High-speed train brake pad state monitoring system and method based on improved graph neural network

Abstract

The application provides a high-speed train brake pad state monitoring system and method based on an improved graph neural network, a node feature matrix and a relationship matrix representing the similarity of each node are constructed according to collected friction block original signals; a feature graph and a topology graph are constructed according to the node feature matrix and the relationship matrix; the constructed feature graph and topology graph are optimized by using Bayesian optimization; first features and second features are extracted according to the feature graph and the topology graph respectively, and the first features and the second features are combined with adaptive weights to obtain fused features through weighted fusion; and the fused features are used to determine the fault category. The application extracts abstract features from a topology and a feature dual space, optimizes the number of neighbor nodes in the topology space and the feature space by using a Bayesian algorithm, improves the quality of the constructed graph, uses the powerful graph structure processing advantage of the GCN, fuses information from multiple graph spaces, and realizes the imbalance data state monitoring of the brake pad.

Description

Technical Field

[0001] This invention belongs to the field of rail transit technology and relates to the monitoring of high-speed train operation status, especially the technology for monitoring the status of high-speed train brake pads during high-speed train operation under unbalanced data. Background Technology

[0002] With the rapid development of high-speed trains, higher demands have been placed on train braking. Mechanical braking, as one of the main braking methods for trains, relies on the friction between the brake disc and brake pads. Studies show that trains frequently operate under extremely complex and variable braking conditions, such as high-speed emergency braking and repeated braking under heavy loads. Due to the effects of high speed, heavy loads, and the resulting high temperatures, brake friction pads are prone to abnormal wear, leading to deterioration phenomena such as uneven wear, peeling, and cracks. When deteriorated friction pads rub against the brake disc interface, it causes more severe noise and vibration, affecting ride comfort, reducing the service life of components, and even threatening the safe operation of high-speed trains. Therefore, timely condition monitoring of the brake friction components is a crucial technical guarantee for the safe and reliable operation of high-speed trains.

[0003] With the rapid rise of artificial intelligence technology, using deep learning to monitor the braking status of high-speed trains has become a research hotspot. Zhang et al. proposed an ensemble model based on convolutional neural networks (CNN) and bidirectional gated recurrent units to achieve intelligent monitoring of uneven wear of brakes by fusing multi-source information [Zhang M, Zhang X, Mo J, et al. Brake uneven wear of high-speed train intelligent monitoring using an ensemble model based on multi-sensor feature fusion and deep learning; Liu et al. proposed a model based on multidimensional time-series data feature selection and feature fusion to achieve fault diagnosis of heavy-haul train braking system [Liu Z, Zhang M, Liu F, et al. Multidimensional Feature Fusion and Ensemble Learning-Based Fault Diagnosis for the Braking System of Heavy-Haul Train[J]. IEEE Transactions on Industrial Informatics, 2021, 17(1): 41-51]; Hu et al. proposed a deep subdomain generalization network (DSGN), which learns and generalizes the time-frequency domain acceleration signals of uneven wear at different positions of the friction block through DSGN, and realizes intelligent state recognition of target brake pads in an unsupervised manner [Hu R, Zhang M, Meng X, et al.]. The previous research focused on monitoring the braking status of high-speed train brake pads under balanced state category data. However, during actual braking of high-speed trains, the friction pads operate in a normal state for extended periods, resulting in a predominance of normal samples collected by sensors and a scarcity of fault state data, exhibiting a data imbalance problem. If the imbalanced data is input into a deep network for learning, the resulting features become more biased towards normal state samples, leading to underfitting of fault state data and identification errors. Therefore, monitoring the brake pad status under imbalanced datasets has become a significant challenge in achieving high-speed train braking friction status monitoring.

[0004] Oversampling, undersampling, and downsampling strategies for synthesizing minority class samples have been proposed, providing a way to address the imbalance of sample data. However, these methods tend to focus on data preprocessing techniques, facing problems such as weak stability, difficulty in parameter selection, and the learned features being biased towards normal sample features. In recent years, researchers have used nearest neighbor algorithms and deep neural networks to extract features from a small number of samples, effectively improving the quality of samples while increasing sample diversity. However, these methods still focus on the data itself, lacking exploration of the relationships and structures hidden between data. In real-time brake pad status detection, analyzing the correlation between data in different states of the brake pad through the collected signals, and constructing and learning the correlation between signal data, helps in the status monitoring of the brake pad.

[0005] To address the aforementioned issues, the Graph Convolutional Network (GCN) method extends convolution operations from traditional data (such as text, images, and structured data) to graph data (unstructured data). Based on graph theory, it constructs a convolutional network for the data stored in the graph. In data processing, it considers the correlation of samples, increasing the identifiable features of a few imbalanced samples. Furthermore, it performs neural network operations on this basis, ensuring that the learned features of each sample are equal. Different fault states of friction blocks are aggregated using a graph to combine their own features with related features. The geometric structure of the data can provide better information, making it more practical for monitoring the imbalanced state of brake pads. Liu et al. proposed a SuperGraph feature learning method based on an autoencoder, which performs graph convolution on the constructed SuperGraph and achieves good results in fault diagnosis of unbalanced training datasets for rotating machinery [Liu J, Zhou K, Yang C, et al. Imbalanced fault diagnosis of rotating machinery using autoencoder-based SuperGraph feature learning[J]. Frontiers Of Mechanical Engineering, 2021, 16(4): 829-839]; Jongmin et al. proposed an Edge Labeling Graph Neural Network (EGNN), which iteratively updates edge labels by directly utilizing intra-cluster similarity and inter-cluster inconsistency, and obtains a good generalization model for handling imbalanced data problems [Jongmin K, Taesup K, Sungwoong K, et al. Edge-Labeling Graph Neural Network for Few-Shot Learning[M]. 2019]; Deng et al. proposed a Directed Graph Convolutional Neural Network (DGCNN), which enhances the salient region features of a small number of target samples by aggregating the signals of neighbors and outside neighbors [Kenning M, Deng JJ, Edwards M, et al.]. [al. A directed graph convolutional neural network for edge-structured signals in link-fault detection[J]. PatternRecognition Letters,2022,153:100-106]. Although the above literature finds the connection between different fault types by constructing edges in the graph, it ignores the influence of the differences in node features and the topological relationships between adjacent nodes when constructing the graph. In the data of brake pad imbalance, relying solely on the feature space to construct the relationship graph will confuse the minority class sample data and fail to correctly obtain the important information to distinguish the fault categories.

[0006] In conclusion, the problem of brake pad imbalance data has not yet been resolved, which will seriously affect the effective monitoring of brake pad status in high-speed trains, and thus pose a hidden danger to the safe operation of high-speed trains. Summary of the Invention

[0007] The purpose of this invention is to provide a high-speed train brake pad status monitoring system and method based on an improved graph neural network, which can effectively solve the problem of unbalanced data of high-speed train brake pads, thereby realizing effective monitoring of the status of high-speed train brake pads.

[0008] To address the aforementioned problems, this invention proposes a high-speed train brake pad condition monitoring system based on improved graph neural networks (FT-GCN). This system is an adaptive multi-channel graph convolutional network that fuses topology and feature spaces to construct the input network graph. A Bayesian optimization algorithm is used to optimize the number of neighbor nodes in both graph structures, and the graph constructed with the optimal values ​​is input into the graph convolutional network. The network then extracts the feature embeddings of the braking signal and uses the Bayesian optimization algorithm to determine the weights of each graph space. This invention integrates these weights to achieve the fusion of multiple graph spaces, thus enabling fault diagnosis of brake pads.

[0009] The present invention provides a high-speed train brake pad status monitoring system based on an improved graph neural network, comprising:

[0010] The input matrix construction module is used to construct a node feature matrix and a relationship matrix representing the similarity of each node based on the collected original signals of the friction block.

[0011] The input graph construction module is used to construct feature graphs and topology graphs based on the node feature matrix and relation matrix; and to optimize the constructed feature graphs and topology graphs using Bayesian optimization.

[0012] The feature extraction module includes two parallel, identical graph convolutional units and a feature fusion unit. The two graph convolutional units are used to extract the first feature and the second feature based on the feature map and the topology map, respectively. The feature fusion unit is used to combine the first feature and the second feature with adaptive weights to obtain the fused feature.

[0013] The fault identification module is used to determine the fault category based on the obtained fused features.

[0014] The above-mentioned input matrix construction module states that the original signal includes N nodes, each node containing n feature vectors; the N node features constitute the node feature matrix X; the relationship between two nodes is calculated using cosine similarity, that is, the cosine value of the angle between two vectors in the vector space is used as a measure of the magnitude of the difference between two individuals.

[0015]

[0016] In the formula, X i and X j Let A represent two node vectors respectively. ij This represents the similarity between two node vectors.

[0017] Initially, the cosine similarity between all nodes forms a relation matrix A, which contains K1 nodes; K1 = N.

[0018] The above input graph construction module includes the following sub-modules:

[0019] The topology graph construction submodule generates a topology graph from the relation matrix and the node feature matrix;

[0020] The first Bayesian optimization module is used to optimize the generated topology graph;

[0021] The feature map construction submodule first uses the KNN algorithm to obtain the adjacency matrix based on the similarity of each node in the relation matrix, and then generates the feature map from the adjacency matrix and the node feature matrix.

[0022] The second Bayesian optimization module is used to optimize the generated feature map.

[0023] The topology graph construction submodule generates a topology graph G = (A, X) from the relation matrix A and the node feature matrix X.

[0024] The feature map construction submodule, based on the relation matrix A, sorts the nodes in descending order of cosine similarity, and selects the top K2 similar nodes for each node to set edges, thus obtaining the adjacency matrix A. f :

[0025]

[0026] In the formula, m and n represent two nodes, and E represents the set of edges associated with the nodes.

[0027] Then from relation matrix A f The feature map GF = (A) is generated from the feature matrix X of the nodes. f (X).

[0028] The first and second Bayesian modules utilize the Bayesian optimization algorithm to optimize the topology graph and feature graph (i.e., optimize K1 and K2) respectively, obtaining topology graphs and feature graphs with optimal K1 and K2. Bayesian optimization is also known as the sequence model-based optimization method (SMBO). The Bayesian optimization method uses a Gaussian process regression model to obtain the objective function f(K) of the feature graph and topology graph respectively, and obtains the globally optimal solution K that satisfies the objective function according to the following optimization function. *(See Shahriari B, Swersky K, Wang Z, et al. Taking the Human Out of the Loop: A Review of Bayesian Optimization[J]. Proceedings Of the Ieee, 2016,104(1):148-75):

[0029]

[0030] K* represents the global optimum of the constraint domain of f(K).

[0031] The aforementioned feature extraction module includes two parallel, identically structured graph convolutional units (GCNs) and a feature fusion unit.

[0032] The graph convolutional unit consists of two identical graph convolutional layers arranged sequentially. Each graph convolutional layer includes several filtering layers, followed by activation functions and pooling layers. The graph convolutional layers are implemented according to the Laplacian and Chebyshev polynomials of the graph (see Kan Guo; Yongli Hu; Zhen Qian, et al. Civil and Environmental Engineering and H. John Heinz III College, Carnegie Mellon University, Pittsburgh, PA, USA Dynamic Graph Convolution Network for Traffic Forecasting Based on Latent Network of Laplace Matrix Estimation), where the filtering and activation layers are designed to generate better node features. The output of the second convolutional layer is normalized to fix its range. The pooling layer summarizes the node features, generating high-level features that capture the information of the entire graph. After a series of graph filtering and activation layers, the pooling layer produces a coarsened graph with more abstract and higher-level node features. Through multiple iterations of the above image convolutional unit, multiple blocks are used to extract fault features from the input.

[0033] After normalization, the learned features are processed. Considering the varying correlations between node labels and the topology and feature maps extracted by the graph convolutional units, a feature fusion unit is placed after two parallel graph convolutional units. The feature fusion unit uses a Bayesian network to extract weights for these classification features (see Shahriari, Bobak; Swersky, Kevin; Wang, Ziyu et al. Taking the Human Out of the Loop: A Review of Bayesian Optimization). The topology and feature maps are then weighted and summed to obtain the fused features.

[0034] The aforementioned fault identification module can use the argmax function to determine the fault category based on fused features.

[0035] This invention further provides a training method for a high-speed train brake pad status monitoring system based on an improved graph neural network, which includes the following steps:

[0036] A1 divides the training data samples into a training set and a validation set;

[0037] A2 uses the input matrix construction module to process the training data samples and obtain the corresponding node feature matrix and the relationship matrix representing the similarity of each node.

[0038] A3 first uses the node feature matrix and relation matrix of the training data samples to obtain the topology graph and feature graph using the input graph construction module; then, using the obtained topology graph and feature graph, it obtains the node parameters K1 of the topology graph and K2 of the feature graph based on Bayesian optimization.

[0039] A4 uses the training set data processed by the input matrix construction module and the trained input graph construction module to train the graph convolutional unit and the fault identification module, respectively, which are suitable for topology graphs and feature graphs, and fixes the network parameters of the graph convolutional unit.

[0040] A5 uses the validation set data processed by the input matrix construction module and the trained input graph construction module to obtain the weights of the output features of the two graph convolution units in the feature extraction module composed of two graph convolution units, and completes the training of the high-speed train brake pad status monitoring system through Bayesian optimization to obtain the weights of the feature fusion unit in the output features of the two graph convolution units.

[0041] In step A3 above, the first Bayesian optimization module and the second Bayesian optimization module are trained using the optuna library with the obtained topology map and feature map to obtain the optimal parameters K1 of the first Bayesian optimization module and the optimal parameters K2 of the second Bayesian optimization module.

[0042] This invention further provides a method for monitoring the status of brake pads on high-speed trains based on an improved graph neural network, which includes the following steps:

[0043] S1 constructs a node feature matrix and a relationship matrix representing the similarity of each node based on the collected original signals of the friction block;

[0044] S2 constructs feature maps and topology maps based on node feature matrices and relation matrices; and optimizes the constructed feature maps and topology maps using Bayesian optimization.

[0045] S3 extracts the first feature and the second feature based on the feature map and the topology map respectively, and then combines the first feature and the second feature with adaptive weights to perform weighted fusion to obtain the fused feature;

[0046] S4 determines the fault category based on the obtained fusion features.

[0047] In step S1 above, the similarity of each node is calculated from the cosine similarity between the nodes. Step S2 above includes the following sub-steps:

[0048] S21 generates a topology graph from the relation matrix and the node feature matrix, and then optimizes the generated topology graph using a Bayesian optimization algorithm.

[0049] S22 first uses the KNN algorithm to obtain the adjacency matrix based on the similarity of each node in the relation matrix. Then, it generates a feature map from the adjacency matrix and the node feature matrix. Finally, it optimizes the generated feature map using the Bayesian optimization algorithm.

[0050] Step S3 above includes the following sub-steps:

[0051] S31 performs graph convolution processing on the optimized topology graph and feature graph respectively to extract the first feature and the second feature;

[0052] S32 obtains the fused feature by weighting and summing the first and second features based on their weights.

[0053] In step S3 above, the argmax function can be used to determine the fault category based on the fusion features.

[0054] The high-speed train brake pad status monitoring system based on improved graph neural networks provided by this invention has the following beneficial effects:

[0055] (1) Extract abstract features from the topology and feature spaces, and use the Bayesian algorithm to optimize the number of neighbor nodes in the topology and feature spaces to improve the quality of the constructed graph.

[0056] (2) By leveraging the powerful graph structure processing capabilities of GCN, information from multiple graph spaces is integrated to achieve monitoring of brake pad imbalance data status.

[0057] (3) Determine the optimal weights for each spatial branch through a Bayesian network to improve the accuracy of the system's final decision. Attached Figure Description

[0058] Figure 1 A schematic diagram illustrating the principle of building a module for the input graph.

[0059] Figure 2 This is a schematic diagram illustrating the principle of feature fusion.

[0060] Figure 3 This is a schematic diagram of the training and testing process framework for a high-speed train brake pad status monitoring system based on an improved graph neural network.

[0061] Figure 4 This is a schematic diagram of friction block failure types.

[0062] Figure 5 This is a schematic diagram illustrating the working principle of braking friction.

[0063] Figure 6 The visualization results of the extracted features in a two-dimensional plane are shown below; (a) corresponds to the original signal input to the feature extraction module, (b) corresponds to the visualization of the first layer graph convolution signal in a single space, taking the feature space as an example, and (c) corresponds to the fused features obtained by the feature extraction module.

[0064] Figure 7 The results show the accuracy of brake pad state recognition under different models.

[0065] Figure 8 The confusion matrix represents the brake pad state identification results under different models.

[0066] Figure 9 The search results are for the topology space and feature space weight optimization process in FT-GCN.

[0067] Figure 10 The results show the brake pad status identification for various models with an unbalanced data ratio of 10:1.

[0068] Figure 11 The results show the accuracy of brake pad state identification for various models under different balance data ratios. Detailed Implementation

[0069] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] Example 1

[0071] The high-speed train brake pad status monitoring system based on improved graph neural network provided in this embodiment includes an input matrix construction module, an input graph construction module, a feature extraction module, and a fault identification module.

[0072] 1. Input matrix construction module

[0073] The input matrix construction module is used to construct a node feature matrix and a relationship matrix representing the similarity of each node based on the collected original signals of the friction block.

[0074] The original signal is set to include N nodes, each containing n feature vectors; the features of the N nodes constitute the node feature matrix X; the relationship between two nodes is calculated using cosine similarity, that is, the cosine value of the angle between two vectors in the vector space is used as a measure of the difference between the two individuals.

[0075]

[0076] In the formula, X i and X j Let A represent two node vectors respectively. ij This represents the similarity between two node vectors.

[0077] Initially, the cosine similarity between all nodes forms a relation matrix A, which contains K1 nodes; K1 = N.

[0078] 2. Input graph construction module

[0079] The input graph construction module is used to construct feature graphs and topology graphs based on the node feature matrix and relation matrix; and to optimize the constructed feature graphs and topology graphs using Bayesian optimization. Figure 1 The schematic diagram of the input graph construction module is given.

[0080] The input graph construction module includes the following sub-modules:

[0081] The topology graph construction submodule generates a topology graph from the relation matrix and the node feature matrix.

[0082] Here, the topology graph G = (A, X) is generated from the relation matrix A and the node feature matrix X.

[0083] The first Bayesian optimization module is used to optimize the generated topology graph.

[0084] The feature map construction submodule first uses the KNN algorithm to obtain the adjacency matrix based on the similarity of each node in the relation matrix, and then generates the feature map from the adjacency matrix and the node feature matrix.

[0085] Here, based on the relation matrix A, the nodes are sorted in descending order of cosine similarity. The top K2 similar nodes are selected for each node to form edges, resulting in the adjacency matrix A. f :

[0086]

[0087] In the formula, m and n represent two nodes, and E represents the set of edges associated with the nodes.

[0088] Then from relation matrix A f The feature map GF = (A) is generated from the feature matrix X of the nodes. f (X).

[0089] The second Bayesian optimization module is used to optimize the generated feature map.

[0090] Here, the first Bayesian module and the second Bayesian module use the Bayesian optimization algorithm to optimize the topology graph and the feature graph respectively (i.e., optimize K1 and K2) to obtain the topology graph and the feature graph with optimal K1 and K2.

[0091] In this embodiment, labeled samples can be used to implement Bayesian optimization using the optuna library to obtain the optimal K1 and K2.

[0092] 3. Feature Extraction Module

[0093] The feature extraction module includes two parallel, identically structured graph convolutional units and a feature fusion unit. The two graph convolutional units are used to extract the first feature and the second feature based on the feature map and the topology map, respectively. The feature fusion unit is used to combine the first feature and the second feature with adaptive weights to obtain the fused feature.

[0094] The feature extraction module includes two parallel, structurally identical graph convolutional units (GCNs) and feature fusion units.

[0095] The graph convolutional unit consists of two identical graph convolutional layers arranged sequentially. Each convolutional layer includes several filtering layers, followed by activation functions and pooling layers. The output of the second convolutional layer is normalized to fix its range. The pooling layer summarizes the node features, generating high-level features that capture the information of the entire graph. After a series of graph filtering and activation layers, the pooling layer passes through a graph pooling layer to produce a coarsened graph with more abstract and higher-level node features. Through multiple iterations of the above image convolutional unit, multiple blocks are used to extract fault features from the input.

[0096] A feature fusion unit is placed after two parallel graph convolutional units. The feature fusion unit uses a Bayesian network to extract weights for these classification features. Then, the topology graph and feature map are weighted and summed to obtain the fused features. For example... Figure 2 As shown, the features (a) of the normalized topology graph and the features (b) of the feature graph are in the range of 0-1. The weights (W1 and W2) corresponding to the two are extracted by a Bayesian network, and the features are then assigned weights to obtain the fused features (c). This helps to improve the accuracy of system decision-making.

[0097] 4. Fault Identification Module

[0098] The fault identification module is used to determine the fault category based on the obtained fused features. Here, the argmax function is used to determine the fault category based on the fused features.

[0099] The training method for the high-speed train brake pad state monitoring system based on the improved graph neural network includes the following steps:

[0100] A1 divides the training data samples into a training set and a validation set.

[0101] A2 uses the input matrix construction module to process the training data samples and obtain the corresponding node feature matrix and the relationship matrix representing the similarity of each node.

[0102] A3 first uses the node feature matrix and relation matrix of the training data samples to obtain the topology graph and feature graph using the input graph construction module; then, using the obtained topology graph and feature graph, it obtains the node parameters K1 of the topology graph and K2 of the feature graph based on Bayesian optimization.

[0103] Here, the optimization range of K1 is set to [2-16], and the optimization range of K2 is set to [10-100]. Using the obtained topology map and feature map, the first Bayesian optimization module and the second Bayesian optimization module are trained using the optuna library to obtain the optimal parameters K1 of the first Bayesian optimization module and the optimal parameters K2 of the second Bayesian optimization module.

[0104] A4 uses the training set data processed by the input matrix construction module and the trained input graph construction module to train the graph convolutional unit and the fault identification module, respectively, which are suitable for topology graphs and feature graphs, and fixes the network parameters of the graph convolutional unit.

[0105] Here, the graph convolutional unit consists of two identical graph convolutional layers set sequentially; the number of feature maps in the graph convolutional layer is set to 100. During training, the Adam optimizer is used with a learning rate of 0.001 and the categorical_crossentropy loss function; the training loop iterations are 150; after the loop iterations are completed, the network parameters of the graph convolutional unit are fixed.

[0106] A5 uses the validation set data processed by the input matrix construction module and the trained input graph construction module to obtain the weights of the output features of the two graph convolution units in the feature extraction module composed of two graph convolution units, and completes the training of the high-speed train brake pad status monitoring system through Bayesian optimization to obtain the weights of the feature fusion unit in the output features of the two graph convolution units.

[0107] Here, the weights of the output features of the two graph convolutional units are set as W1 (topology graph feature extraction weight) and W2 (feature map feature extraction weight), respectively. The optimization range of weights W1 and W2 is [0-1] and W1+W2=1, and the number of optimizations is set. Through Bayesian optimization, the weights of the output features of the two graph convolutional units are adaptively adjusted to extract the weights W1 and W2 corresponding to the highest accuracy of the system on the validation set.

[0108] Example 2

[0109] This embodiment provides a method for monitoring the status of brake pads on high-speed trains based on an improved graph neural network, which includes the following steps:

[0110] S1 constructs a node feature matrix and a relationship matrix representing the similarity of each node based on the collected original signals of the friction block.

[0111] In this step, the similarity of each node is calculated from the cosine similarity between each node.

[0112] S2 constructs feature maps and topology maps based on the node feature matrix and relation matrix; and optimizes the constructed feature maps and topology maps using Bayesian optimization; specifically, it includes the following steps:

[0113] S21 generates a topology graph from the relation matrix and the node feature matrix, and then optimizes the generated topology graph using a Bayesian optimization algorithm.

[0114] Here, we select the topology graph constructed based on node parameter K1.

[0115] S22 first uses the KNN algorithm to obtain the adjacency matrix based on the similarity of each node in the relation matrix. Then, it generates a feature map from the adjacency matrix and the node feature matrix. Finally, it optimizes the generated feature map using the Bayesian optimization algorithm.

[0116] Here, we select the feature map constructed based on node parameter K2.

[0117] S3 extracts the first and second features based on the feature map and the topology map respectively, and then combines the first and second features with adaptive weights to obtain the fused feature; specifically, it includes the following steps:

[0118] S31 performs graph convolution processing on the optimized topology graph and feature graph respectively to extract the first feature and the second feature;

[0119] S32 obtains the fused feature by weighting and summing the first and second features based on their weights.

[0120] S4 determines the fault category based on the obtained fusion features.

[0121] In this step, the argmax function can be used to determine the fault category based on the fusion characteristics.

[0122] Application examples

[0123] This application example collects fault status signal data on friction block imbalance in the high-speed train braking system, and uses this data to train and test the high-speed train brake pad condition monitoring system (FT-GCN network) based on improved graph neural networks provided in Example 1. Figure 3 As shown, the overall solution of this application example consists of four main steps: signal acquisition, multi-space graph construction, model training and integration, and verification and application.

[0124] (1) Collect raw data (using vibration acceleration as an indicator) of friction blocks during braking under different fault conditions. Divide each dataset into training, validation, and test sets according to different proportions, and ensure that the labels of each dataset are consistent. The friction blocks are brake pads from high-speed train brakes and scaled proportionally.

[0125] (2) Establish different spatial graphs. The relationships between points in the dataset are derived using cosine relations, constructing a topological graph and a feature graph. Bayesian optimization is then used to optimize the parameters of the topological and feature graphs, ultimately selecting the optimal parameters to construct the final graph. In this step, the differences in node relationships and topological structure are considered, and graph signals are constructed based on different spaces (feature space and topological space). When constructing the input graph, node features propagate simultaneously from both the feature space and the topological space, extracting the information most relevant to the node labels from both spaces.

[0126] (3) The network of the high-speed train brake pad status monitoring system provided in Example 1 is trained using the training set of each dataset, and the weight parameters of the model are saved; then the network is integrated, and the optimal integrated weights are determined by combining the validation set of each dataset with the Bayesian optimization method.

[0127] (4) Use the optimal combination of ensemble weights to perform the final ensemble of the model, and finally test the monitoring effect on the test set.

[0128] (I) Signal Acquisition

[0129] In the experiment, to simulate different braking failure states, based on previous literature review and data collection, four actual failure types were selected to process the friction block, including normal wear, uneven wear, non-uniform wear, and spalling. Figure 4 As shown. Based on the leading edge characteristics of different friction blocks, the specific experimental settings are as follows:

[0130] (1) Normal: The friction block is in normal working condition, such as Figure 4 As shown in (a).

[0131] (2) Uneven Wear: When the friction block is under conditions of ramp and uneven braking force, the stress at the cut end is relatively large. Prolonged unilateral high stress leads to uneven wear of the friction block, a common phenomenon that reduces the effective contact area between the friction block and the brake disc. Uneven wear accelerates the wear of the friction pad itself, thus affecting braking performance. In this application example, the uneven wear of the friction block is set at 1mm. Figure 4 As shown in (b).

[0132] (3) Uniform Wear: In actual high-speed trains, the installation methods of friction blocks vary, and a single brake pad contains multiple friction blocks. Due to installation methods and arrangements, some friction blocks experience uniform force at both ends, preventing uneven wear. However, the thickness of the friction blocks decreases after long-term operation, leading to a decline in friction performance. To simulate this phenomenon, this paper sets the uniform wear of the friction blocks at 3mm. Figure 4 As shown in (c).

[0133] (4) Peeling: Due to unstable friction-induced vibration or thermo-coupling, the friction block may form a hardened layer on its surface, which easily leads to cracks on the friction surface, and these cracks are perpendicular to the friction surface. When the temperature is high, the friction block will peel off. In this application example, a peeled friction block after long-term operation is selected as the simulated friction block, such as... Figure 4 As shown in (d).

[0134] This simulates the braking process of a high-speed train, where the brake disc and friction blocks come into contact, generating friction on the contact surface, accompanied by wear and noise from the friction blocks. Figure 5As shown, braking experiments were conducted on different types of friction blocks. Vibration acceleration signals of the friction blocks were collected (measured by a triaxial accelerometer mounted on the back plate of the friction block), mechanical signals acquired during the friction process (measured by a triaxial force sensor mounted between the connecting rod and the push rod), and noise signals generated during the friction process (measured by a microphone 100mm away from the braking contact surface). This yielded the results of the influence of friction on the friction blocks. Since acceleration signals are more representative than force and noise signals, and tangential vibration acceleration has higher excitation energy in acceleration signals, this paper selected tangential vibration acceleration as the representative signal of the friction blocks.

[0135] (II) Data Preprocessing

[0136] Experimental data were collected for friction blocks with different fault types: (1) Normal; (2) Unevenwear; (3) Uniform wear; and (4) Exfoliation. To better represent the fault information of different types of data, each group of signals collected in the experiment was divided into 1024 segments for feature extraction, with tangential vibration acceleration as the feature index. When constructing the imbalanced dataset, 2000 samples were selected for each fault type, and datasets A, B, C, D, and E were constructed according to Table 1. Each dataset included a training set, a validation set, and a test set. Considering multiple imbalance states, the data imbalance ratio was set to 100:1, 50:1, 20:1, 10:1, and 5:1, respectively. The number of samples for each fault type was the same in the validation set and the test set, as shown in Table 1.

[0137] Table 1 Dataset Classification

[0138]

[0139] (III) System Training and Testing

[0140] First, following the system training method provided in Example 1, the high-speed train brake pad status monitoring system based on the improved graph neural network was trained on an NVIDIA GeForce RTX 2060 graphics processor.

[0141] Then, following the improved graph neural network-based high-speed train brake pad state monitoring method provided in Example 2, the trained high-speed train brake pad state monitoring system was tested using a test set. To reduce the influence of randomness, each experiment was repeated 20 times, and the average value was taken as the final fault monitoring result. The monitoring results are shown in Table 2. As can be seen from the table, under different braking imbalance datasets, the system's fault identification rate reached over 85% in terms of macro precision, macro recall, macro F1-score, and accuracy. The system's recognition accuracy increased as the dataset ratio decreased, reaching a maximum of 99.73% when the training set ratio was 5:1.

[0142] Table 2 Test results of test sets for different fault types

[0143]

[0144] The t-SNE dimensionality reduction technique is used to visualize the features extracted during the testing process of test set D in a two-dimensional plane, such as... Figure 6 As shown. Labels 0, 1, 2, and 3 correspond to four fault types: normal operating condition, uneven wear, uniform wear, and peeling. Figure 6 It can be seen that, in terms of separability, the fault features representing uneven wear and peeling, normal wear and uniform wear fault conditions only have a small number of spatial clusters and confusions, indicating that the high-speed train brake pad condition monitoring system based on improved graph neural network provided by the present invention can effectively distinguish the features of each fault condition. In addition, it also has high performance in clustering.

[0145] 1. Optimize effectiveness analysis

[0146] To verify the effectiveness of considering feature and topology space fusion and graph optimization when constructing the network input graph, the high-speed train brake pad condition monitoring system based on improved graph neural networks (FT-GCN) is compared with models that only consider a single space and do not optimize the graph on dataset D. The structural settings of each model are shown in Table 3. Among them, F-GCN-a is the GCN model that only considers the feature space as input, T-GCN-a is the GCN model that only considers the topology space as input, FT-GCN-a is the GCN model that considers both topology and feature space as input, and the remaining three models are the corresponding optimization models.

[0147] Table 3 Comparison of Model Structures

[0148]

[0149] 1.1 Effectiveness Analysis of Graph Construction and Graph Optimization

[0150] Each model architecture was trained using the training set, and then the accuracy of brake pad state recognition for each model was analyzed on the test set D. The results are as follows: Figure 7 As shown in the figure. From the figure, we can conclude that:

[0151] (1) The results of FT-GCN are consistently better than all other five models, with an accuracy rate of over 98%. This demonstrates the effectiveness of the monitoring system provided by this invention in simultaneously considering the feature space and topological space and optimizing the graph space.

[0152] (2) The accuracy results of F-GCN and T-GCN are usually better than those of F / T-GCN-a, which verifies that optimizing graph construction can significantly improve the accuracy of the model.

[0153] (3) The diagnostic results of FT / F / T-GCN are more concentrated than those of the unoptimized network, indicating that Bayesian optimization plays an efficient aggregation role in this framework.

[0154] (4) By comparing FT-GCN-a with the other four variants, it can be found that even with only spatial fusion and no optimization, the proposed framework still achieves more competitive performance than single-space fusion, demonstrating its stability and competitiveness.

[0155] The recognition results of FT-GCN, F-GCN, and T-GCN are displayed in the confusion matrix, such as Figure 8 As shown in the figure, the vertical axis represents the true label of the sample, and the horizontal axis represents the network's predicted label. Labels 0, 1, 2, and 3 represent the four states of the friction block: normal, uneven wear, uniform wear, and peeling. The figure shows that the accuracy of class 0 samples decreases slightly with fusion, because some features of the topological space are lost, resulting in certain defects in the fused features. However, the accuracy of the remaining three classes significantly improves after fusion, and the overall accuracy is better than that of the single-space GCN.

[0156] 1.2 Effectiveness Analysis of Weight Optimization

[0157] To verify the superiority of the fusion of topology and feature space, Figure 9 The final search results for topology graph feature extraction weight W1 and feature map feature extraction weight W2 on the validation set D are presented. In the figure, the x-axis represents the weight of topology graph feature extraction, the y-axis represents the weight of feature map feature extraction, and the points represent the accuracy of the monitoring system on the validation set D. From... Figure 9 The trend at the midpoint shows that the accuracy reaches its maximum of 99.1% when the weights are W1 = 0.59 and W2 = 0.41. To achieve higher accuracy, the features trained in the feature space are assigned more weights, meaning that information in the feature space is more important than information in the topological space. The result of the weight allocation also corresponds to... Figure 7The diagnostic performance of each single-space model is improved because the feature space contains more fault information; selecting the feature space as the main space can optimize model performance. Figure 9 As can be seen, the FT-GCN of the monitoring system provided by the present invention can improve the diagnostic performance of GCN in the fusion space, while adaptively allocating better weight values ​​for more important information.

[0158] To further verify the necessity and adaptability of considering weights after feature fusion, different experiments were conducted for comparative analysis. Test sets with different imbalance ratios were re-input into the trained monitoring system and integrated according to the determined weights. The final diagnostic results of the model on the test set are shown in Table 4 (Experiments 1 and 2 represent the accuracy of GCN in handling different imbalance ratios after considering only a single space for graph construction; Experiment 3 represents the accuracy of GCN when considering both spaces simultaneously and with a fixed weight ratio of 1:1; Experiment 4 represents the accuracy of the improved GCN provided in this invention with the optimized weights). Comparing Experiments 1, 2, and 3, it can be seen that multi-source spaces can generate more comprehensive feature information than single-source spaces, and the proposed monitoring system can achieve higher diagnostic accuracy. However, comparing Experiments 1 and 2 shows that the superiority of the two networks fluctuates when processing datasets with different proportions. Therefore, optimizing the weights can further improve the diagnostic accuracy of the system, thereby enabling the system to reach its optimal state.

[0159] Table 4. Accuracy of different networks under different imbalance ratios

[0160]

[0161] 2. Comparative Analysis of Different Networks

[0162] To verify the superiority of the proposed network structure, it is compared with the current mainstream imbalance diagnosis models: CNN without sample processing (No sample), CNN based on sample oversampling (SMOTE) (Chawla NV, Bowyer KW, Hall LO, et al. SMOTE: synthetic minority over-sampling technique[J]. Journal of artificial intelligence research, 2002, 16: 321-357), CNN based on sample undersampling (under-sampling) (using random undersampling, randomly selecting some samples from the majority class to remove), CNN based on mixed sampling (SMOTEENN) (Batista GE, Prati RC, Monard M CA study of the behavior of several methods for balancing machine learning training data[J]. ACM SIGKDD explorations newsletter, 2004, 6(1): 20-29) and the F / T-GCN (representing the ordinary GCN model) proposed in Table 3. The first four parameters all use standard CNN networks, while the parameters for the last two are consistent with those described above. The weights of the monitoring system model proposed in this invention are selected based on the optimal weight ratio during training. First, a representative imbalanced dataset D is used for analysis, and the experimental results are shown in Table 5. The results of the normal state diagnosis show that, under imbalanced data, each network can accurately distinguish data with a large sample size. Table 5 shows that the diagnostic accuracy of ordinary neural networks and different data processing methods for brake pad data faults is only 89.1%, 95.3%, 90.7%, and 95.7%, respectively, while GCN achieves accuracy of 92.57% and 94.85%. This indicates that graph neural networks have a greater advantage in processing imbalanced data, but still have shortcomings compared to other networks after data processing. To compensate for this deficiency, this paper considers fusing two spatial systems to improve the feature information extracted by graph neural networks and incorporates optimization during the spatial construction process. The experimental results are significantly better than ordinary neural networks after data processing, achieving a diagnostic accuracy of 98.7%.

[0163] Table 5. Braking diagnostic accuracy (%) for an imbalance ratio of 10:1

[0164]

[0165] Figure 10The figure shows the identification results at an imbalance ratio of 10:1. It can be seen from the figure that the ordinary GCN model has relatively clear diagnostic results for all fault types, but some misclassifications still exist. This is because braking friction is a special and complex friction condition, and single-space fault information is difficult to accurately reflect the corresponding fault condition. There are various methods for constructing GCN graphs, and different graph spaces calculate the relationships between adjacent nodes differently, resulting in incomplete feature information and limited feature knowledge. Based on... Figure 10 (g) It can be seen that the number of misjudged samples is significantly reduced after combining two different spaces.

[0166] To comprehensively evaluate the model performance, accuracy, recall, F1 score, and precision were selected as evaluation metrics to compare and analyze the test results of each model before and after integration. Fault diagnosis results under different imbalance rates were tested, with ten tests performed for each experiment. The results are shown in Table 6. Figure 11 As shown, GCN generally outperforms other neural networks when handling imbalanced data. However, this model is less reliable and fault-tolerant in diagnosing and evaluating data with different balance ratios in a single space compared to a multi-space model. Furthermore, single-space information exhibits varying degrees of uncertainty for each type of fault wear, indicating that different graph spatial information significantly impacts diagnostic accuracy. This is primarily due to the incompleteness of information constructed from a single-space graph, leading to uncertainty in diagnostic results. The monitoring system model FT-GCN provided by this invention effectively integrates multi-space information, resulting in significantly better diagnostic results than other models and substantial improvements in multiple evaluation metrics.

[0167] Table 6. Recall, F1 score, and precision (%) for different models

[0168]

[0169]

[0170] from Figure 11 It can be seen that as the sample size of imbalanced fault data increases, the diagnostic accuracy of various methods for imbalanced faults also gradually increases, but graph neural networks (GCNs) maintain a high accuracy even at a ratio of 100:1. This indicates that the proposed GCN has more advantages in handling severely imbalanced data. Furthermore, a comparison between F-GCN and T-GCN shows that the graph neural network in the feature space performs worse than the GCN in the topological space when handling highly imbalanced data. However, as the data volume increases, the performance of the GCN in the feature space gradually surpasses that of the topological space, indicating that the superiority of GCNs constructed from different spatial graphs varies under different imbalance ratios. Therefore, considering the fusion of the two spatial systems is necessary.

[0171] In summary, obtaining fault state data of friction blocks during the actual braking process of high-speed train brake pads is quite difficult. Normal samples are abundant, while fault state data is scarce, leading to a data imbalance problem when monitoring the health status of brake pads online. This invention collects data on friction blocks under various states and proposes a targeted method for identifying the state of friction blocks under imbalanced data. Unlike other existing studies, FT-GCN considers the relationships between feature space and topological space to expand the identifiable information of nodes when constructing the network input graph. The proposed network optimizes and selects the best edge relationships using a Bayesian algorithm, while simultaneously integrating weight features from multiple spaces for learning. Experimental results show that this invention can effectively monitor the state of brake pads under imbalanced conditions and exhibits higher recognition performance even on balanced datasets compared to conventional networks. The following conclusions can be drawn from the experiment: (1) The present invention can effectively integrate the topological relationship between the two spaces and accurately extract the features in the signal data; by weighting and fusing the features after processing the two spaces, a high accuracy can be achieved; (2) Compared with other mainstream intelligent diagnostic models, the present invention has better performance in terms of stability and accuracy when processing brake pad imbalance data.

[0172] In summary, FT-GCN can effectively identify brake imbalance data during condition monitoring, thus providing a new approach for monitoring the condition of brake pads on high-speed trains.

Claims

1. A high-speed train brake pad state monitoring system based on improved graph neural network, characterized in that, Comprise: The input matrix construction module is used for constructing the node feature matrix and the relationship matrix representing the similarity of each node according to the collected friction block original signal; The input graph construction module is used for constructing the feature graph and the topology graph according to the node feature matrix and the relationship matrix, and optimizing the constructed feature graph and topology graph by using Bayesian optimization; The input graph construction module comprises the following sub-modules: The topology graph construction submodule generates the topology graph from the relationship matrix and the node feature matrix; The first Bayesian optimization module is used for optimizing the generated topology graph; The feature graph construction submodule first obtains an adjacency matrix according to the similarity of each node in the relationship matrix based on the KNN algorithm, and then generates the feature graph from the adjacency matrix and the node feature matrix; The second Bayesian optimization module is used for optimizing the generated feature graph; The feature extraction module comprises two parallelly arranged graph convolution units and a feature fusion unit which are of the same structure; the two graph convolution units are respectively used for extracting first features and second features according to the feature graph and the topology graph; and the feature fusion unit is used for combining the first features and the second features to obtain fused features by weighted fusion with adaptive weights; The fault recognition module is used for determining the fault category according to the obtained fused features.

2. The high-speed train brake pad state monitoring system based on improved graph neural according to claim 1, wherein, The input matrix construction module, the original signal comprises N nodes, each node contains n feature vectors; N node features constitute a node feature matrix X; the relationship between two nodes is calculated by using cosine similarity, that is, the cosine value of the angle between two vectors in the vector space is used as a size measurement for measuring the difference between two individuals; ； In the formula, X i and X j respectively represent two node vectors, and Aij represents the similarity between the two node vectors.

3. The high-speed train brake pad state monitoring system based on improved graph neural according to claim 1, wherein, The graph convolution unit is composed of two graph convolution layers which are arranged in sequence and are of the same structure; each convolution layer comprises a plurality of filtering layers, an activation function located after each filtering layer, and a pooling layer.

4. The training method of the high-speed train brake pad state monitoring system based on the improved graph neural network according to any one of claims 1 to 3, comprising the following steps: A1 divide the training data sample into a training set and a validation set; A2 process the training data sample by using the input matrix construction module to obtain the corresponding node feature matrix and the relationship matrix representing the similarity of each node; A3 first, the node feature matrix and the relationship matrix of the training data sample are used to obtain the topology graph and the feature graph by using the input graph construction module; then, the node parameters K1 of the topology graph and the node parameters K2 of the feature graph are obtained based on Bayesian optimization by using the obtained topology graph and feature graph; A4 train the graph convolution unit and the fault recognition module suitable for the topology graph and the feature graph by using the training set data processed by the input matrix construction module and the trained input graph construction module, and fix the network parameters of the graph convolution unit; A5 the feature extraction module composed of two graph convolution units and a feature fusion unit is trained by using the validation set data processed by the input matrix construction module and the trained input graph construction module, and the weight of the feature fusion unit to the output features of the two graph convolution units is obtained by Bayesian optimization, thereby completing the training of the high-speed train brake pad state monitoring system.

5. A high-speed train brake pad state monitoring method based on improved graph neural, characterized in that, Comprise the following steps: S1, constructing a node feature matrix and a relation matrix representing the similarity of each node according to the collected original signal of the friction block; S2, constructing a feature map and a topology map according to the node feature matrix and the relation matrix; and optimizing the constructed feature map and topology map by using Bayesian optimization; step S2 includes the following sub-steps: S21, generating a topology map from the relation matrix and the node feature matrix, and optimizing the generated topology map by using a Bayesian optimization algorithm; S22, first, obtaining an adjacency matrix based on the KNN algorithm according to the similarity of each node in the relation matrix, and then generating a feature map from the adjacency matrix and the node feature matrix; then, optimizing the generated feature map by using a Bayesian optimization algorithm; S3, extracting first features and second features according to the feature map and the topology map respectively, and combining the first features and the second features with adaptive weights to obtain a fused feature by weighted fusion; S4, determining the fault category according to the obtained fused feature.

6. The high-speed train brake pad state monitoring method based on the improved graph neural network according to claim 5, characterized in that, In step S1, the similarity of each node is calculated by the cosine similarity between each node.

7. The high-speed train brake pad state monitoring method based on the improved graph neural network according to claim 5, characterized in that, Step S3 includes the following sub-steps: S31, performing graph convolution processing on the optimized topology map and feature map respectively to extract first features and second features; S32, weighting and summing the first features and the second features based on the first feature weights and the second feature weights to obtain a fused feature. ​

Images