High-speed train suspension fault diagnosis method and system based on time-frequency sparse attention

Abstract

The present application relates to the field of rail transit vehicle state monitoring and intelligent fault diagnosis, and particularly relates to a high-speed train suspension fault diagnosis method and system based on time-frequency sparse attention. The method obtains vibration signals of a high-speed train suspension system under normal working conditions and multiple fault working conditions, performs noise enhancement, standardization, time sequence division and sliding window segmentation on the vibration signals to obtain sample sequences; the sample sequences are input into a time-frequency mapping feature extraction module, a time-frequency joint feature is constructed using a Fourier basis mapping, and a fixed length feature sequence is formed through a convolution structure; the fixed length feature sequence is input into a sparse attention fusion module, global coarse-grained, key sparse points and local neighborhood features are extracted in parallel through compressed attention, selected attention and sliding window attention, and are fused in a gating manner.

Description

Technical Field

[0001] This invention relates to the fields of rail transit vehicle condition monitoring, vibration signal processing, and intelligent fault diagnosis technology, and particularly to a method and system for diagnosing suspension faults in high-speed trains based on time-frequency sparse attention. Background Technology

[0002] The suspension system of a high-speed train, located between the wheelsets and the car body, is a crucial system for ensuring the stability, safety, and ride comfort of the train at high speeds. The suspension system absorbs shocks and vibrations from track irregularities, wheel-rail interactions, and the vehicle's movement, controlling the relative motion between the car body, bogies, and wheelsets. If the suspension system experiences performance degradation or component failure, it can lead to increased vehicle vibration, decreased ride comfort, and in severe cases, compromise train operational safety.

[0003] During the long-term service of high-speed trains, the suspension system faces the influence of factors such as large-scale road networks, complex geographical and climatic conditions, long-distance continuous high-speed operation, wheel-rail excitation differences, and different speed levels of operation. Its vibration response exhibits characteristics of multiple operating conditions, multiple modes, strong nonlinearity, and significant noise interference. Therefore, how to accurately identify the health status of the suspension system from complex vibration signals is an important technical problem in the intelligent operation and maintenance and proactive safety assurance of high-speed trains.

[0004] Existing suspension system fault monitoring and diagnosis technologies can be broadly categorized into model-based methods and data-driven methods. Model-based methods establish a vehicle system dynamics model and analyze the residuals or state deviations between the model output and actual monitoring signals to achieve fault diagnosis, offering the advantage of strong physical interpretability. However, these methods place high demands on the accuracy of vehicle dynamics modeling, model parameter identification, computational efficiency, and hardware conditions, making them challenging to apply in complex service environments and under multiple fault combinations.

[0005] Data-driven approaches typically extract statistical, frequency domain, time-frequency, or deep features directly from sensor vibration signals and utilize machine learning or deep learning models for state recognition. Traditional machine learning methods rely on manual feature design, making it difficult to adequately adapt to the complex responses of high-speed train suspension systems under varying speeds, loads, and high noise conditions. Deep learning methods can automatically extract high-level features through convolutional networks, recurrent networks, or Transformer structures, but existing single-structure models often struggle to simultaneously consider local time-frequency information, long-range temporal dependencies, and sparse focusing on critical fault segments.

[0006] Specifically, simple convolutional neural networks can extract local morphological features, but their ability to model long-range dependencies and global correlations is limited; recurrent neural networks can process time series, but they are prone to information decay or decreased training efficiency under long sequence conditions; while ordinary Transformers have global modeling capabilities, they involve large computational costs for long vibration sequences and are easily affected by a large number of irrelevant background segments. For the fault diagnosis task of high-speed train suspension systems, fault characteristics are often simultaneously reflected in frequency component changes, local transient responses, and cross-time step correlation structures, making it difficult to reliably extract these multi-level information using only a single model.

[0007] Therefore, there is a need in this field for a fault diagnosis method that can organically combine physically meaningful time-frequency mapping, selective sparse attention mechanism, and time series model with global modeling capability for vibration signals of high-speed train suspension systems, so as to improve the accuracy, robustness, and interpretability of suspension system fault identification under complex working conditions and noise interference. Summary of the Invention

[0008] This invention provides a method and system for diagnosing high-speed train suspension faults based on time-frequency sparse attention. It aims to solve the problem that a single model in the existing fault diagnosis of high-speed train suspension systems is difficult to extract time-frequency features, key sparse segments, local neighborhood associations and global temporal dependencies at the same time, thereby improving the diagnostic accuracy and stability under complex working conditions and noise interference.

[0009] To achieve the objectives of this invention, the technical solution adopted is: a fault diagnosis method for high-speed train suspension systems based on time-frequency mapping and sparse attention time-series modeling, comprising: S1. Acquire vibration signals of the high-speed train suspension system under normal operating conditions and various fault conditions, and set corresponding category labels for different operating conditions; S2. Perform data preprocessing on the vibration signal to obtain a sample sequence for training or inference of the fault diagnosis model; S3. Input the sample sequence into the time-frequency mapping feature extraction module, construct time-frequency joint features using Fourier basis mapping, and obtain a fixed-length feature sequence through convolution dimensionality reduction; S4. Input the fixed-length feature sequence into the sparse attention fusion module, extract temporal correlation features of different scales through compressed attention, selective attention and sliding window attention respectively, and obtain the fused feature sequence through gating fusion; S5. Input the fused feature sequence into the time series modeling module, and perform time series dependency and global correlation modeling through patch embedding, global token and exogenous token to obtain the fault representation vector; S6. Input the fault characterization vector into the classifier and output the fault category of the high-speed train suspension system.

[0010] As an optimized solution of the present invention, in step S1, the multiple fault conditions include at least one or more of the following: air spring deflating, primary vertical damper malfunction, secondary lateral damper malfunction, anti-roll torsion bar malfunction, and anti-hunting damper malfunction. Preferably, the data collected under normal operating conditions is denoted as Label 0, the air spring deflating condition as Label 1, the primary vertical damper malfunction as Label 2, the secondary lateral damper malfunction as Label 3, the anti-roll torsion bar malfunction as Label 4, and the anti-hunting damper malfunction as Label 5.

[0011] As an optimized solution of the present invention, in step S2, the data preprocessing includes adding a preset level of Gaussian noise to the vibration signal to simulate random interference in the actual operating environment; using Z-score standardization to subtract the mean and divide by the standard deviation to reduce the influence of differences in the dimensions and amplitudes of data from different channels and different speed levels; dividing the original data of each speed level into training set, validation set and test set in chronological order to avoid data leakage in time series modeling; and performing sliding window segmentation on each segment of data to obtain a fixed-length sample sequence.

[0012] As an optimized solution of the present invention, in step S3, the time-frequency mapping feature extraction module is preferably a Fourier basis mapping module. This module performs a real-valued fast Fourier transform on the input sample sequence to obtain the real and imaginary coefficients of the frequency components, and then uses pre-computed cosine basis functions and sine basis functions for time-domain expansion to form a time-frequency joint feature map. Subsequently, the time-frequency joint feature map is subjected to dimensionality reduction through multi-layer convolution, batch normalization, activation and pooling structures, and the feature sequence length is unified to a preset length through adaptive average pooling to obtain a fixed-length feature sequence.

[0013] As an optimized solution of the present invention, in step S4, the sparse attention fusion module is preferably a native sparse attention module. This module generates a query matrix Q, a key matrix K, and a value matrix V through linear projection, and executes compressed attention, selective attention, and sliding window attention in parallel. Specifically, compressed attention performs intra-block average compression of keys and values ​​from adjacent time steps to obtain a global coarse-grained representation; selective attention retains only the top K most relevant key-value pairs for each query to focus on critical sparse fragments related to faults; and sliding window attention restricts attention calculation to a local neighborhood to enhance local temporal continuity and neighborhood interaction.

[0014] As an optimization of the present invention, the sparse attention fusion module further includes a gated linear layer. The gated linear layer learns the fusion weights of three outputs—compressed attention, selection attention, and sliding window attention—based on the input features at each time step, and performs weighted fusion of the three attention outputs, thereby achieving an adaptive balance between global coarse-grained information, key sparse information, and local neighborhood information.

[0015] As an optimized solution of the present invention, in step S5, the time-series modeling module is preferably the TimeXer module. This module divides the fused feature sequence into preset patch lengths, flattens the features within each patch and maps them to patch embeddings, and adds a learnable global token to the front of the patch token sequence; simultaneously, it compresses the entire fused feature sequence into exogenous tokens to provide supplementary information for the global context. The patch tokens, global tokens, and exogenous tokens are jointly input into the Transformer encoder, which models the complex dependencies between feature channels and time segments through alternating self-attention and cross-attention.

[0016] To achieve the objectives of this invention, the technical solution also includes a fault diagnosis system for a high-speed train suspension system based on time-frequency mapping and sparse attention temporal modeling, comprising a data acquisition module, a data preprocessing module, a time-frequency mapping feature extraction module, a sparse attention fusion module, a temporal modeling module, and a classification output module.

[0017] The present invention has the following beneficial effects: 1) By explicitly constructing time-frequency joint features through Fourier basis mapping, the feature extraction at the front end of the model has a clear time-frequency physical meaning, and can utilize the frequency component change information of the vibration signal more directly than randomly initializing the convolution kernel; 2) By using parallel modeling of compressed attention, selective attention, and sliding window attention, the expression of fault-related information can be enhanced from three perspectives: global coarse-grained, key sparse points, and local neighborhood, respectively, reducing the interference of a large number of irrelevant background time steps on the diagnostic results; 3) By adaptively allocating the contributions of the three attention branches through a gating fusion mechanism, the model can automatically adjust the attention scale for different fault types and different speed level response characteristics; 4) By introducing patch tokens, global tokens, and exogenous tokens through the TimeXer time series modeling module, the interaction between local segment features, global state representation, and contextual information is realized, thereby improving the ability to model complex time series dependencies. 5) Through a hierarchical architecture of front-end time-frequency feature extraction, mid-end multi-scale sparse screening, and back-end global time series modeling, the accuracy, robustness, and convergence efficiency of fault diagnosis of high-speed train suspension system under noise interference and variable working conditions are improved. Attached Figure Description

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1 This is a technical flowchart of the high-speed train suspension system fault diagnosis method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the FBM-NSA-TimeXer hybrid neural network structure provided in an embodiment of the present invention; Figure 3 This is a diagram showing the confusion matrix results of the proposed method on the test set in this embodiment of the invention; Figure 4 This is a t-SNE visualization result of the method proposed in this embodiment on the test set.

[0020] Figure 5 This is a graph showing the trend of accuracy changes of each model in the ablation experiment training set and validation set in the embodiments of the present invention; Figure 6 This is a comparison chart of the accuracy of the method proposed in the embodiments of the present invention with different comparison models. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or improvements made to related steps, parameters, and implementation forms within the spirit and principles of the present invention should all fall within the scope of protection of the present invention.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or improvements made to related steps, modules, network structures, parameter values, and experimental conditions within the spirit and principles of the present invention should all fall within the scope of protection of the present invention.

[0023] This embodiment uses vibration signals obtained from a high-speed train suspension system fault simulation test as an example to illustrate a fault diagnosis method for high-speed train suspension systems based on time-frequency mapping and sparse attention time-series modeling. Figure 1 As shown, the method of the present invention mainly includes steps such as data acquisition and label setting, data preprocessing, time-frequency mapping feature extraction, sparse attention fusion, TimeXer time series modeling, classification output, and training evaluation.

[0024] 1. Data Acquisition and Tag Setting In this embodiment, a high-speed train suspension system fault simulation test is conducted on a rolling vibration test bench for locomotives and rolling stock, and vibration signals under different operating conditions are collected. Based on vehicle dynamics calculations and fault response characteristics, a set of sensors capable of characterizing the fault state of the high-speed train suspension system is selected, and the vibration signals collected by this set of sensors are used as raw data.

[0025] Preferably, six types of operating condition labels are set: data collected under normal operating conditions is denoted as Label 0; data collected under the condition of one air spring deflating is denoted as Label 1; data collected under the condition of removing eight primary vertical dampers is denoted as Label 2; data collected under the condition of removing one secondary lateral damper is denoted as Label 3; data collected under the condition of removing one anti-roll torsion bar is denoted as Label 4; and data collected under the condition of removing one anti-hunting damper is denoted as Label 5. These operating condition labels are used to construct the category labels required for supervised learning.

[0026] In this embodiment, the sensor sampling frequency is 1000Hz, and the sampling duration for each speed level is 60s. Since removing the anti-hunting damper may cause bogie instability or even derailment, a differentiated speed sampling strategy is adopted: the first five operating conditions each contain five speed levels with 300,000 data points; the sixth operating condition contains three speed levels with 180,000 data points. The above experimental setup is used to illustrate one embodiment of the present invention, and the present invention is not limited to the specific number of speed levels, sampling frequency, or sampling duration described above.

[0027] 2. Data Preprocessing and Sample Construction To simulate the random interference that may be present in sensor data during actual high-speed train operation and to improve the robustness of the model under noisy conditions, this embodiment adds a preset level of Gaussian noise to the original sample data. Preferably, the Gaussian noise level is 1%. In other embodiments, the noise type and noise level can also be adjusted according to the actual track environment, sensor noise characteristics, or data augmentation requirements.

[0028] To reduce the impact of outliers and amplitude differences at different velocity levels on model training, this embodiment employs Z-score standardization. Let the original sample sequence be x, with a mean of μ and a standard deviation of σ. The standardized sample can then be expressed as: x′=(x-μ) / (σ+ε), where ε is a constant term to prevent the denominator from being zero. This process adjusts the data distribution to have a mean close to 0 and a standard deviation close to 1.

[0029] To avoid data leakage issues common in time series modeling, this embodiment first divides the raw data at each velocity level into training, validation, and test sets sequentially according to time. Then, it performs sliding window segmentation on each data segment and finally shuffles the data to form a sample set. This approach avoids highly similar data segments within the same continuous time period from appearing simultaneously in the training and test sets, thereby improving the reliability of the model evaluation results.

[0030] Preferably, after the sliding window segmentation, 2996 samples are obtained, each sample having a length of 2048, and the ratio of training set, validation set, and test set is 0.7:0.15:0.15. The input size of a single sample can be represented as [Batch, Seq_len=2048, 1], where Batch represents the batch size, Seq_len represents the time series length, and 1 represents a single-channel input. The above sample length and segmentation ratio are preferred values ​​in this embodiment, and the present invention is not limited thereto.

[0031] like Figure 2 As shown, the fault diagnosis model in this embodiment preferably adopts an FBM-NSA-TimeXer hybrid neural network structure. After normalization, the original vibration signal is first input into the FBM module, which uses Fourier basis mapping to construct a joint time-frequency representation and extracts local time-frequency features through a multi-layer convolutional structure. Subsequently, the feature sequence enters the NSA module, which performs compressed attention, selective attention, and sliding window attention in parallel, and adaptively fuses the three types of attention outputs through a gating aggregation mechanism. The fused features are further input into the TimeXer module, which mines global temporal dependencies and channel correlations through structures such as Embedding, Self-Attention, Cross-Attention, Feed-forward, and Layer Norm. Finally, the classification head outputs the fault category of the high-speed train suspension system.

[0032] 3. Time-frequency feature extraction using Fourier basis mapping After the sample construction is completed, the sample sequence is input into the Fourier basis mapping feature extraction module. This module is used to explicitly construct physically meaningful time-frequency joint features at the front end of the model to enhance the model's ability to express changes in the frequency components and temporal structure of vibration signals.

[0033] Specifically, for the original vibration signal with input size [Batch, Seq_len=2048, 1], a real-valued Fast Fourier Transform is first performed to obtain the real and imaginary coefficients of the frequency components. Then, the real and imaginary coefficients are extended in the time domain using pre-computed cosine and sine basis functions to form a time-frequency joint feature map. Through the above processing, the model can obtain an explicit representation corresponding to the signal spectral structure before entering the deep network.

[0034] Subsequently, the time-frequency joint feature map is input into a multi-layer convolutional dimensionality reduction structure. Preferably, the convolutional dimensionality reduction structure includes three layers of convolution-batch normalization-activation-pooling units, used to extract local time-frequency patterns and progressively reduce the feature dimension. Finally, adaptive average pooling is used to unify the feature sequence length to 64, outputting a fixed-length feature sequence with a size of [Batch, Time_Steps=64, d_model=128].

[0035] Compared to directly inputting the original vibration sequence into a regular convolutional network, the Fourier basis mapping module can introduce explicit frequency basis functions in the early stages of feature extraction, making it easier for the model to capture the frequency components, energy distribution, and local vibration morphology changes caused by suspension system component failures.

[0036] 4. Sparse Attention Fusion Module A fixed-length feature sequence is input into the sparse attention fusion module. This module is used to filter and enhance the feature sequence from three dimensions: global, sparse key points, and local neighborhood, in order to improve the ability to distinguish fault-related features.

[0037] Specifically, the query matrix Q, key matrix K, and value matrix V are first generated from the input feature sequence through linear projection. Then, three attention mechanisms are executed in parallel: compressed attention, selection attention, and sliding window attention. Compressed attention averages and compresses the keys and values ​​of adjacent time steps within a block, forming a shorter, coarse-grained global representation, thereby reducing the computational burden of long sequence attention while preserving the overall trend of change.

[0038] Attention is selected based on the relevance between the query and the key. For each query, only the top K most relevant key-value pairs are retained, and attention is calculated for these pairs. In this way, the model can focus on the key segments most likely to contain fault information from a large number of time steps, reducing the interference of irrelevant background vibrations on the diagnostic results.

[0039] Sliding window attention restricts the attention calculation at each time step to its neighboring time window, emphasizing local continuous responses and short-term vibration modes. For vibration signals of high-speed train suspension systems, shocks, attenuations, oscillations, and transient changes in the local neighborhood often have important diagnostic significance. Therefore, sliding window attention can enhance the model's representation of local temporal structures.

[0040] After obtaining the three attention outputs, a gated linear layer is used to learn the fusion weights of the three outputs at each time step. The dimension of the gated fused output remains [Batch, 64, 128], and is used as the input to the subsequent TimeXer time series modeling module. Through the above structure, the sparse attention fusion module can adaptively balance the three types of information: global trend, key sparse points, and local neighborhood.

[0041] 5. TimeXer time series modeling and classification output The feature sequence fused by sparse attention is input into the TimeXer temporal modeling module. This module is used to deeply explore the temporal dependencies and global correlations between feature channels, thereby forming a high-level fault representation for fault classification.

[0042] Specifically, the fused feature sequence is first divided into segments according to a preset patch length, and the features within each patch are flattened and mapped to patch embeddings. These patch embeddings can aggregate local continuous time segments into more compact token representations, thereby reducing the modeling complexity of the subsequent Transformer encoder.

[0043] Then, a learnable global token is added to the front of the patch token sequence. This global token is used to aggregate global fault information throughout the sequence and serves as the input representation for the classifier after encoding. Simultaneously, the entire fused feature sequence is compressed into exogenous tokens to provide the model with overall contextual information.

[0044] The patch token sequence, global token, and exogenous token are input into a single-layer Transformer encoder. This encoder performs self-attention and cross-attention operations sequentially. Self-attention models the relationships between patch tokens and between global and patch tokens, while cross-attention enables information exchange between exogenous and global tokens. The encoded token representations are then transformed by a feedforward network.

[0045] Finally, the encoded global token vector is taken as the fault representation vector, and a predicted category is output through a fully connected classifier. The predicted category corresponds to normal operating conditions or different suspension system fault conditions, thereby realizing fault diagnosis of high-speed train suspension systems.

[0046] 6. Model training, ablation experiments, and comparative experiments During model training, training set samples are input into a hybrid neural network consisting of a Fourier basis mapping module, a sparse attention fusion module, and a TimeXer temporal modeling module, and parameters are optimized using the cross-entropy loss function. The validation set is used for model parameter tuning and training process monitoring, while the test set is used to ultimately evaluate the model's diagnostic performance. Evaluation metrics include Accuracy, Precision, Recall, and F1-score, used to evaluate model performance from the perspectives of overall classification correctness, category recognition completeness, and comprehensive classification performance.

[0047] In this embodiment, the complete model FBM-NSA-TimeXer achieved good diagnostic results on the test set. The test set accuracy was 0.9923, precision was 0.9932, recall was 0.9929, and F1-score was 0.9929. Furthermore, five repeated experiments were conducted on the complete model, yielding accuracy of 0.9964±0.0034, precision of 0.9968±0.0031, recall of 0.9967±0.0032, and F1-score of 0.9967±0.0032. These results demonstrate that the proposed model exhibits high classification accuracy and good stability in repeated experiments on the test set.

[0048] like Figure 3 As shown, the proposed method exhibits a clear diagonal advantage in the confusion matrix results on the test set, indicating that most samples are correctly classified. Among the six operating conditions, the normalized recognition results for Label 0, Label 1, Label 2, Label 4, and Label 5 all reach 1.00. A small number of samples in Label 3 are identified as Label 4, with a corresponding proportion of 0.04, but the overall misclassification rate is low. These results demonstrate that the proposed FBM-NSA-TimeXer model can effectively distinguish between normal operating conditions and various suspension system fault conditions, and has good applicability for multi-category fault diagnosis tasks in high-speed train suspension systems.

[0049] like Figure 4As shown, after t-SNE dimensionality reduction, the high-level fault representations extracted by the proposed method form relatively clear cluster distributions of different categories of samples in two-dimensional space, and the sample clusters of each category have good spatial separation. This result indicates that after Fourier basis mapping, sparse attention fusion, and TimeXer temporal modeling, the model can learn fault feature representations with strong discriminative ability. In other words, different suspension system operating conditions have formed distinguishable category structures in the model feature space, further validating the effectiveness of the proposed method from the perspective of feature visualization.

[0050] To verify the effectiveness of each submodule, this embodiment constructs an ablation experiment. In the ablation experiment, the complete Proposed Method is compared with models without FBM, without NSA, without TimeXer, FBM, and TimeXer to analyze the contributions of the Fourier basis mapping module, sparse attention fusion module, and TimeXer temporal modeling module to fault diagnosis accuracy, convergence speed, and generalization ability, respectively.

[0051] like Figure 5 As shown, in the ablation experiments, the complete FBM-NSA-TimeXer model was compared with models without the FBM module, without the NSA module, without the TimeXer module, and with only some modules retained. The ablation experiment results show that the accuracies of the models without FBM, without NSA, without TimeXer, FBM, and TimeXer are 0.929, 0.974, 0.975, 0.923, and 0.859, respectively, with corresponding convergence rounds of 14, 10, 21, 18, and 20. Compared with the aforementioned ablation models, the complete model is superior in both accuracy and convergence performance, indicating that the time-frequency joint representation provided by the FBM module, the multi-scale sparse attention filtering provided by the NSA module, and the global temporal modeling provided by the TimeXer module all contribute to fault feature extraction and classification.

[0052] Specifically, after removing the FBM module, the model front end lacks explicit time-frequency joint representation, and the accuracy drops to 0.929, indicating that Fourier basis mapping can provide effective frequency domain structure information for vibration signal fault diagnosis. After removing the NSA module, the model lacks adaptive filtering of global coarse-grained information, key sparse segments, and local neighborhood information, and the accuracy drops to 0.974, indicating that the sparse attention fusion mechanism helps enhance the discrimination ability of fault-related segments. After removing the TimeXer module, the accuracy drops to 0.975, indicating that the back-end time-series modeling module can further mine the global correlation in the feature sequence. When only FBM or only TimeXer is used, the model accuracy is 0.923 and 0.859, respectively, indicating that a single module is difficult to fully cover the time-frequency features, key segments, and global time-series dependencies in the vibration signal of the suspension system fault. Therefore, the hierarchical structure of front-end time-frequency mapping, mid-end sparse attention fusion, and back-end time-series modeling adopted in this invention has synergistic gains.

[0053] To evaluate the effectiveness of the proposed method compared to existing methods, this embodiment further constructs comparative experiments, comparing the proposed FBM-NSA-TimeXer hybrid neural network with models such as SVM, CNN, CNN-LSTM, CNN-Transformer, TFN, WCCN-BiLSTM, and TimeMIL. The performance of each model is evaluated by averaging multiple experiments. Evaluation metrics include accuracy, precision, recall, and F1-score.

[0054] like Figure 6As shown, this embodiment compares the proposed FBM-NSA-TimeXer model with various comparative models through repeated experiments. The average values ​​of each evaluation metric for different models are shown in Table 1. Table 1 shows that the proposed method outperforms the listed comparative methods in terms of Accuracy, Precision, Recall, and F1-score. Taking Accuracy as an example, the average accuracies of SVM, CNN, CNN-LSTM, CNN-Transformer, TFN, WCCN-BiLSTM, and TimeMIL are 0.758, 0.845, 0.911, 0.920, 0.972, 0.917, and 0.944, respectively. The proposed method improves the accuracy by about 2.7% compared to the best comparison method TFN, indicating that the proposed combination of time-frequency mapping, sparse attention fusion and TimeXer time series modeling can make fuller use of the local time-frequency features, key time segment information and global time series dependencies in the vibration signal of the high-speed train suspension system, thereby improving the accuracy and robustness of fault diagnosis under complex working conditions.

[0055] Table 1. Average values ​​of each evaluation index in the comparison model

[0056] The above experimental results are used to illustrate the diagnostic effect of the present invention in specific embodiments. It should be noted that the scope of protection of the present invention is not limited to specific experimental accuracy, training rounds, data scale, or evaluation index values. As long as the time-frequency mapping, sparse attention fusion, and temporal modeling diagnostic process described in the present invention are used, they should all fall within the scope of protection of the present invention.

[0057] 7. System Implementation Examples This embodiment also provides a fault diagnosis system for a high-speed train suspension system based on time-frequency mapping and sparse attention temporal modeling. The system includes a data acquisition module, a data preprocessing module, a time-frequency mapping feature extraction module, a sparse attention fusion module, a temporal modeling module, a classification output module, and a training and evaluation module.

[0058] The data acquisition module is used to acquire vibration signals of the high-speed train suspension system under normal operating conditions and various fault conditions, and to assign corresponding category labels to different operating conditions. The vibration signals can come from rolling vibration test benches, track operation monitoring systems, on-board sensor systems, or other devices capable of acquiring the vibration response of the suspension system.

[0059] The data preprocessing module is used to perform data augmentation, standardization, time-sequential partitioning, and sliding window segmentation on the vibration signal to obtain a sample sequence. The data augmentation may include adding Gaussian noise, random perturbation, amplitude scaling, or time shifting; the standardization may be Z-score standardization, max-min normalization, or other amplitude normalization processing.

[0060] The time-frequency mapping feature extraction module is used to perform Fourier basis mapping on the sample sequence, construct joint time-frequency features, and obtain a fixed-length feature sequence through convolution dimensionality reduction. In alternative embodiments, the time-frequency mapping can also employ short-time Fourier transform, wavelet transform, empirical mode decomposition, variational mode decomposition, or other mapping methods capable of constructing time-frequency features.

[0061] The sparse attention fusion module is used to extract temporally correlated features at different scales through compressed attention, selective attention, and sliding window attention, and then fuses them through a gating mechanism. In alternative embodiments, the sparse attention mechanism can also be replaced by local attention, multi-head attention, channel attention, temporal attention, or other attention structures that can achieve feature selection and weight allocation.

[0062] The temporal modeling module is used to model the fused feature sequence through patch embedding, global tokens, exogenous tokens, self-attention, and cross-attention to obtain a fault representation vector. In alternative embodiments, the temporal modeling module may also employ Transformer, TCN, LSTM, GRU, Informer, Autoformer, or other network structures with temporal dependency modeling capabilities.

[0063] The classification output module is used to output the fault category of the high-speed train suspension system based on the fault representation vector. The training and evaluation module is used to construct ablation experiments and comparison experiments, and output evaluation results such as accuracy, precision, recall, F1 score, loss curve, convergence speed or confusion matrix.

[0064] 8. Alternative Implementation Methods This invention is not limited to the specific embodiments described above. Regarding data sources, in addition to data collected from a rolling vibration test bench, actual track operation data, hardware-in-the-loop simulation data, or vehicle dynamics simulation data can also be used, as long as the data can reflect the vibration response of the high-speed train suspension system under different health conditions, it can be used as input data for this invention.

[0065] In terms of fault categories, in addition to air spring deflating, primary vertical damper malfunction, secondary lateral damper malfunction, anti-roll torsion bar malfunction, and anti-hunting damper malfunction, other categories related to the suspension system status may also be included, such as spring stiffness degradation, damper performance degradation, component loosening, sensor malfunction, or combined faults.

[0066] For time-frequency feature extraction, Fourier basis mapping can be replaced by short-time Fourier transform, wavelet packet transform, continuous wavelet transform, empirical mode decomposition, variational mode decomposition, Hilbert-Huang transform, or learning filter bank. As long as the original vibration signal can be converted into a feature representation containing time and frequency domain information, the same or similar invention objectives can be achieved.

[0067] For sparse attention, the specific implementation methods, block length, Top-K values, and window length of compressed attention, selective attention, and sliding window attention can all be adjusted according to the data length, computing resources, and fault feature distribution; other sparse attention or multi-scale attention structures can also be used to achieve the fusion of global, key point, and local information.

[0068] For temporal modeling, the TimeXer module can be replaced with other Transformer variants, temporal convolutional networks, recurrent neural networks, or hybrid neural networks. For the classifier, fully connected layers, Softmax classifiers, support vector machine classifiers, or other classifier head structures can be used. All of these alternatives can achieve fault diagnosis of high-speed train suspension systems without departing from the inventive concept.

Claims

1. A fault diagnosis method for high-speed train suspension systems based on time-frequency mapping and sparse attention temporal modeling, characterized in that, Includes the following steps: S1. Acquire vibration signals of the high-speed train suspension system under normal operating conditions and various fault conditions, and set corresponding category labels for different operating conditions; S2. Perform data preprocessing on the vibration signal to obtain a sample sequence for training or inference of the fault diagnosis model; S3. Input the sample sequence into the time-frequency mapping feature extraction module, construct time-frequency joint features using Fourier basis mapping, and obtain a fixed-length feature sequence through convolution dimensionality reduction; S4. Input the fixed-length feature sequence into the sparse attention fusion module, extract temporal correlation features of different scales through compressed attention, selective attention and sliding window attention respectively, and obtain the fused feature sequence through gating fusion; S5. Input the fused feature sequence into the time series modeling module, and perform time series dependency and global correlation modeling through patch embedding, global token and exogenous token to obtain the fault representation vector; S6. Input the fault characterization vector into the classifier and output the fault category of the high-speed train suspension system.

2. The method for diagnosing faults in a high-speed train suspension system according to claim 1, characterized in that: In step S1, the various fault conditions include at least one or more of the following: air spring deflating, primary vertical damper malfunction, secondary lateral damper malfunction, anti-roll torsion bar malfunction, and anti-hunting damper malfunction.

3. The method for diagnosing faults in a high-speed train suspension system according to claim 1, characterized in that: In step S2, the data preprocessing includes adding a preset level of Gaussian noise to the vibration signal, using Z-score normalization, dividing the raw data of each velocity level into training set, validation set and test set in chronological order, and performing sliding window segmentation on the divided data respectively.

4. The high-speed train suspension system fault diagnosis method according to claim 1, characterized in that: In step S3, the time-frequency mapping feature extraction module performs a real-valued fast Fourier transform on the sample sequence with input size [Batch, Seq_len, Channel] to obtain the real and imaginary coefficients of the frequency components, and uses pre-computed cosine basis functions and sine basis functions to perform time-domain extension to form a time-frequency joint feature map.

5. The high-speed train suspension system fault diagnosis method according to claim 4, characterized in that: The time-frequency mapping feature extraction module includes multi-layer convolution, batch normalization, activation and pooling structures, and uses adaptive average pooling to unify the feature sequence length to a preset length, resulting in a fixed-length feature sequence with dimensions [Batch, Time_Steps, d_model].

6. The method for diagnosing faults in a high-speed train suspension system according to claim 1, characterized in that: In step S4, the sparse attention fusion module generates a query matrix Q, a key matrix K, and a value matrix V through linear projection; the compressed attention performs intra-block average compression of the keys and values ​​of adjacent time steps to obtain a global coarse-grained representation. The selective attention retains the top K most relevant key-value pairs for each query to obtain a sparse focused representation; The sliding window attention method restricts attention calculation to a preset local neighborhood to obtain a local interaction representation.

7. The high-speed train suspension system fault diagnosis method according to claim 6, characterized in that: The sparse attention fusion module also includes a gated linear layer, which learns the weights of the three outputs—compressed attention, selection attention, and sliding window attention—at each time step, and performs weighted fusion of the three outputs to obtain a fusion feature sequence with unchanged dimensions.

8. The method for diagnosing faults in a high-speed train suspension system according to claim 1, characterized in that: In step S5, the temporal modeling module divides the fused feature sequence into preset patch lengths, flattens the features within each patch and maps them to patch embeddings, adds a learnable global token to the front of the patch token sequence, and compresses the fused feature sequence into an exogenous token. The patch token sequence, the global token, and the exogenous token are input into the Transformer encoder, which sequentially performs self-attention, cross-attention, and feedforward network transformations, and outputs the fault representation vector corresponding to the global token.

9. The method for diagnosing faults in a high-speed train suspension system according to claim 1, characterized in that: The fault diagnosis model uses the cross-entropy loss function for parameter optimization during training, and evaluates the contributions of the time-frequency mapping feature extraction module, sparse attention fusion module, and temporal modeling module to the diagnostic results through ablation experiments and comparative experiments.

10. A fault diagnosis system for high-speed train suspension systems based on time-frequency mapping and sparse attention temporal modeling, characterized in that, include: The data acquisition module is used to acquire vibration signals of the high-speed train suspension system under normal operating conditions and various fault conditions, and to set corresponding category labels for different operating conditions. The data preprocessing module is used to perform data enhancement, standardization, time-sequence partitioning, and sliding window segmentation on the vibration signal to obtain a sample sequence; The time-frequency mapping feature extraction module is used to perform Fourier basis mapping on the sample sequence, construct time-frequency joint features, and obtain a fixed-length feature sequence through convolution dimensionality reduction; The sparse attention fusion module is used to extract temporal correlation features at different scales based on compressed attention, selection attention and sliding window attention, and obtain a fused feature sequence through gating fusion; The temporal modeling module is used to model the fused feature sequence through patch embedding, global tokens, exogenous tokens, self-attention and cross-attention to obtain a fault representation vector; The classification output module is used to output the fault category of the high-speed train suspension system based on the fault characterization vector.

Images