Lightweight train bogie transmission system fault diagnosis method and system

Abstract

The invention discloses a lightweight train bogie transmission system fault diagnosis method, which comprises the steps of S1, collecting sensor data through signal acquisition equipment, and dividing samples through a sliding window with an overlapping rate for training, verification and testing; s2, folding the original one-dimensional data collected in the S1 into two-dimensional data along a time sequence dimension, and inputting the two-dimensional data into a sparse self-attention network for extracting attention features of fused attention information; s3, inputting the attention features extracted in the S2 into a KANs classifier for classification so as to enhance the internal interpretability of the model; s4, quantifying the feature contribution degree of the attention features enhanced in the S3 through an SHAP method, so as to improve the extrinsic interpretability of the model; and S5, finally outputting a visual diagnosis result and interpretability analysis according to the quantized feature contribution degree in the step S4.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit, in particular to a light-weight train bogie transmission system fault diagnosis method and system. BACKGROUND

[0002] As the main artery of urban transportation, the safe and efficient operation of the rail transit system is directly related to passenger safety and urban operational efficiency. With the increase of train running speed, real-time fault diagnosis and rapid response capability of the bogie transmission system under high-speed working conditions have become a key technical bottleneck. Under the background of Industry 4.0, the innovation of intelligent manufacturing and automation technology poses a double challenge to the fault diagnosis model: it needs to optimize the occupation of computing resources, and it needs to strengthen the interpretability of the model. Although the existing fault diagnosis model has high diagnostic accuracy, its high computational complexity and lack of interpretability hinder industrial deployment. Therefore, the development of a fault diagnosis method with light weight and high interpretability has become an urgent breakthrough direction to ensure the transparency and rapid response capability of the safety decision of the high-speed train bogie transmission system.

[0003] CONTENT

[0004] The purpose of the present application is to provide a light-weight train bogie transmission system fault diagnosis method and system, and the specific technical solutions are as follows:

[0005] A light-weight train bogie transmission system fault diagnosis method includes: S1, collecting sensor data through a signal acquisition device, and dividing samples through a sliding window with a set overlap rate for training, verification and testing; S2, inputting the original one-dimensional data collected in S1 into a sparse self-attention network after folding the data along the time sequence dimension into two-dimensional data, for extracting attention features that fuse attention information; S3, inputting the attention features extracted in S2 into a KANs classifier for classification, for enhancing the intrinsic interpretability of the model; S4, quantifying the feature contribution of the enhanced attention features in S3 through a SHAP method, for improving the extrinsic interpretability of the model; S5, according to the quantified feature contribution in S4, finally outputting a visual diagnosis result and an interpretability analysis.

[0006] The sparse self-attention network in S2 is a sparse processing of the attention mechanism, wherein the attention mechanism calculation steps include: constructing a weight matrix W Q ,W K ,W V of the model query (Q), key (K) and value (V), and the dimensions are all (d model ×d model ), and the input data matrix has a dimension of (L seq ×d model)) respectively multiplied by each weight matrix to generate corresponding Q, K, V matrices, the number of floating point operations in this step is 3·L seq ·d model 2 , the parameter quantity is Calculate the dot product of each row of the query matrix and the transpose key matrix, and get the attention weight by softmax normalization, the number of floating point operations in this step is L seq 2 ·d model ; the attention weight is weighted multiplied with the value matrix to generate the final output, the number of floating point operations in this step is L seq 2 ·d model ; the total floating point operation number of the attention mechanism is:

[0007] FLOPs(Attention)

[0008] =FLOPs(Q,K,V)+FLOPs(softmax(Q,K))+FLOPs(Attention(Q,K,V))

[0009] =FLOPs(3·L seq ·d model 2 +2·L seq 2 ·d model );

[0010] The total parameter quantity of the attention mechanism ignoring the influence of the bias term on the parameter quantity is:

[0011]

[0012] The sparse processing of the attention mechanism includes: obtaining attention scores through the dot product operation of the query matrix and the key matrix, and selecting the top k elements with the highest scores for subsequent calculation, that is, the number of floating point operations required is reduced from (L seq 2 ·d model ) to (L seq ·k·d model ).

[0013] The folding of the original one-dimensional data into two-dimensional data along the time dimension in S2 is: by folding and converting the dimension of the input data from (batch_size, N) to (batch_size, L seq ,d model ), where N=L seq ×d model; that is, assuming the dimension of the original input data is (batch_size, 512), it can be folded into (batch_size, 32, 16), which means that there are 32 sequences in each batch sample, and each sequence has 16 feature dimensions.

[0014] In S3, KANs parameterize each weight by B-spline spline function, which realizes the nonlinear expression of the weight, and the specific implementation is as follows:

[0015] The input vector x = (x1, x2, ···, x n ), and the function f can be expressed as:

[0016]

[0017] where Φ q is the outer function, and φ q,p is the inner function, both of which are learnable parameterized one-dimensional functions; the inner function φ q,p is realized by B-spline function, which is a piecewise polynomial function, and is a polynomial in each grid interval, and is smooth and continuous in the entire definition domain, each φ q,p is parameterized as a B-spline function with trainable coefficients;

[0018] φ(x) = w b b(x) + w s spline(x)

[0019] where w b and w s are trainable coefficients of the base activation function b(x) and the spline function spline(x) respectively, and the base activation function b(x) is usually selected as the silu function:

[0020] b(x) = silu(x) = x / (1 + e -x ),

[0021] The spline function spline(x) is defined as a linear combination of B-spline basis functions:

[0022]

[0023] where c i is a trainable coefficient, and B i (x) is a B-spline basis function defined recursively, and the 0-order B-spline basis function is defined as follows:

[0024]

[0025] Here, x is the grid interval [t i ,t i+1 ] node, The value is 1 in the node interval and 0 elsewhere;

[0026] When the order is greater than 1, the basis function is defined as:

[0027]

[0028] Among them, w left is the left weight coefficient, w right is the right weight coefficient. For any point x, the weight coefficient represents the position ratio of x in the interval, w left The larger it is, the closer x is to node t. i+p , the basis function on the left has a large weight, w right The larger it is, the closer x is to node t. i+p+1 , the basis function on the right contributes most to the B-spline function. By adjusting these weights, the continuity and smoothness of the basis function at the nodes are ensured. At the same time, the curvature of the spline curve in a specific interval can be controlled to better fit the data. The outer function Φ q It is also implemented by B-spline, which takes the output of the inner function as input and performs further transformation, so each Φ q It is also a learnable one-dimensional function.

[0029] The SHAP method in S4 determines the SHAP value by evaluating the average marginal contribution of each feature in all possible feature subsets. Specifically:

[0030] For a specific sample x and model f, the prediction result f(x) can be formulated as:

[0031]

[0032] Among them, φ0 is the baseline value, which reflects the average prediction made by the model in the absence of any feature information (when all feature values ​​are 0), φ j (x) represents the SHAP value of the jth feature, indicating the individual contribution of this feature to the model prediction f(x);

[0033] For each feature x in sample x j , whose SHAP value φ j (x) is defined as:

[0034]

[0035] Among them, S means that feature x is not included j|S| is the number of features in the subset S, n is the total number of features in the sample x, represents the result when the model prediction includes the subset S and the feature x j S F(x) is the prediction value of the model only including the subset S, for each subset S not including x j is the change in the model prediction after adding x j j represents the marginal contribution of x j to the prediction result, is the weight of the marginal contribution, which reflects the average influence of x j in all possible subset combinations, the SHAP value of x j is obtained by summing the weighted marginal contributions of all subsets, that is, its average marginal contribution to the model prediction.

[0036] A lightweight train bogie transmission system fault diagnosis system, comprising: a data preprocessing module, the data preprocessing module is used for responsible for data acquisition and sample preparation, using signal acquisition equipment to collect sensor data, and then dividing samples through setting sliding window overlap rate for training, verification and testing; a sparse self-attention module, the sparse self-attention module is used for folding and compressing one-dimensional time series signal into two-dimensional data, and inputting into sparse self-attention network, the sparse self-attention module reduces the complexity and parameter amount of the model while extracting and fusing attention information deep features; an interpretable diagnosis module, the interpretable diagnosis module is used for outputting diagnosis results and providing interpretable analysis, first inputting the extracted attention features into KANs classifier for classification, KANs enhances the internal explainability of the model by dynamically adjusting the shape of the activation function, and then quantifies the feature contribution degree by means of SHAP method, visualizes the correlation between attention features and fault categories, so as to enhance the external explainability of the model.

[0037] The beneficial effects of the present application are that the sparse self-attention module significantly reduces the computational complexity and parameter amount by extracting key features, which are then input into the KANs classifier for classification. KANs enhances the internal explainability of the model by dynamically adjusting the shape of the activation function and visualizing its changes, while the SHAP method quantifies the contribution degree of the sparse self-attention features to enhance the external explainability of the model.

[0038] Drawings

[0039] Figure 1 is a schematic diagram of the sparse self-attention network in the present application, wherein the value of k is set to 3, that is, only three score positions are retained in each query row, and all other positions are set to 0;

[0040] Figure 2 ​Fig. 1 is a schematic diagram of folding original one-dimensional data into two-dimensional data along the time dimension in the present application.

[0041] Figure 3 Fig. 2 is a schematic diagram of the structure of KANs in the present application.

[0042] Figure 4 Fig. 3 is a schematic diagram of the relationship between model parameters and data dimensions in Example 1.

[0043] Figure 5 Fig. 4 is a heatmap of the relationship between model complexity, k, and data dimensions in Example 1.

[0044] Figure 6 Fig. 5 is a schematic diagram of the relationship between fault diagnosis accuracy, k value, and data dimensions in Example 1, where the horizontal axis represents different parameter dimension configurations, the vertical axis shows the diagnostic accuracy index, and the highlighted markers in each subgraph indicate the parameter dimension configuration that achieves the peak diagnostic accuracy.

[0045] Figure 7 Fig. 6 is a schematic diagram of the activation function of the KAN classifier module in Example 1.

[0046] Figure 8 Fig. 7 is a schematic diagram of the SHAP visualization results in Example 1.

[0047] Figure 9 Fig. 8 is a schematic diagram of the SHAP visualization results for other prediction categories in Example 1.

[0048] Figure 10 Fig. 9 is a schematic diagram of the loss and accuracy during training and validation of the method in the present application and comparative methods in Example 1.

[0049] Figure 11 Fig. 10 is a schematic diagram of the confusion matrix results of the method in the present application on the test set in Example 1.

[0050] Figure 12 Fig. 11 is a schematic diagram of the loss curve and diagnostic accuracy of the method in the present application and comparative methods in the training set and validation set in Example 2.

[0051] Figure 13 Fig. 12 is a schematic diagram of the ROC curve in Example 2.

[0052] Figure 14 Fig. 13 is a schematic diagram of the activation function of the KAN classifier module in Example 2.

[0053] Figure 15 Fig. 14 is a schematic diagram of the SHAP visualization results for the normal category in the CWRU data in Example 2.

[0054] Figure 16 Fig. 15 is a schematic diagram of the SHAP visualization results of the method in the present application for the other nine fault categories on the CWRU data set in Example 2. DETAILED DESCRIPTION

[0055] For the purposes of the present invention, the technical solutions and advantages will be more clearly apparent from the following detailed description in conjunction with the accompanying drawings. It should be understood that the description is only exemplary and is not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and techniques are omitted to avoid unnecessarily obscuring the concept of the present invention.

[0056] A light train bogie transmission system fault diagnosis method comprises: S1, collecting sensor data through a signal acquisition device, and dividing samples through a sliding window with an overlap rate for training, verification and testing; S2, inputting the original one-dimensional data collected in S1 into a sparse self-attention network after folding along the time sequence dimension to two-dimensional data, for extracting attention features that fuse attention information; S3, inputting the attention features extracted in S2 into a KANs classifier for classification, for enhancing the intrinsic explainability of the model; S4, quantifying the feature contribution of the enhanced attention features in S3 through a SHAP method, for improving the extrinsic explainability of the model; S5, according to the quantified feature contribution in S4, finally outputting a visual diagnosis result and an explainability analysis.

[0057] As shown in Figure 1 , the sparse self-attention network in S2 is a sparse processing of the attention mechanism, wherein the attention mechanism calculation steps include: constructing a weight matrix W Q ,W K ,W V of the model query (Q), key (K) and value (V), all of which have a dimension of (d model ×d model ), and the input data matrix (dimension (L seq ×d model )) is multiplied by each weight matrix to generate corresponding Q, K and V matrices, and the number of floating point operations in this step is 3·L seq ·d model 2 , and the parameter quantity is The dot product of each row of the query matrix and the transposed key matrix is calculated, and the attention weight is obtained by softmax normalization, and the number of floating point operations in this step is L seq 2 ·d model ; the attention weight is weighted multiplied with the value matrix to generate the final output, and the number of floating point operations in this step is L seq 2 ·d model ; the total number of floating point operations of the attention mechanism is:

[0058] FLOPs(Attention)

[0059] =FLOPs(Q,K,V)+FLOPs(softmax(Q,K))+FLOPs(Attention(Q,K,V))

[0060] = FLOPs(3·L seq ·d model 2 +2·L seq 2 ·d model );

[0061] The total number of parameters of the attention mechanism ignoring the influence of the bias term on the number of parameters is:

[0062]

[0063] The sparse processing of the attention mechanism includes: obtaining the attention score by performing the dot product operation of the query matrix and the key matrix, and selecting the k elements with the highest scores for subsequent calculations, that is, reducing the number of floating-point operations required from (L seq 2 ·d model ) down to (L seq ·k·d model ).

[0064] like Figure 2 As shown, in S2, the original one-dimensional data is folded into two-dimensional data along the time series dimension: by folding the input data dimension from (batch_size, N) to (batch_size, L seq ,d model ), where N = L seq ×d model That is, assuming that the dimension of the original input data is (batch_size, 512), it can be folded into (batch_size, 32, 16), which means that 32 sequences are contained in each batch sample, and each sequence has 16 feature dimensions.

[0065] like Figure 3 As shown in Figure 2, KANs in S3 parameterizes each weight by parameterizing a one-dimensional function using a B-spline function, thereby achieving a nonlinear expression of the weight. Specifically:

[0066] Input vector x=(x1,x2,···,x n ), the function f can be expressed as:

[0067]

[0068] Among them, Φ q is the outer function, φq,p are inner functions, which are all learnable parameterized one-dimensional functions; the inner function φ q,p It is realized by B-spline function, which is a piecewise polynomial function. It is a polynomial in each grid interval and remains smooth and continuous in the entire domain. q,p It is parameterized as a B-spline function with trainable coefficients;

[0069] φ(x)=w b b(x)+w s spline(x)

[0070] Among them, w b ,w s They are the trainable coefficients of the basic activation function b(x) and the spline function spline(x). The basic activation function b(x) is usually chosen as the silu function:

[0071] b(x)=silu(x)=x / (1+e -x ),

[0072] The spline function spline(x) is defined as a linear combination of B-spline basis functions:

[0073]

[0074] Among them, c i is a trainable coefficient, B i (x) is the B-spline basis function defined recursively, the 0th order B-spline basis function The mathematical expression of is defined as follows:

[0075]

[0076] Here, x is the grid interval [t i ,t i+1 ] node, The value is 1 in the node interval and 0 elsewhere;

[0077] When the order is greater than 1, the basis function is defined as:

[0078]

[0079] Among them, w left is the left weight coefficient, w right is the right weight coefficient. For any point x, the weight coefficient represents the position ratio of x in the interval, w left The larger it is, the closer x is to node t. i+p , the basis function on the left has a large weight, w rightThe larger, the closer x is to node t i+p+1 The basis functions on the right side contribute more to the B-spline function. By adjusting these weights, the continuity and smoothness of the basis functions at the nodes are ensured, and the bending degree of the spline curve in a specific interval can be controlled, so as to better fit the data; the outer function Φ q is also realized by B-spline, which takes the output of the inner function as input and performs further transformation, therefore, each Φ q is also a learnable one-dimensional function.

[0080] In S4, the SHAP method is to determine the SHAP value of each feature by evaluating the average marginal contribution of each feature in all possible feature subsets, specifically:

[0081] For a specific sample x and a model f, the prediction result f(x) can be formulaically represented as:

[0082]

[0083] where φ0 is the baseline value, which reflects the average prediction made by the model without any feature information (when all feature values are 0), φ j (x) represents the SHAP value of the jth feature, indicating the individual contribution of the feature to the model prediction f(x);

[0084] For each feature x j in the sample x, its SHAP value φ j (x) is defined as:

[0085]

[0086] where S represents a feature subset that does not contain the feature x j , |S| is the number of features in the subset S, n is the total number of features in the sample x, represents the result of the model prediction containing the subset S and the feature x j , F(x S ) is the prediction value of the model containing only the subset S, for each feature subset S that does not contain x j , is the change in the model prediction after adding x j , representing the marginal contribution of x j to the prediction result, is the weight of the marginal contribution, which reflects the average influence of x j in all possible subset combinations, by summing the weighted marginal contributions of all subsets, the SHAP value of x j is obtained, that is, its average marginal contribution to the model prediction.

[0087] The lightweight train bogie transmission system fault diagnosis system comprises a data preprocessing module, a sparse self-attention module and an explainable diagnosis module. The data preprocessing module is responsible for data acquisition and sample preparation, collects sensor data by using a signal acquisition device, and then divides samples by setting a sliding window overlap rate for training, verification and testing. The sparse self-attention module is used for folding and compressing one-dimensional time series signals into two-dimensional data, and inputting the two-dimensional data into a sparse self-attention network. The sparse self-attention module reduces the complexity and parameter amount of the model while extracting and fusing deep features of attention information. The explainable diagnosis module is used for outputting diagnosis results and providing explainable analysis. First, the extracted attention features are input into a KAN classifier for classification. The KAN enhances the internal explainability of the model by dynamically adjusting the shape of the activation function. Then, the SHAP method is used to quantify the feature contribution degree, and the association between the visualized attention features and the fault categories is visualized, so as to improve the external explainability of the model. The beneficial effects of the present application are that the sparse self-attention module is used to significantly reduce the calculation complexity and parameter amount by extracting key features. These features are then input into a KANs classifier for classification. The KANs enhance the internal explainability of the model by dynamically adjusting the shape of the activation function and visualizing the changes, and the SHAP method quantifies the contribution degree of the sparse self-attention features to improve the external explainability of the model.

[0088] In order to make the present application easier to be understood, the following will be further described in combination with actual experiments.

[0089] Example one

[0090] The BJTU-RAO bogie dataset

[39] is published by the Beijing University of Technology Advanced Rail Autonomous Running National Key Laboratory, and is the first publicly available rail transit bogie transmission system multi-sensor fault simulation dataset. The dataset covers 51 health states of metro bogie transmission system data. The real metro bogie is reproduced at a scale of 1:2. 24 channels of data such as vibration, current, speed and sound are collected. The sampling frequency of each channel is 64 kHz, the single sampling time is 10 seconds, and a total of 640000 data points are obtained. The power transmission chain of the test bench includes a motor, a reduction gear box and an axle box. The motor bearing type is SKF 6205-2RSH; the reduction gear box adopts a helical gear (drive gear tooth number 16, driven gear tooth number 107); the drive gear support bearing type is HRB32305; and the axle box bearing type is HRB 352213.

[0091] We focus on the CH16 channel data beside the gearbox output shaft under the condition of 20 Hz motor speed and 0 kN transverse load. The selection of this condition and channel is based on: this condition provides a stable and simplified environment, which can clearly identify fault features; low speed and no transverse load make the system more sensitive to faults, which is convenient for signal capture and provides ideal conditions for fault diagnosis algorithm verification. The data set contains 9 health states, including 1 normal condition and 8 typical fault types (Table 1).

[0092]

[0093] Table 1

[0094] To provide sufficient data volume and facilitate subsequent model training and testing, the original data was preprocessed. A sliding window was used to extract samples from the vibration signals of each health state, with a window length of 512 and an overlap rate of 0.5. 2499 samples were generated for each health state, and 22491 samples were generated for the 9 classes, corresponding to the generated labels. After randomizing the samples, they were divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer was used for training, with a learning rate of 0.005, a batch size of 64, and 100 epochs. The number of attention heads was set to 8. The KAN network structure was [feature_dim, 9], where feature_dim represented the deep feature dimension extracted by sparse self-attention, and 9 corresponded to the number of categories. The random seed for all experiments was fixed at 20 to ensure reproducibility.

[0095] The operating equipment and environment are as follows: the operating system is Windows 11. The CPU is i5-9600KF. The RAM is 16.0GB. The GPU is Nvidia GeForce RTX 2070SUPER. The software is Pytorch 2.5.1.

[0096] To study the effects of different k values and data dimensions on model parameter size, computational complexity, and diagnostic accuracy, a series of comparative experiments were conducted. The experiments aimed to determine the optimal k value and dimension combination to build the best diagnostic model. The k value was set to [5, 10, 15, 20], and the folded data dimensions were set to [(8, 64), (16, 32), (32, 16), (64, 8), (128, 4)] according to the constraint condition k ≤ seq_len (the size of the first dimension after data folding), which included 5 dimension configurations.

[0097] To verify the effectiveness of the sparse self-attention module, this section first analyzes the effects of different k values and data dimensions on model parameter size and computational complexity. The parameter size of the sparse self-attention module is mainly affected by the feature dimension d kthe influence of k value. Therefore, we explore the total parameter amount of the model under different data shapes, including the bias term. The experimental results are shown in Figure 4 , where the horizontal axis represents the parameter amount, and the vertical axis represents different data dimensions.

[0098] As shown in Figure 4 , the model parameter amount is significantly positively correlated with the feature dimension: when the feature dimension increases from 4 to 64, the parameter amount increases linearly from 420 to 18240.

[0099] Figure 5 The influence of different k values and data dimensions on FLOPs is shown by a heat map, where the horizontal axis is the k value, the vertical axis is the data dimension, and the grid value directly corresponds to the model calculation complexity. The heat map uses color gradient visualization, where dark areas represent high FLOPs, and light areas represent low FLOPs.

[0100] According to the formula

[0101] FLOPs(Attention)

[0102] = FLOPs(Q, K, V) + FLOPs(softmax(Q, K)) + FLOPs(Attention(Q, K, V))

[0103] = FLOPs(3·L seq ·d model 2 +2·L seq 2 ·d model ),

[0104] The model calculation amount is mainly determined by the parameter dimension and the k value. As shown in Figure 5 , when the k value is fixed, the model calculation amount decreases significantly as the parameter dimension increases (e.g., from (128, 4) to (8, 64)). This indicates that increasing the parameter dimension can effectively reduce the calculation demand. Conversely, when the parameter dimension is fixed, increasing the k value will increase the calculation amount, but the increase is relatively flat. This phenomenon is due to the fact that the calculation amount of the sparse self-attention mechanism is simultaneously affected by the k value, the sequence length, and the feature dimension squared term.

[0105] Analysis Figure 6As can be seen, when the parameter dimension is set to (16, 32), the model generally achieves higher diagnostic accuracy. This is because the self-attention mechanism captures time series features by processing time series data of length seq_len. Sequence lengths that are too short do not provide sufficient information to capture subtle fault features, which must accumulate over a sustained period to become apparent. Conversely, sequences that are too long introduce redundant information. Even with sparse self-attention, the model still struggles to focus on key areas, resulting in the weakening or omission of important features.

[0106] Table 2 shows the number of parameters and FLOPs corresponding to the highest diagnostic accuracy for various values ​​of k in the validation set. The model proposed in this application is represented by SAKAN, and SAKAN-5 in the table represents the SAKAN model when k = 5. Although the SAKAN-20 model has the lowest number of parameters, it also has the lowest accuracy, at only 97.41%. In comparison, the SAKAN-10 model performs superior in both accuracy and FLOPs. Therefore, this paper selects the SAKAN-10 model as the optimal diagnostic model.

[0107]

[0108] Table 2

[0109] like Figure 7 Figure 2 illustrates the dynamic, learnable activation function morphology and evolution of the KAN classifier module within the SAKAN framework. Using B-spline parameterization, KAN's activation function is designed as an adaptively adjusted one-dimensional continuous function, whose morphology is driven by training data. The last layer of the classifier module contains 32 nodes, and the output layer contains 9 nodes. Due to the large number of activation functions between layers, only a subset of these are visualized for clarity. For example, act8_3 represents the activation function between the 8th node in the last layer and the 3rd node in the output layer. As training progresses, the activation function morphology stabilizes and no longer changes dramatically. Significant differences are observed between the activation functions of different nodes (e.g., act8_3 and act14_1): act8_3 exhibits a nonlinear, increasing curve that significantly increases after the input feature exceeds a threshold, indicating its selective enhancement of specific fault signatures. In contrast, act14_1 exhibits a nonlinear, decreasing characteristic, suppressing high-amplitude inputs. This complementary nature helps improve the model's detection sensitivity.

[0110] When using the SHAP method to analyze the external interpretability of the diagnostic model and reveal the key role of sparse self-attention features in fault classification, the trained model structure and parameters are first frozen, and deep features containing attention information are extracted through the sparse self-attention module. The attention features are used as the benchmark dataset for SHAP analysis to quantify and visualize the specific contribution of each feature dimension in the KAN diagnostic decision, thereby improving the model transparency. Figure 8As shown, the SHAP visualization result is shown, which shows that the prediction result is the normal class. The horizontal axis of the figure represents the sample, and 50 background data are selected for each class. The first 50 are normal condition class samples, and the others are samples of the other nine fault classes. The left side of the vertical axis represents the features of each dimension. The higher the feature is in the figure, the greater its influence on the model output. The black bars on the right side of the vertical axis represent the contribution of the feature of this dimension to the fault class. The longer the bar, the greater the contribution of the feature to the state. The color of the SHAP value changes from blue (negative influence) to red (positive influence) to represent the size of the SHAP value. Red indicates that the feature increases the model output, and blue indicates that it reduces the model output. The black line at the top of the figure represents the change in the KAN classification layer output value f(x), which is the prediction probability of the corresponding state class. The dashed line in f(x) is the baseline value, which is the output of the model without any input features. The feature column of the normal class samples (0-49) shows a significant red distribution, indicating that these feature dimensions positively enhance the prediction of the normal class. Feature 22 and Feature 6 at the top of the feature column contribute most significantly to the normal class discrimination. In contrast, the feature column of the other class samples (50-449) shows a blue dominant distribution, indicating that it suppresses the model's tendency to output the normal class. The f(x) value of the 0-49 sample is consistently higher than the baseline and other samples, confirming that the KAN diagnostic accurately identifies normal working conditions. It is worth noting that the 350-400 (bearing rolling element fault) sample has an abnormally high output probability, indicating that the model may misclassify the rolling element fault sample as a normal class.

[0111] Further analysis of the model's prediction results for the eight fault classes, Figure 9 shows that it maintains a high confidence. The feature of the gear crack fault sample (50-99) shows a significant red distribution in the prediction, indicating a strong reinforcement effect on the positive class discrimination. The feature contribution of different fault classes differs significantly: the 24th and 4th features contribute significantly to the prediction of the gear missing tooth class, while the 19th and 25th features play a dominant role in the discrimination of the gear broken tooth class. This shows that SHAP analysis explicitly reveals the dependence pattern of SAKAN on key feature dimensions (such as the significant positive contribution of Feature 22 to the Crack class), while traditional deep models can learn feature weight differences but cannot intuitively present the decision logic of specific dimensions. Figure 9The loss and accuracy curves of the method and the baseline methods in the training and validation phase are compared. SAKAN shows higher initial accuracy at the beginning of training. Within the first 15 epochs, the loss convergence speed and accuracy improvement of the baseline methods are significantly lower than that of SAKAN. SAKAN enters a stable state after the 15th epoch, while the baseline method needs to be stable until the 60th epoch. It is worth noting that the validation set of MobileNet-V2, LSTM and other baseline methods still fluctuates after 60 epochs, while SAKAN remains stable. This phenomenon confirms that SAKAN has better generalization performance and data adaptability, which is due to the precise capture of key information by the sparse self-attention mechanism, making the model converge significantly faster than the baseline method.

[0112] As shown in Figure 10 , the loss and accuracy of the method and the comparison method in the training and validation process are shown. The SAKAN method has a higher initial accuracy than all other methods at the beginning of training. In the first 15 epochs, the loss convergence speed and accuracy improvement speed of the comparison method on the training set and the validation set are slower than SAKAN. After 15 epochs, the loss of SAKAN has begun to stabilize, and other comparison methods begin to stabilize after 60 rounds. After 60 rounds, methods such as MobileNet-V2, LSTM have a certain degree of fluctuation in the validation set, while the SAKAN method is more stable. This shows that the method has better generalization ability and adaptability to data.

[0113] In addition to accuracy and FLOPs, precision, recall, and F1 score are also important indicators for evaluating model performance. For binary classification, precision, recall, and F1 score are defined as follows: TP

[0114]

[0115] where TP,FP,and FN are true positive,false positive,and falsenegative respectively. For multi-classification tasks, the above formula can be applied to each class and take the average value.

[0116] The performance comparison results of the test set are shown in Table 3. Although traditional fault diagnosis methods (1DCNN / 2DCNN / LSTM) can achieve good diagnostic results, they have the problem of excessive parameter and computational complexity. MobileNet-V2 reduces FLOPs to a lower level (16.64k) through depthwise separable convolution, but the number of parameters is still as high as 2234.83k; the number of parameters of the Transformer method can be compressed to 17.45k, but the FLOPs increases significantly to 983.33. The method in this paper achieves 21.10k FLOPs with 6.05k parameters, while obtaining the highest accuracy of 98.48%. The comparison of precision, recall rate and F1 score shows that this method exhibits balanced and excellent comprehensive performance in different fault categories.

[0117]

[0118] Table 3

[0119] like Figure 11 As shown in Figure 2, the confusion matrix results of the SAKAN method on the test set are shown. The method has a low misjudgment rate in each fault type. Specifically, the five bearing cage fault samples are misclassified as gear wear tooth fault and bearing rolling element fault. Figure 9 The conclusions are consistent.

[0120] In order to evaluate the feature extraction capability of this method, t-SNE

[40] was used to perform dimension reduction visualization analysis on the sparse self-attention features. The results are shown in the figure below. Figure 11 As shown in the figure, different colors represent different fault types. After extraction using the sparse self-attention model, t-SNE can effectively distinguish features of each category, demonstrating the model's excellent feature extraction capabilities.

[0121] Example 2

[0122] This embodiment uses the acceleration data of the drive-end bearing from the Case Western Reserve University Data Center (CWRU)

[41] . The drive-end speed is 1797 rpm and the sampling frequency is 12 kHz. The data set contains three types of faults: inner race fault (IR), outer race fault (OR), rolling element fault (B), and normal operating conditions. Each fault contains three damage sizes (unit: inch), as listed in Table 4. The sliding window size is 512 and the coverage is 0.5. The training set, validation set, and test set are divided in a ratio of 7:2:1. The output nodes of the SAKAN model correspond to 10 bearing health states (covering 9 types of faults and normal operating conditions). The model training parameters remain unchanged.

[0123]

[0124] Table 4

[0125] like Figure 12As shown, the loss curves and diagnostic accuracy of SAKAN method and comparative methods in the training set and validation set are shown. The loss of SAKAN method decreases faster than other methods (such as MobileNet-V2, LSTM and Transformer) in the early stage of training (first 15 rounds), which indicates that SAKAN can capture the key features of the data faster in the early stage of training. The loss curve of SAKAN method tends to be stable after 15 rounds, while the loss curves of other methods (such as MobileNet-V2 and LSTM) still have large fluctuations in the later stage. This is consistent with the performance in BJTU-RAO dataset. It shows that the method in this paper is more stable in the training process, has faster convergence speed, and has stronger model generalization ability and data adaptability.

[0126] Table 5 shows that the classification accuracy of SAKAN method in the test set is 99.11%. Although the FLOPs index is slightly higher than MobileNet-V2, the parameter quantity is significantly reduced to 6.37k. The comprehensive evaluation index (precision, recall and F1 value) has advantages. It should be noted that the output node of CWRU dataset is 10 (BJTU-RAO is 9), which makes the parameter quantity and FLOPs of SAKAN in CWRU slightly higher than BJTU-RAO dataset.

[0127]

[0128]

[0129] Table 5

[0130] To further verify the performance of the method in this paper in different categories, the model is analyzed in depth through true positive rate (True Positive Rate, TPR) and false positive rate (False Positive Rate, FPR), and the Receiver Operating Characteristic Curve is drawn as shown in Figure 13 The ROC curve is an effective tool for evaluating the performance of classification model, which can intuitively reflect the classification ability of the model by showing the relationship between TPR and FPR. Specifically, the true positive rate represents the proportion of all actual positive samples that are correctly predicted as positive, and the false positive rate represents the proportion of all actual negative samples that are incorrectly predicted as positive. The calculation formula is:

[0131]

[0132] Where TP is the number of true positives, FN is the number of false negatives, FP is the number of false positives, and TN is the number of true negatives.

[0133] FromFigure 13 As can be seen, the model is close to the upper left corner of the ROC curve of all categories, and the area under the curve value is close to 1, while the performance of identifying the failure type of the damage scale of 21 inches is slightly poor, but the area value is also greater than 0.99. It shows that SAKAN can learn to distinguish high-discrimination sparse attention features, thereby distinguishing various fault types well.

[0134] To improve the interpretability of the SAKAN model, the KAN classifier activation function is visualized. As shown in Figure 14 , different node activation functions present different forms, indicating that the model processes features through a differential nonlinear transformation strategy. The morphological diversity helps to capture complex data patterns. It is observed that the activation function form is dynamically adjusted with training iterations, proving that its adaptive characteristics can improve model performance. When the training period increases to 100 rounds, the activation function form tends to be stable, showing that the model has converged.

[0135] When analyzing the interpretability of SAKAN on the CWRU dataset using the SHAP method. Take the sparse self-attention feature as the SHAP benchmark data, and use the SAKAN model with frozen parameters as the target model to quantify and visualize the contribution of each feature dimension in decision-making. Figure 15 The SHAP results of normal class samples (Sample 0-19) are shown, and Feature 6 / 12 at the top of the feature sequence has the highest SHAP value, indicating that these features significantly promote normal state judgment. For Sample 20-199, Feature 6 / 12 presents a dominant blue color, indicating that its features have an inhibitory effect on non-normal class prediction. At the same time, the f(x) output value of Sample 0-19 is significantly higher than that of other samples, proving that the model has the highest prediction confidence for the normal class.

[0136] As shown in Figure 16As shown, the SHAP visualization results of SAKAN model on the CWRU dataset for the other nine fault categories are displayed. Taking fault categories such as IF7, BF7, OF7, and OF14 as examples, it can be observed that the model can dynamically focus on different key features under different fault categories. For example, in the IF7 and BF7 fault categories, Feature 18 and Feature 8 have a significant contribution to the model's prediction of this fault category, with their SHAP values significantly higher than other features, indicating that they also play a key role in diagnosing IF7 and BF7 faults; for the OF7 fault category, Feature 6 and Feature 12 have higher SHAP values, indicating that they are key features for diagnosing this fault. In the OF14 fault category, Feature 1 and Feature 25 have prominent SHAP values, indicating that they have a decisive role in diagnosing IF14 faults. Through these SHAP visualization results, it can be seen that the SAKAN model can dynamically focus on different key features under different fault categories, rather than uniformly utilizing all features.

[0137] In this paper, a lightweight and interpretable fault diagnosis framework SAKAN is developed based on sparse self-attention mechanism. This framework aims to solve the problems of parameter redundancy, high computational complexity, and lack of interpretability in traditional diagnosis models. The effectiveness and interpretability of this method are verified in multiple datasets. The conclusions are as follows:

[0138] The SAKAN framework achieves lightweight design through sparse self-attention mechanism and dimension folding technology, with extremely low parameter quantity (only 6.05k parameters for BJTU-RAO dataset) and computational complexity (FLOPs as low as 21.10k), while achieving high diagnostic accuracy of 98.48% and 99.11% on BJTU-RAO and CWRU datasets, respectively, fully verifying the feasibility of the synergy optimization of lightweight and high precision.

[0139] Experiments show that SAKAN converges rapidly within 15 rounds, with balanced precision, recall, and F1 value for each fault category. Its sparse self-attention mechanism effectively captures key features, making it suitable for low-resource and high real-time industrial scenarios, providing reliable decision support for intelligent operation and maintenance of rail transit.

[0140] KANs activation function visualization and SHAP analysis explicitly construct a transparent association between "key features and fault categories", providing a methodological framework for the interpretability verification and optimization of complex system diagnosis models, which can be extended to the field of high-dimensional time series signal health monitoring.

Claims

1. A method for diagnosing faults in a lightweight train bogie transmission system, characterized in that: include: S1. Collect sensor data through signal acquisition equipment and divide samples into sliding windows with set overlap rates for training, validation, and testing. S2, folding the original one-dimensional data collected in S1 into two-dimensional data along the time series dimension and inputting it into a sparse self-attention network to extract attention features that fuse the attention information; S3, inputting the attention features extracted in S2 into the KANs classifier for classification to enhance the intrinsic interpretability of the model; S4. quantify the feature contribution of the enhanced attention features in S3 using the SHAP method to improve the external interpretability of the model; S5. Based on the feature contribution quantified in S4, the visual diagnosis results and interpretability analysis are finally output.

2. The lightweight train bogie transmission system fault diagnosis method according to claim 1, characterized in that: The sparse self-attention network in S2 is to perform sparse processing on the attention mechanism, wherein the attention mechanism calculation step includes: The model builds the weight matrix W of query (Q), key (K), and value (V) Q ,W K ,W V , whose dimensions are (d model ×d model ), input data matrix (dimension (L seq ×d model )) are multiplied by each weight matrix to generate the corresponding Q, K, and V matrices. The number of floating-point operations in this step is 3·L seq ·d model 2 , the parameter is Calculate the dot product of each row of the query matrix and the transposed key matrix, and obtain the attention weight by softmax normalization. The number of floating-point operations in this step is L seq 2 ·d model ; Multiply the attention weights by the value matrix to generate the final output. The number of floating-point operations in this step is L seq 2 ·d model ; The total number of floating-point operations for the attention mechanism is: FLOPs (Attention) =FLOPs(Q,K,V)+FLOPs(softmax(Q,K))+FLOPs(Attention(Q,K,V)) =FLOPs(3·L seq ·d model 2 +2·L seq 2 ·d model ); The total number of parameters of the attention mechanism ignoring the influence of the bias term on the parameter amount is: The sparse processing of the attention mechanism includes: The attention score is obtained by performing the dot product operation between the query matrix and the key matrix, and the k elements with the highest scores are selected for subsequent calculations, that is, the number of floating-point operations required is reduced from (L seq 2 ·d model ) down to (L seq ·k·d model ).

3. The lightweight train bogie transmission system fault diagnosis method according to claim 2, characterized in that: In S2, the original one-dimensional data is folded into two-dimensional data along the time series dimension by converting the dimension of the input data from (batch_size, N) to (batch_size, L seq ,d model ), where N = L seq ×d model That is, assuming that the dimension of the original input data is (batch_size, 512), it can be folded into (batch_size, 32, 16), which means that 32 sequences are contained in each batch sample, and each sequence has 16 feature dimensions.

4. The lightweight train bogie transmission system fault diagnosis method according to claim 3, characterized in that: In the S3, KANs parameterizes each weight by parameterizing a one-dimensional function using a B-spline function, thereby achieving a nonlinear expression of the weight. Specifically: Input vector x=(x1,x2,···,x n ), the function f can be expressed as: Among them, Φ q is the outer function, φ q,p are inner functions, which are all learnable parameterized one-dimensional functions; the inner function φ q,p It is realized by B-spline function, which is a piecewise polynomial function. It is a polynomial in each grid interval and remains smooth and continuous in the entire domain. q,p It is parameterized as a B-spline function with trainable coefficients; φ(x)=w b b(x)+w s spline(x), Among them, w b ,w s They are the trainable coefficients of the basic activation function b(x) and the spline function spline(x). The basic activation function b(x) is usually chosen as the silu function: b(x)=shadow(x)=x / (1+e -x ), The spline function spline(x) is defined as a linear combination of B-spline basis functions: Among them, c i is a trainable coefficient, B i (x) is the B-spline basis function defined recursively, the 0th order B-spline basis function The mathematical expression of is defined as follows: Here, x is the grid interval [t i ,t i+1 ] node, The value is 1 in the node interval and 0 elsewhere; When the order is greater than 1, the basis function is defined as: Among them, w left is the left weight coefficient, w right is the right weight coefficient. For any point x, the weight coefficient represents the position ratio of x in the interval, w left The larger it is, the closer x is to node t. i+p , the basis function on the left has a large weight, w right The larger it is, the closer x is to node t. i+p+1 , the basis function on the right contributes more to the B-spline function. By adjusting these weights, the continuity and smoothness of the basis function at the nodes are ensured. At the same time, the curvature of the spline curve in a specific interval can be controlled, thereby better fitting the data; Outer function Φ q It is also implemented by B-spline, which takes the output of the inner function as input and performs further transformation, so each Φ q It is also a learnable one-dimensional function.

5. The lightweight train bogie transmission system fault diagnosis method according to claim 4, characterized in that: The SHAP method in S4 determines the SHAP value of each feature by evaluating the average marginal contribution of each feature in all possible feature subsets. Specifically: For a specific sample x and model f, the prediction result f(x) can be formulated as: Among them, φ0 is the baseline value, which reflects the average prediction made by the model in the absence of any feature information (when all feature values ​​are 0), φ j (x) represents the SHAP value of the jth feature, indicating the individual contribution of this feature to the model prediction f(x); For each feature x in sample x j , whose SHAP value φ j (x) is defined as: Among them, S means that feature x is not included j is a feature subset of , |S| is the number of features in subset S, n is the total number of features in sample x, Indicates that subset S and feature x are included in the model prediction j The result when F(x S ) is the predicted value of the model when only subset S is included. For each model that does not contain x j The feature subset S, Yes Join x j The change in the model prediction after denoting x j The marginal contribution to the prediction result, is the weight that weights the marginal contribution to reflect x j The average impact across all possible subset combinations is obtained by summing the weighted marginal contributions of all subsets to give x j The SHAP value of , that is, its average marginal contribution to the model prediction.

6. A lightweight train bogie transmission system fault diagnosis system, characterized in that: include: A data preprocessing module is responsible for data acquisition and sample preparation. It uses signal acquisition equipment to collect sensor data and then divides the samples into sliding windows with set overlap rates for training, validation, and testing. A sparse self-attention module is used to compress the one-dimensional time series signal into two-dimensional data through dimension folding and input it into the sparse self-attention network. The sparse self-attention module extracts deep features of the fused attention information while reducing model complexity and parameter quantity; The explainable diagnosis module is used to output diagnostic results and provide explainable analysis. First, the extracted attention features are input into the KANs classifier for classification. KAN enhances the intrinsic interpretability of the model by dynamically adjusting the shape of the activation function. Then, the SHAP method is used to quantify the feature contribution and visualize the relationship between the attention features and the fault category, thereby improving the external interpretability of the model.

Images