) is RSS R The F-statistic test formula is as follows: If the F-test conclusion rejects the null hypothesis H0: a1=a2=...=a i =0, then sensor i is the Granger cause of sensor j, that is, from vertex v i To vertex v j It is a directed edge; otherwise, sensor i is not a Granger cause of sensor j; similarly, if the F-test conclusion rejects the null hypothesis H0: c1=c2=...=c i =0, then sensor i is the Granger cause of sensor j, that is, from vertex v j To vertex v i It is a directed edge; furthermore, the stronger the Granger causality, the smaller the value of F. Therefore, the edge E = (e...) in the high-speed train graph structure... ij ) n×n The following conditions must be met: Step 5, RS-GAT fault diagnosis model Based on the RSNet model framework, the directed graph for bogie fault diagnosis is input, and the original convolutional layers are replaced with GAT, which has stronger aggregation capabilities, to construct the RS-GAT fault diagnosis model. Since GAT already has aggregation capabilities in graph neural networks, the RS-GAT model does not use the Squeeze operation in the RSGAT unit, but uses the GAP operation in the final output to ensure that the average value of all input feature maps is taken as the output. The RSNet model is a CNN model designed based on the residual learning framework of ResNet, inspired by the Squeeze operation of DCNN and SENet, and its input is a one-dimensional time-series signal.

2. The high-speed train bogie fault diagnosis method based on residual neural network according to claim 1, characterized in that, The seven types of faults include 24 sensor measurement points with the most significant impact. The specific fault names are as follows: Structural failures: cracks, wear, bending deformation; Bearing failures: fatigue spalling, wear, plastic deformation, corrosion, and galling; Wheelset failures: surface wear, surface peeling, axle cracks, axle overheating; Gearbox malfunctions: broken teeth, pitting, wear, scratches. Traction motor malfunctions: short circuit, excessive voltage, rotor breakage; Air spring malfunctions: cracking, weakening, corrosion.

3. The high-speed train bogie fault diagnosis method based on residual neural network according to claim 1, characterized in that, The RSNet model uses the ResNet network as its framework, replacing the original residual learning units with convolutional layers. Each convolutional layer includes a BN layer, a ReLU layer, and a convolution operation. After passing through three convolutional layers with 16 kernels and three convolutional layers with 32 kernels, the RSNet model performs global average pooling (GAP) on all features. Finally, it passes through a fully connected layer (FC) and a softmax operation to obtain the final result.

4. The high-speed train bogie fault diagnosis method based on residual neural network according to claim 1, characterized in that, The RS-GAT model consists of three parts: The first part is the input unit of the RS-GAT model. The model input is the directed graph data X for bogie fault diagnosis and the graph structure A. After passing through a GAT, the normalized attention coefficient a of node j with respect to node i is obtained. ij The specific calculation formula is as follows: e ij =Attention(Wx i ,Wx j )=LeakyReLU(w T [Wx i PWx j ]) Among them, LeakyReLU(wT[Wx i PWx j ]) is the activation function; exp(LeakyReLU(w T [Wx i PWx j ])) represents an exponential operation, and the data at time i and time j are x i and x j The first-order neighborhood of nodes i and j is N. i and N j The learning parameter is W, and the training parameter is w; using the attention coefficient a ij Compute the reconstructed vector x′ of node i i The calculation formula is as follows: Where f(·) is a non-linear activation function; multiple attention units are concatenated to improve the feature capture capability of node i, as shown in the following formula: Where P represents the splicing operation and K represents the number of attention units; The second part consists of l RSGAT units. The RSGAT unit is designed with six layers based on the residual network framework: BN layer, ReLU layer, GAT layer, BN layer, ReLU layer, and GAT layer. After every three RSGAT units, the output H matrix is ​​pooled, and then the graph attention units of the next three RSGAT units are doubled. After processing by l RSGAT units, the final output matrix H is obtained. l l is a multiple of 3; The third part is the output unit of the RS-GAT model, which is the H output from the second part. l It contains a large amount of fault characteristic information, H l Pooling operations are performed, followed by GAP, FC, and Softmax operations to finally obtain the fault classification results.

Images