A Fault Diagnosis Method and Equipment for Railway Bogie Bearings Based on Feature Extraction Networks

Abstract

The application discloses a railway bogie bearing fault diagnosis method and equipment based on a feature extraction network, and relates to the field of intelligent diagnosis and maintenance guarantee of urban rail trains. The method comprises the following steps: data acquisition: collecting vibration signals of the bogie axle box bearing under the same speed and load combination; data preprocessing: performing frequency spectrum adaptive decomposition on each signal based on an improved empirical wavelet transform, introducing a Fisher score and a frequency spectrum entropy joint scoring mechanism, screening out key modal signals, and then performing image mapping and reconstruction into a diagnosis sequence; model construction: introducing a topological perception attention mechanism module on the basis of a lightweight convolutional neural network to construct a fault diagnosis model; model training: training the fault diagnosis model to obtain an optimized fault diagnosis model; and diagnosis result output: using the optimized fault diagnosis model to diagnose the fault signals of the bogie axle box bearing and outputting a diagnosis result. The method can improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis and maintenance of urban rail trains, and in particular to a method and equipment for diagnosing railway bogie bearing faults based on feature extraction networks. Background Technology

[0002] As a core component of the urban rail train transmission system, bogie axle box bearings endure 15-20 ton axle loads and high-speed operation at 120 km / h for extended periods. Under the combined effects of vibration, shock, temperature rise, and metal fatigue, they are prone to progressive failure. The long-term high load, high-speed operation, and complex working conditions of the bogie make bearings susceptible to wear, fatigue, and even failure, which not only reduces the train's service life but may also compromise operational safety and even endanger personnel safety. Therefore, conducting health monitoring and fault diagnosis research on bogie axle box bearings to comprehensively understand their operating status and provide timely warnings of potential faults is a crucial measure to ensure the safe operation of rail vehicles and reduce accident risks. Traditional manual tapping inspections and periodic disassembly flaw detection have blind spots. Technicians need to indirectly assess the bearing's health status by observing multi-dimensional parameters such as vibration spectrum characteristics, temperature rise curves, and acoustic emission energy values. However, manual interpretation suffers from difficulties in multi-source data fusion, delayed response to abnormal features, and strong subjectivity in setting empirical thresholds. Summary of the Invention

[0003] The technical problem to be solved by the present invention is how to provide a railway bogie bearing fault diagnosis method and equipment based on feature extraction network that can improve the accuracy of fault diagnosis.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a railway bogie bearing fault diagnosis method based on feature extraction network, comprising the following steps:

[0005] Data acquisition: Vibration signals of bogie axle box bearings under the same speed and load combination are collected. The vibration signals include normal bearing condition signals and signals of various typical bearing fault types.

[0006] Data preprocessing: Based on the improved empirical wavelet transform, each signal segment is subjected to spectral adaptive decomposition. A joint scoring mechanism of Fisher score and spectral entropy is introduced to screen out key modal signals. Then, image mapping is performed to reconstruct the diagnostic sequence.

[0007] Model Construction: A topology-aware attention mechanism module is introduced on the basis of a lightweight convolutional neural network to construct a fault diagnosis model LWCNN-TDA-Attn;

[0008] Model training: Train the fault diagnosis model LWCNN-TDA-Attn, optimize the parameters in the fault diagnosis model, and obtain the optimized fault diagnosis model LWCNN-TDA-Attn;

[0009] Diagnostic results output: The optimized fault diagnosis model LWCNN-TDA-Attn is used to diagnose the fault signals of the bogie axle box bearings, and the diagnostic results are output.

[0010] This invention also discloses a railway bogie bearing fault diagnosis device based on a feature extraction network, comprising:

[0011] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the bogie axle box bearing fault diagnosis method.

[0012] The beneficial effects of adopting the above technical solution are as follows: The method described in this application first sends the original vibration signal into the IEWT module based on random matrix theory RMT and minimum eigenvalue spacing detection for processing, which can realize adaptive sub-band division of the vibration signal spectrum. The boundary detection method driven by RMT is more robust in noisy environments and can effectively distinguish between signal and noise components, thereby obtaining a purer and physically meaningful IMF mode.

[0013] For the IMF modes obtained from decomposition, an FS–SE screening strategy was designed. This strategy first calculates the Fisher score (FS) for each IMF to measure its ability to distinguish fault types, and then introduces spectral entropy (SE) to evaluate the concentration of its frequency domain energy distribution. Finally, the IMF modes are scored and ranked through a nonlinear combination method to select the mode components with the most diagnostic value.

[0014] Next, the reconstructed key modal signals are mapped into two-dimensional polar coordinate projection images, allowing for a more intuitive and comprehensive representation of vibration characteristics in circumferential and radial spaces. Finally, the polar coordinate images are input into a topology-aware attention masking network, TDA-Attn. This network extracts topological toroidal hole structures from the images using a persistent cohomology method, generating attention masks to dynamically highlight small but critical fault mode regions, thereby improving the accuracy and robustness of diagnosis. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0016] Figure 1 This is the main flowchart of the method described in the embodiments of the present invention;

[0017] Figure 2 This is a schematic diagram of the bogie axle box bearing data acquisition platform in the method described in this embodiment of the invention;

[0018] Figure 3 This is a waveform diagram of the bearing vibration signal in the axle box of the method described in the embodiment of the present invention;

[0019] Figure 4 This is an IEWT decomposition effect diagram of the outer ring fault in the method described in the embodiment of the present invention;

[0020] Figure 5 This is an IEWT (Extended Isolation Time) diagram of cage failure in the method described in this embodiment of the invention;

[0021] Figure 6a This is a polar coordinate projection diagram of the outer ring fault in an embodiment of the present invention;

[0022] Figure 6b This is a polar coordinate projection diagram of a holding device failure in an embodiment of the present invention;

[0023] Figure 7 This is a principle block diagram of the fault diagnosis model in the method described in the embodiments of the present invention;

[0024] Figure 8a This is an accuracy curve of the fault diagnosis model in an embodiment of the present invention;

[0025] Figure 8b This is a loss value curve of the fault diagnosis model in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0028] like Figure 1 As shown in the figure, this invention discloses a railway bogie bearing fault diagnosis method based on feature extraction networks, including the following steps:

[0029] S1, Data Acquisition: Acquire vibration signals of bogie axle box bearings under the same speed and load combination. The vibration signals include normal bearing condition signals and signals of various typical bearing fault types.

[0030] S2, Data Preprocessing: Based on the improved empirical wavelet transform, each signal segment is subjected to spectral adaptive decomposition. A joint scoring mechanism of Fisher score and spectral entropy is introduced to screen out key modal signals. Then, image mapping is performed to reconstruct the diagnostic sequence.

[0031] S3, Model Construction: Based on a lightweight convolutional neural network, a topology-aware attention mechanism module is introduced to construct the fault diagnosis model LWCNN-TDA-Attn;

[0032] S4, Model Training: Train the fault diagnosis model LWCNN-TDA-Attn, optimize the parameters in the fault diagnosis model, and obtain the optimized fault diagnosis model LWCNN-TDA-Attn;

[0033] S5, Diagnostic Result Output: The optimized fault diagnosis model LWCNN-TDA-Attn is used to diagnose the fault signals of the bogie axle box bearings, and the diagnostic results are output.

[0034] The above steps will be explained in detail below with reference to specific content:

[0035] Step S1, data acquisition, specifically includes:

[0036] (1) Data acquisition is performed using a hardware-in-the-loop simulation platform for bogie axle box bearings. This platform mainly consists of the bogie running components, axle box composite sensors, a pre-processor, a 3U embedded host, an on-board diagnostic instrument, and a fault data simulation and visualization interface, such as... Figure 2 As shown, the platform simulates the actual operating environment of urban rail vehicles and has high simulation accuracy and data acquisition capabilities.

[0037] (2) During the experiment, the axle box bearing was connected to the simulated running mechanism of the bogie test bench via a mechanical connection to simulate the actual working conditions of the vehicle under running conditions. The axle box composite sensor was installed at a symmetrical position on the outside of the bearing structure to realize the synchronous acquisition of multi-dimensional vibration information. The sensor integrates multiple measurement units such as acceleration and velocity, and has high sensitivity and wide bandwidth response capability, which can effectively capture early fault characteristics. The acquired original vibration signal is first processed by the pre-processor to perform analog signal conditioning, including filtering, amplification and anti-aliasing processing. Then the signal is converted by A / D and enters the 3U embedded host system to complete the digitization, caching and classification management of the data. At the same time, the on-board diagnostic instrument runs synchronously, and monitors the signal waveform, amplitude and spectrum information in real time through the software interface to realize the visual monitoring of the test process.

[0038] The experimental platform supports dynamic adjustment of operating parameters through a simulation interface, including control of motor speed, load status, and environmental variables. In this experiment, the motor speed was set to 3500 rpm, and the load was maintained at AW0 (no load) to ensure the consistency and comparability of data acquisition. Specific data are shown in Table 1.

[0039] Table 1: Experimental bearing data

[0040]

[0041] (3) After completing the equipment deployment and setting of operating parameters, data acquisition is carried out sequentially under different fault types and operating conditions. The acquisition system is started under stable operating conditions. The vibration signal collected by the axle box composite sensor is continuously recorded by the 3U embedded host at a sampling frequency of 12kHz. The signal is stored in segments according to the set time window (each segment is 1024 points long). The sampling adopts a non-overlapping sliding window strategy to avoid data redundancy. To ensure the integrity and traceability of the data, the data management module of the platform automatically classifies each signal according to the label information, including the fault type (normal, outer ring fault, inner ring fault, rolling element fault, cage fault) and the corresponding operating parameters (speed, load, etc.), and records the timestamp and sensor number simultaneously. All acquired data is finally stored in the local storage device of the embedded host in a unified multi-channel time series data file format, which is convenient for subsequent feature extraction and model training. The entire data acquisition process is monitored in real time through the software visualization interface of the vehicle diagnostic instrument. The operator can dynamically observe the time domain waveform, spectrum changes and system operating status to ensure that the acquisition process is stable and free from abnormal interference. One-dimensional timing fault data of bogie axle box bearings, such as Figure 3 As shown;

[0042] Step S2, Data Preprocessing Stage:

[0043] This application employs a three-stage processing method consisting of adaptive spectral decomposition, modal selection, and image mapping. First, by introducing random matrix theory (RMT) and an improved empirical wavelet transform (IEWT) based on the minimum eigenvalue gap detection method, adaptive sub-band division of the vibration signal spectrum is achieved. Then, Fisher score and spectral entropy (FS-SE) are combined to score each IMF modal component, selecting the component with the highest diagnostic value. Finally, the reconstructed signal is converted into a two-dimensional polar coordinate image to enhance the spatial representation capability of periodic structures, laying the foundation for subsequent topology-aware modeling.

[0044] This invention provides a vibration signal processing method for strong nonstationarity and multi-source interference problems. It achieves a step-by-step enhancement from signal decomposition and feature purification to structural reconstruction, significantly improving the expressive power and separability of weak fault features in vibration signals. A dataset is then constructed based on this method, specifically including: Step S1, segmenting the vibration signals collected under different fault types, with each sample containing 1024 sampling points. Five fault states are involved: normal, outer race fault, inner race fault, rolling element fault, and cage fault. 1000 samples are collected for each state, totaling 5000 samples. During signal acquisition, the motor speed is maintained at 3500 RPM, and the load state is AW0. Step S2, performing adaptive spectral decomposition on the collected time-series signals using improved empirical wavelet transform (IEWT) to effectively extract key modal signals and enhance the diagnostic feature expression capability of the signal. Step S3 involves reconstructing the selected key modal signals and mapping them to two-dimensional images through polar coordinate transformation to construct a fault image dataset oriented towards topology perception, providing input data support for subsequent fault diagnosis models based on topology-aware attention masking networks (TDA-Attn).

[0045] Among them, the improved empirical wavelet transform (IEWT) combining random matrix theory (RMT) and minimum eigenvalue spacing detection algorithm is used to identify spectral mutation points and perform sub-band adaptive division analysis on the sample data. After the key modal signals are screened, they are reconstructed and mapped to two-dimensional images through polar coordinate transformation, resulting in 5000 polar coordinate projection maps. The specific calculation process is as follows.

[0046] The RMT and minimum eigenvalue spacing detection algorithm consists of two parts: constructing the covariance matrix and calculating the spacing between adjacent eigenvalues. The steps and formulas for constructing the covariance matrix are as follows:

[0047] Step 1: Perform a short-time Fourier transform on the original bogie axle box bearing vibration signal to obtain the complex spectrum matrix:

[0048] ;

[0049] in, The original vibration signal, For Hamming window functions, This is the current center position of the sliding window. Angular frequency, For the Fourier kernel used for frequency conversion; at all times and frequency Upsampling yields the complex spectral matrix. .

[0050] Step 2: Construct local covariance using candidate boundary points. Take as the center, and take to the left and right respectively. Each frequency point constitutes a left and right sub-band, i.e., the left segment. and right segment :

[0051] ;

[0052] The sample covariance matrix of the two segments can then be expressed as:

[0053] ;

[0054] in, The sample covariance matrix for the corresponding sub-segment. The number of samples for each spectrum segment. It is the conjugate transpose;

[0055] The construction steps and calculation formula of the minimum eigenvalue gap detection algorithm are as follows:

[0056] Step 1: First, calculate the distance between adjacent feature values. Finding the eigenvalues ​​and sorting them in ascending order can be represented as follows:

[0057] ;

[0058] in, For the corresponding eigenvalues, The dimension of the covariance matrix is... Let be the rank. Next, for each segment of the feature value sequence, calculate the distance between adjacent segments, which can be specifically represented as:

[0059] ;

[0060] in, This represents the corresponding eigenvalue spacing. To eliminate the influence of amplitude scaling, the maximum eigenvalue spacing in each segment is normalized, which can be expressed as:

[0061] ;

[0062] in, This corresponds to the normalized value. The maximum interval is ultimately taken as the "mutation index".

[0063] ;

[0064] in, This represents the maximum spacing.

[0065] Step 2: Boundary candidate localization: within the effective spectrum range Within this range, a sliding window search is used to find frequency points that satisfy the following equation as spectral abrupt change boundaries:

[0066]

[0067] in, For the threshold, This is the minimum bandwidth.

[0068] Step 3: Combine the detected boundary sets Substituting the spectral segmentation formula of the Empirical Wavelet Transform (EWT), we construct a Meyer-type filter:

[0069] ;

[0070] in, This is the frequency domain response of the filter. For the current frequency, The center frequency of the current sub-band. For bandwidth coefficient, This is a smooth transition function.

[0071] After the above steps are completed, the IEWT adaptive decomposition result is as follows:

[0072] Based on the IEWT decomposition steps, the bogie axle box bearing vibration signal is subjected to IEWT adaptive decomposition using RMT and minimum eigenvalue spacing detection to obtain a set of intrinsic mode functions. Figure 4 This demonstrates the IEWT decomposition effect under outer ring failure and cage failure conditions. Figure 5 It can be seen that the modes after IEWT decomposition have good frequency band separation. Through adaptive frequency band division, the key fault signal is more prominent in specific mode components, while noise and other irrelevant components are separated.

[0073] Next, the best IMF signal is selected by combining Fisher's score and spectral entropy:

[0074] Combining the Fisher Score (FS) and Spectral Entropy (SE) into the FS-SE screening strategy can effectively select the optimal modal components. The following are the specific design and implementation steps of FS-SE:

[0075] To avoid an imbalance in weight allocation between FS and SE due to their different units, we perform a normalization process on both, as follows:

[0076] ;

[0077] ;

[0078] Then define a comprehensive score. The details are as follows:

[0079] ;

[0080] in, For the first The higher the FS score of an IMF score, the better. For the first For each IMF SE (Self-Effective Factor), the smaller the value, the better. To control the influence of spectral entropy on the scoring; when When the frequency is relatively small (concentrated distribution), When the score is close to 1, the FS score dominates. When the frequency distribution is large (dispersed), the spectral entropy has a significant impact on the score and suppresses low-quality modes.

[0081] Calculation of all IMFs after IEWT decomposition of outer ring faults and cage faults. , =1.0, sorted by score, the highest-scoring modalities are selected for subsequent analysis. Specific comprehensive score. As shown in Table 2:

[0082] Table 2: Overall score for each modality

[0083]

[0084] As shown in Table 2, IMF2 and IMF5 have higher overall scores for outer ring faults, indicating that these two modes contain more key fault information. For cage faults, IMF3 and IMF5 have higher overall scores, demonstrating their significant advantage in feature extraction from cage fault signals. Therefore, to ensure that signal reconstruction preserves key fault features to the greatest extent while suppressing irrelevant components and noise, this application selects the top two IMF components with the highest overall scores for different fault types during signal reconstruction: IMF2 and IMF5 for outer ring faults, and IMF3 and IMF5 for cage faults. This strategy ensures the integrity of feature information and improves the accuracy and robustness of subsequent diagnostic classification.

[0085] The screened key modal signals are reconstructed and mapped to a two-dimensional image using polar coordinate transformation:

[0086] After reconstructing the bogie axle box bearing FS-SE filter using polar coordinate projection Mapping data to two-dimensional polar coordinates can visually reveal the periodicity, patterns, and anomalous characteristics of signals. The basic process includes several key steps:

[0087] Step 1, Angle Mapping: Map the vibration signal sample points to angles. ,in The sample length is 1024;

[0088] Step 2, Radius Mapping: Mapping the vibration signal Mapped to radius ;

[0089] Step 3, Plot the polar graph: Plot the graph in the polar coordinate system. This allows us to obtain the polar coordinate projection of each sample data.

[0090] The reconstructed signals from outer ring faults and cage faults are converted into two-dimensional polar coordinate projection diagrams, as shown below. Figures 6a-6b As shown. Figures 6a-6b It can be seen that outer ring faults and cage faults form regular rings in the graph. This symmetry helps to quickly locate key fault features, making relatively hidden fault modes clearly visible.

[0091] Step S3, Design the fault diagnosis model:

[0092] A fault diagnosis model, LWCNN-TDA-Attn, is constructed by introducing a Topological-aware Attention (TDA-Attn) module on top of a Lightweight Convolutional Neural Network (LWCNN). The TDA-Attn module guides the model to focus on key regions with topological saliency in the polar coordinate map, enhancing its ability to perceive subtle fault features. The LWCNN-TDA-Attn fault diagnosis model aims to improve the expressive power of input features and classification performance. The model structure includes three main modules:

[0093] Multi-layer feature extraction module: The feature extraction module adopts a convolutional design optimized for small sample data, combined with adaptive batch normalization, to dynamically adjust the feature distribution, and further enhances the multi-scale expression capability through local pooling.

[0094] Topology-aware attention module: This module adjusts the weight distribution of the feature map using adaptively generated topological features. Guided by feature similarity and global correlation, it dynamically generates attention masks to enhance key features and suppress redundant features.

[0095] Classification Module: Based on the extracted global features, a fully connected layer is used to determine the fault category. The module design incorporates global average pooling and random deactivation techniques, effectively reducing the risk of overfitting and improving generalization ability.

[0096] The entire network fully integrates topological information and the characteristics of a lightweight model, focusing on solving the problem of multi-category fault diagnosis under complex operating conditions. Compared with traditional methods, this network is not only structurally simple, but also has significant advantages in utilizing global information and adaptive modeling.

[0097] Step S3 provides a fault diagnosis model for the vibration signal of the bogie axle box bearing. (See [link]) Figure 7 The present invention provides a fault diagnosis model for vibration signals of bogie axle box bearings, which can accurately identify various fault types. The specific implementation method includes the following steps:

[0098] Step S3-1: Dataset construction. Use a certain graph to construct the dataset...etc., refer to the previous summary. The 5000 samples are divided into training set, validation set and test set in a ratio of 6:2:2.

[0099] Step S3-2: The specific components and functions of the fault diagnosis model are shown in Table 3, and mainly include the following steps:

[0100] Table 3: Composition and Function of Fault Diagnosis Model

[0101]

[0102] Step S3-2-1: First, the multi-layer feature extraction module employs two sets of convolutional-pooling structures. The convolutional layers are configured with 3×3 kernels, and the number of channels changes sequentially from 3 to 16 and from 16 to 32. Combined with adaptive batch normalization and ReLU activation, this effectively and dynamically adjusts the feature distribution, enhancing the model's adaptability and generalization ability to small sample data. Max pooling layers are used to reduce feature dimensionality and enhance multi-scale representation. This design provides a rich and robust multi-scale feature foundation for the subsequent topology-aware attention module, thereby improving the overall accuracy and robustness of fault diagnosis.

[0103] Step S3-2-2: To enhance the model's ability to perceive detailed faults, a Topology Aware Attention Module (TDA-Attn) is introduced after the backbone feature extraction network. This module, based on Persistent Homology theory, systematically mines high-order topological structures in intermediate feature maps, especially information on loops and holes with significant persistence. First, the feature map of each channel is treated as a discrete grayscale image, and a subset filtering operation is performed. As the threshold increases, the generation and disappearance processes of connected components (0-homology) and loops (1-homology) are gradually identified and recorded, resulting in a set of topological barcodes. :

[0104] ;

[0105] in and The first The birth and death thresholds of each annular hole correspond to its persistence. K represents the total number of holes; then the one with the highest durability... Select one annular hole and calculate its geometric center:

[0106] ;

[0107] in, Indicates the first A set of pixels for each ring hole. Based on this, a spatial attention mask is constructed:

[0108]

[0109] in Controlling the diffusion radius, For the first The persistence of each hole is considered, and the mask is weighted and diffused in space with high-persistence centers as the cores. Finally, this attention mask is unfolded into a vector and element-wise multiplied with the attention weight matrix in the Transformer.

[0110] ;

[0111] By analyzing the topological barcode of the feature map, the module dynamically identifies topological structures with significant spatial and scale persistence, which typically correspond to key patterns in fault signals. Based on the extracted topological features, the module adaptively generates a spatial attention mask. This mask combines the similarity and global correlation between features, focusing on enhancing the weights of feature regions associated with fault modes while suppressing noise and redundant information, significantly improving the model's sensitivity and robustness to subtle fault features. This mechanism overcomes the limitation of traditional attention mechanisms that rely solely on local weight adjustments.

[0112] Step S3-2-3: The feature vectors weighted by topological attention are flattened and fed into the fully connected layer to fuse information and output the predicted probabilities of each fault category. During training, the model uses the AdaBelief optimization algorithm instead of the traditional stochastic gradient descent method. AdaBelief, while maintaining the fast convergence characteristic of the Adam optimizer, introduces the deviation between the gradient and its exponential moving average as an uncertainty measure, thereby achieving a more reasonable adaptive update of the learning rate. Its update formula is as follows:

[0113] ;

[0114] in, This represents the first-order moment estimate. It is an updated second-order estimate. For learning rate, To prevent small constants from being divided by zero, For rate reduction rate, For the parameters of the iterative model, The loss function is represented at the th... During the next iteration, the parameters are... The gradient is calculated. This optimization strategy enables rapid convergence in the early training phase and suppresses gradient fluctuations in the later stages, effectively enhancing the model's ability to identify minor faults and its robustness to noise.

[0115] Step S4: Training and comparative analysis of the fault diagnosis model.

[0116] Step S4-1: Figure 8a This indicates how the accuracy changes with the number of iterations during model training. Figure 8b The corresponding loss value variation curves are then displayed. Training results show that the proposed TDA-Attn model can rapidly improve recognition accuracy in the early iterations, and the loss function converges and stabilizes in the later stages, demonstrating good convergence and fault feature learning ability throughout the training process. These training results verify the effectiveness and stability of the proposed network structure in multi-class fault recognition tasks.

[0117] Step S4-2: To verify the effectiveness of the proposed LWCNN-TDA-Attn model in bogie axlebox bearing fault identification, two representative deep learning methods, ResNet18 and DenseNet, were selected as comparison models, and classification performance tests were conducted on each. Under a unified test set, the accuracy, recall, and F1-Score of the three models were quantitatively analyzed and compared to comprehensively evaluate the overall performance of the proposed model in terms of fault mode identification accuracy and robustness.

[0118] The test results are shown in Table 4. It can be seen that the LWCNN TDA-Attn model proposed in this invention performs excellently in the fault diagnosis task, achieving an accuracy of 99.32%, a recall of 99.34%, and an F1-Score as high as 0.993. Compared with existing comparative models, this method significantly improves the model's recognition accuracy and robustness for multi-class bearing faults while maintaining low parameter complexity, verifying the effectiveness and practical value of the proposed topology-aware attention mechanism and lightweight convolutional structure.

[0119] Table 4: Fault Classification Evaluation Indicators for Different Methods

[0120]

[0121] Step S4-3: Performance comparison analysis of module combination methods. In this step, to verify the synergistic effect of the topology-aware module and the attention mechanism in the fault diagnosis network of this invention, a control example was set up, introducing only a single module (topology-aware module or attention mechanism) for ablation experiments. The experimental results are shown in Table 5. When the topology-aware module (TDA) is introduced alone, the model can mine the topological invariants of the input signal, which improves the identification ability of spatial structural features to a certain extent, but still lacks selective attention to key information regions, with an accuracy of 97.45%. When the attention mechanism (Attn) is introduced alone, the model can dynamically weight and strengthen local features, effectively suppressing interference, but lacks structural prior support, with an accuracy of 98.67%. The joint model (TDA-Attn) proposed in this invention synergistically integrates the topology-aware module and the attention mechanism, enhancing the response ability to key regions while maintaining spatial connectivity sensitivity, ultimately achieving a model accuracy of 99.32%, significantly better than the above single module configurations. This result verifies the significant advantages of this invention in multi-angle modeling and diagnostic robustness.

[0122] Table 5 Ablation Experiment Results

[0123]

[0124] Step S5: Output the fault diagnosis results of the bogie axle box bearings.

[0125] Furthermore, the present invention also discloses a railway bogie bearing fault diagnosis device based on a feature extraction network, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the bogie axle box bearing fault diagnosis method.

[0126] In summary, this invention provides a fault diagnosis method for multiple types of faults in railway bogie axle box bearings, particularly suitable for fault identification tasks involving strong non-stationarity and multi-source interference. First, vibration signals of the bearings under the same speed and load combination are collected, covering normal conditions and various typical fault types, including outer race faults, inner race faults, rolling element faults, and cage faults, constructing a multi-condition original signal sample set. Then, based on the improved empirical wavelet transform (IEWT), each signal segment is spectrally decomposed. The decomposition strategy integrates random matrix theory and the minimum eigenvalue spacing detection method, achieving adaptive identification of spectral abrupt changes and sub-band division, thereby improving the stability and adaptability of the decomposition. On this basis, a joint scoring mechanism of Fisher score and spectral entropy is introduced to screen key modal signals and reconstruct them into diagnostic sequences. Furthermore, the one-dimensional reconstructed signal is transformed into a two-dimensional image through polar coordinate mapping, making its periodicity and topological structure more explicit, facilitating feature extraction by subsequent deep models.

[0127] In terms of model design, this invention constructs a lightweight convolutional neural network (LWCNN) as the backbone for feature extraction, employing a two-layer convolutional and pooling structure, combined with adaptive batch normalization to enhance robustness to small sample perturbations. The key lies in introducing a topology-aware attention module (TDA-Attn), which performs one-dimensional homology analysis on intermediate feature maps based on persistent homology theory, extracts the ring hole structure with maximum persistence, constructs an attention mask, and embeds it into the self-attention channel, achieving focus on key topological features and suppression of redundant information. Finally, fault classification output is completed through a fully connected layer. During model training, the AdaBelief optimization algorithm is introduced. This algorithm, while retaining the fast convergence characteristics of the Adam optimizer, effectively improves the model's generalization performance and ability to identify weak faults by introducing confidence weighting to suppress gradient fluctuations.

[0128] Experimental results show that the diagnostic method proposed in this invention can accurately extract key features with physical structural significance from fault signals while maintaining a lightweight structure. The accuracy and robustness of the identification are better than those of traditional deep network models, and it is suitable for axle box bearing condition monitoring and intelligent early warning systems in urban rail transit scenarios.

Claims

1. A railway bogie bearing fault diagnosis method based on a feature extraction network, characterized by Comprising the following steps: Data acquisition: collect the vibration signals of the axle box bearing of the bogie under the same speed and load combination, the vibration signals including the normal state signal of the bearing and the signals of various typical fault types of the bearing; Data preprocessing: based on the improved empirical wavelet transform, each signal is adaptively decomposed in frequency spectrum, a scoring mechanism combining Fisher score and spectral entropy FS-SE is introduced, and the key modal signal is screened out, and then image mapping is performed for reconstruction into a diagnostic sequence; Model construction: On the basis of lightweight convolutional neural network, topological perception attention mechanism module is introduced to construct fault diagnosis model -Attn; Model training: training a fault diagnosis model - Attn is trained, parameters in the fault diagnosis model are optimized, and an optimized fault diagnosis model is obtained - Attn; Diagnosis result output: using the optimized fault diagnosis model -Attn diagnoses the fault signal of the bogie axle box bearing and outputs the diagnosis result The method of improved empirical wavelet transform IEWT comprises the following steps: (1) the original bogie axle box bearing vibration signal is subjected to short-time Fourier transform to obtain a complex spectrum matrix: ； wherein, is the original vibration signal, is a Hamming window function, is the position of the center of the current sliding window, is the angular frequency, is a Fourier kernel for frequency conversion; at all times and frequencies an up-sampled complex spectral matrix is obtained; (2) Local covariance construction, taking the candidate boundary point as the center, and taking a frequency point to the left and right respectively to form left and right subbands, i.e. left segment and right segment : ； The sample covariance matrix of the two segments is represented as: ； wherein is the sample covariance matrix for the corresponding sub-segment, is the number of samples per segment of the spectrum, is the conjugate transpose; (3) Firstly, the interval statistics of adjacent eigenvalues are calculated, and the eigenvalues of each eigenvalues are calculated and arranged in ascending order, which is specifically represented as: ； wherein, is a corresponding eigenvalue, is a covariance matrix dimension, is a rank; Next, the distance between adjacent eigenvalue sequences of each segment is calculated, represented as: ； wherein is the corresponding eigenvalue spacing; In order to eliminate the influence of amplitude scale, the maximum eigenvalue distance in each segment is normalized, specifically represented as: ； wherein is the corresponding normalized value; Finally, the maximum distance is taken as the mutation index: ； wherein is the maximum distance; (4) Boundary candidate positioning: In the spectral effective interval, a sliding window is used to search for a frequency point satisfying the following equation as the spectral mutation boundary: f (n) - f (n - 1) ； wherein is a threshold value, is a minimum bandwidth; (5) the detected boundary set into the spectral division formula of the empirical wavelet transform (EWT) to construct a Meyer-type filter: ； wherein is the frequency domain response of the filter, is the current frequency, is the center frequency of the current subband, is the wideband coefficient, is a smoothing transition function. 2.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 1, characterized in that, The data acquisition comprises the following steps: (1) a semi-physical simulation platform of the bogie axle box bearing is constructed; (2) during the experiment, the axle box bearing is connected to the simulated running mechanism of the bogie test bed through mechanical connection to simulate the actual working condition under the vehicle running state; the axle box composite sensor is installed at the symmetric position outside the bearing structure to realize synchronous acquisition of multi-dimensional vibration information; the original vibration signal collected is first subjected to analog signal conditioning through a preprocessor, and then the signal is converted into digital after A / D conversion and enters the 3U embedded host system to complete data digitization, buffering and classification management; (3) after completing the device deployment and setting the running parameters, the data acquisition work is carried out under different fault types and working conditions. 3.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 1, characterized in that, The data preprocessing specifically comprises the following steps: The improved empirical wavelet transform IEWT introducing random matrix RMT and minimum eigenvalue distance detection method is used to realize adaptive decomposition of the vibration signal frequency spectrum, obtain the IMF modal component: Fisher score FS and spectral entropy FS-SE are combined to score each IMF modal component, and the key modal signal component with diagnostic value is screened out; the screened key modal signal component is reconstructed and mapped into a two-dimensional image through polar coordinate transformation to construct a fault image dataset facing topological structure perception. 4.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 3, characterized in that, The method of screening out the key modal signal component with diagnostic value comprises the following steps: The Fisher score FS and the spectral entropy SE form the FS-SE screening strategy to screen out the optimal modal component: In order to avoid the weight distribution imbalance between FS-SE due to different dimensions, normalization is performed between them, specifically as follows: ； ； A composite score is then defined as follows: ； wherein, FS score of the nth IMF, the larger the value, the better; FS score of the nth IMF, the larger the value, the better; SE value of the nth IMF, the smaller the value, the better; SE value of the nth IMF, the smaller the value, the better; is the control spectrum entropy on the impact of the degree of score. 5.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 3, characterized in that, The method of mapping the screened key modal signal component into a two-dimensional image through polar coordinate transformation comprises the following steps: (1) Angle mapping: mapping the vibration signal sample points to angles wherein is the sample length 1024; (2) Radius mapping: mapping the vibration signal as a radius ; (3) Draw the polar coordinate graph: draw the polar coordinate projection graph of each sample data under the polar coordinate system. , obtain the polar coordinate projection graph of each sample data. 6.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 1, characterized in that, The fault diagnosis model -Attn comprises: Multi-layer feature extraction module: a convolution design optimized for small sample data is used, combined with adaptive batch normalization to dynamically adjust the feature distribution and enhance the multi-scale expression ability through local pooling; Topological perception attention module: used to adjust the weight distribution of the feature map through adaptively generated topological features; Classification module: based on the extracted global features, global average pooling and random inactivation are fused, and a full connection layer is used to complete the discrimination of the fault categories.

7. The feature extraction network-based railway bogie bearing fault diagnosis method of claim 1, wherein, The fault diagnosis model The construction method of Attn includes the following steps: (1) The multi-layer feature extraction module adopts two groups of convolution-pooling structures. The convolution layer is configured with a 3×3 convolution kernel, and the number of channels is 3→16 and 16→32 in turn. Adaptive batch normalization and ReLU activation are used. The max-pooling layer is used to reduce the feature dimension and enhance the multi-scale expression; (2) Topology-aware attention module (TDA-Attn) is introduced after the backbone feature extraction network. First, the feature map of each channel is regarded as a discrete gray image, and a sub-level set filtering operation is performed. As the threshold value increases, the generation and disappearance process of connected components and ring holes are gradually identified and recorded, and a topology barcode set is obtained : ； wherein, with are the birth and death thresholds of the first ring holes, respectively, corresponding to their persistence , K represents the total number of ring holes; then the first ring holes with the highest persistence are selected and their geometric centers are calculated: ； wherein, represents a set of pixels of the ring hole, according to which the spatial attention mask is constructed: ； wherein control the diffusion radius, for the first the persistence of the hole of the ring, the mask is weighted and diffused in space with a high-persistence center, and finally the attention mask is unfolded into a vector, and an element-level product operation is performed with the attention weight matrix in the Transformer: ​ ； By analyzing the topological barcode of the feature map, the topological structure with significant persistence in space and scale is dynamically identified, which corresponds to the key patterns in the fault signal; based on the extracted topological features, the module adaptively generates a spatial attention mask; the mask combines the similarity and global correlation between features, enhances the feature region weight related to the fault pattern, and suppresses noise and redundant information at the same time; (3) The feature vector weighted by the topological attention is flattened and sent to the fully connected layer to fuse information and output the prediction probability of each fault class. 8.The railway bogie bearing fault diagnosis method based on feature extraction network according to claim 1, characterized in that, The fault diagnosis model The optimization method of Attn includes the following steps: In the training process, the AdaBelief optimization algorithm is used, and its update formula is as follows: ； where, denotes the first moment estimate, is the updated second moment estimate, is the learning rate, is a small constant to prevent division by zero, is the rate reduction factor, is the parameter of the iterative model, denotes the gradient of the loss function at the th iteration with respect to the parameter . 9.A railway bogie bearing fault diagnosis device based on a feature extraction network, characterized in that: It includes: At least one processor; And a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the railway bogie bearing fault diagnosis method based on the feature extraction network as claimed in any one of claims 1 to 8.

Images