A rail train gearbox gear fault diagnosis method based on virtual-real feature fusion
By using a virtual-real feature fusion method, combining measured and simulated data, optimizing convolution kernel weights, suppressing noise, and enhancing weak fault features, the problem of strong noise and weak features in the fault diagnosis of railway train gearboxes is solved, and high-precision fault diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-02
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196527A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transmission gearbox fault diagnosis, and specifically to a method for diagnosing gear faults in railway train gearboxes based on the fusion of virtual and real features. Background Technology
[0002] As rail trains develop towards higher speeds and greater intelligence, their operational safety faces even greater challenges. As the core component of the mechanical transmission system of rail trains, gearboxes are prone to faults such as cracks, broken teeth, wear, and missing or broken teeth. These faults can lead to serious safety accidents. Therefore, research on the accurate diagnosis of rail train gearbox faults has significant engineering value and practical implications.
[0003] Despite significant advancements in deep learning-based methods for gearbox inspection, gear detection in rail transit systems remains constrained by the complex operating environment. Internal and external disturbances, such as the operation of other train components and wheel-rail excitation, can mask fault signals. Furthermore, the long propagation paths of the acquired signals lead to signal attenuation, further weakening fault characteristics. The noise-dominated, weak signal characteristics make it difficult for existing diagnostic models to extract effective information, thus impacting diagnostic accuracy. Current diagnostic methods rely on periodically collecting fault operation signals using multiple sensors to ensure data integrity. However, the heterogeneity in distribution and scale of data from different sources can introduce artifacts or weaken high-frequency details, thus failing to deliver the expected advantages in downstream diagnostic tasks and limiting practical applications.
[0004] Existing diagnostic methods suffer from poor applicability and insufficient diagnostic accuracy when faced with complex operating environments and background noise conditions, such as those encountered by rail trains. Achieving high-precision fault diagnosis under conditions of "high noise and weak characteristic features" remains a critical problem that urgently needs to be addressed. Summary of the Invention
[0005] This invention aims to address the shortcomings of deep learning in gearbox fault diagnosis, particularly the difficulty in guaranteeing the diagnostic accuracy of the model when the acquired signal is characterized by "strong noise and weak features".
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] This invention proposes a method for diagnosing gear faults in rail train gearboxes based on the fusion of virtual and real features, characterized by the following steps:
[0008] Step S1: Collect measured samples of vibration acceleration signals from the train bogie gearbox under various gear operating conditions. Extract and label the measured samples to obtain the original measured sample set. ;
[0009] Step S2: Construct a gearbox dynamic model DM based on the actual train bogie gearbox. Simulate the gearbox using the DM to obtain vibration acceleration signal simulation samples of the gearbox under various gear operating states. Extract and label the simulation samples to obtain the original simulation sample set. ;
[0010] Step S3: For the original measured sample set Original simulation sample set Signal preprocessing is performed to obtain the processed measured sample set. and the processed simulation sample set ;
[0011] Step S4: Measure the sample set Simulation sample set The virtual-real feature fusion network model FN is input to perform virtual-real feature fusion. The virtual-real feature fusion network model FN includes a weighted convolutional unit C, a feature extraction module ME, and a virtual-real feature fusion module PF.
[0012] The virtual-real feature fusion operation process includes:
[0013] Step S41: Prepare the measured sample set Simulation sample set The input weighted convolutional unit C is used to perform a convolution operation between the input sample set and the weights of the convolutional kernel of the weighted convolutional unit C, thereby weighting the sample set to obtain a weighted measured sample set. and weighted simulation sample set .
[0014] The kernel weights of the weighted convolutional unit C are:
[0015]
[0016]
[0017] In the formula, These are the scaled and translated discrete time axis coordinates. t is the initial offset of the discrete time axis coordinates, a is the width scaling factor, and b is the position translation amount, which is updated as the model is trained. For scaling control factors, a ≠ ; The amplitude of the sample signal; The convolution kernel weight stationarity coefficient, , The characteristic frequency of gear faults; These are the kernel weights.
[0018] The weighted convolutional unit of this invention optimizes the weight of the convolutional kernel by introducing the gear fault characteristic frequency of the gearbox, which enables the convolutional layer to pay more attention to the gear fault characteristic frequency range of the sample set. The weighted signal is more distinguishable in different categories, and high-frequency noise is significantly suppressed, thereby improving the accuracy of fault diagnosis.
[0019] Step S42: Process the weighted measured sample set using the feature extraction module ME. Weighted simulation sample set Feature extraction was performed separately to obtain the measured features. Simulation features .
[0020] Preferably, the feature extraction module ME includes a semi-soft thresholding and noise reduction feature extraction submodule NR and a multi-scale feature extraction submodule PE, which extracts weighted simulation sample sets. The simulation features are extracted by the PE submodule of the multi-scale feature extraction module. The weighted measured sample set Simultaneously, the semi-soft threshold denoising module NR and the multi-scale feature extraction module submodule PE are input to extract the measured features after denoising. and multi-scale measured features The measured features are obtained by adding the results after assigning path weights. ;
[0021] The method of this invention obtains the measured features after noise reduction through semi-soft thresholding. This effectively reduces noise interference and improves the accuracy of feature extraction.
[0022] The semi-soft threshold noise reduction process is as follows:
[0023]
[0024]
[0025] In the formula, It is a two-channel attention mechanism; For adaptive average pooling; For feature extraction operations; This is a binary masking operation; The positive values are taken for the weighted measured samples; This represents the original weighted measured sample signal; The learnable compensation coefficient.
[0026] Step S43: Transfer the measured features Simulation features Input the virtual-real feature fusion module PF to obtain the fused features. ;
[0027] Optionally, virtual-real feature fusion includes reconstructing simulation features. The reconstructed simulation features Compared with measured features To integrate.
[0028] The virtual-real feature fusion operation is as follows:
[0029]
[0030]
[0031]
[0032] In the formula, For the first Class-weighted simulation sample set; The total number of samples corresponding to class c; For simulation features The class prototype vector; This is a temperature coefficient, adjusted based on the magnitude of the difference between the fused virtual and real samples; the greater the difference, the better. The larger the value; For the number of running status categories, Let i be the simulation feature of the i-th sample. To reconstruct simulation features; For measured features, This is the fusion coefficient.
[0033] Reconstructing simulation features using class prototype vectors Compared with measured features The fusion process enhances the detection of subtle faults in the measured features.
[0034] Step S5: Merge the features Input the classification module to optimize and train the diagnostic model;
[0035] Optionally, the fusion features The input classification module is based on the cross-entropy loss function. The diagnostic model is optimized and trained using the cross-entropy loss function. for:
[0036]
[0037] In the formula, Number of fault categories; For the first Similar to real-world labels; The model predicts that the sample belongs to the first... The probability of a class.
[0038] Step S6: Repeat steps S4 and S5 to train the diagnostic model. When the number of training rounds E is reached, save the trained diagnostic model.
[0039] S7: Input the dataset of the bogie gearbox of the train under test into the trained diagnostic model for processing to obtain the fault diagnosis results.
[0040] Compared with existing technologies, the advantages of this invention are as follows: The rail train gearbox gear fault diagnosis method based on virtual-real feature fusion constructs a weighted operation that incorporates physical information. By introducing gearbox gear fault feature frequencies to optimize the weights of the convolutional kernel, the convolutional layer can give higher attention to the gear fault feature frequency range of the sample set. The weighted signal is more distinguishable between different categories, and high-frequency noise is significantly suppressed, thereby improving the accuracy of fault diagnosis. Semi-soft thresholding noise reduction processing is embedded in the feature extraction module to effectively reduce noise interference. The fusion of measured features and simulation features enhances the detection of weak faults in the measured features. Therefore, the rail train gearbox gear fault diagnosis method based on virtual-real feature fusion can achieve high-precision fault diagnosis under conditions of "strong noise and weak features". Attached Figure Description
[0041] Figure 1 This is a flowchart of the method of the present invention;
[0042] Figure 2 This is a flowchart of an embodiment of the method of the present invention;
[0043] Figure 3 This is a comparison diagram after feature fusion of the method of the present invention;
[0044] Figure 4 This is a comparison chart of the diagnostic accuracy of the method of the present invention under different noise levels; Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. It should be understood that the illustrative embodiments and descriptions of the present invention are only for explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined in any way as long as they do not conflict with each other.
[0046] See Figure 1 This invention proposes a method for diagnosing gear faults in rail train gearboxes based on the fusion of virtual and real features, characterized by the following steps:
[0047] Step S1: Collect measured samples of vibration acceleration signals from the train bogie gearbox under various gear operating conditions. Extract and label the measured samples to obtain the original measured sample set. ;
[0048] Step S2: Construct a gearbox dynamic model DM based on the actual train bogie gearbox. Simulate the gearbox using the DM to obtain vibration acceleration signal simulation samples of the gearbox under various gear operating states. Extract and label the simulation samples to obtain the original simulation sample set. ;
[0049] Optionally, the operating states of the train bogie gearbox gears include five types: normal, cracked, broken teeth, worn and broken teeth, and missing teeth. The construction of the gearbox dynamics model (DM) includes 3D model establishment, model excitation, and fault modeling.
[0050] Step S3: For the original measured sample set Original simulation sample set Signal preprocessing is performed to obtain the processed measured sample set. and the processed simulation sample set ;
[0051] Optionally, the signal preprocessing includes outlier processing, zero-mean processing, fast Fourier transform, and normalization.
[0052] Step S4: Measure the sample set Simulation sample set The virtual-real feature fusion network model FN is input to perform virtual-real feature fusion. The virtual-real feature fusion network model FN includes a weighted convolutional unit C, a feature extraction module ME, and a virtual-real feature fusion module PF.
[0053] The virtual-real feature fusion network model process includes:
[0054] Step S41: Prepare the measured sample set Simulation sample set The input weighted convolutional unit C is used to perform a convolution operation between the input sample set and the weights of the convolutional kernel of the weighted convolutional unit C, thereby weighting the sample set to obtain a weighted measured sample set. and weighted simulation sample set .
[0055] The kernel weights of the weighted convolutional unit C are:
[0056]
[0057]
[0058] In the formula, These are the scaled and translated discrete time axis coordinates. t is the initial offset of the discrete time axis coordinates, a is the width scaling factor, and b is the position translation amount, which is updated as the model is trained. For scaling control factors, a ≠ ; The amplitude of the sample signal; The weights of the convolution kernel are stationary coefficients, selected based on the adaptability to the operating environment. , The characteristic frequency of gear faults; These are the kernel weights.
[0059] The weighted convolutional unit of this invention optimizes the weight of the convolutional kernel by introducing the gear fault characteristic frequency of the gearbox, which enables the convolutional layer to pay more attention to the gear fault characteristic frequency range of the sample set. The weighted signal is more distinguishable in different categories, and high-frequency noise is significantly suppressed, thereby improving the accuracy of fault diagnosis.
[0060] Step S42: Process the weighted measured sample set using the feature extraction module ME. Weighted simulation sample set Feature extraction was performed separately to obtain the measured features. Simulation features .
[0061] Preferably, the feature extraction module ME includes a semi-soft thresholding and noise reduction feature extraction submodule NR and a multi-scale feature extraction submodule PE, which extracts weighted simulation sample sets. The simulation features are extracted by the PE submodule of the multi-scale feature extraction module. The weighted measured sample set Simultaneously, the semi-soft threshold denoising module NR and the multi-scale feature extraction module submodule PE are input to extract the measured features after denoising. and multi-scale measured features The measured features are obtained by adding the results after assigning path weights. ;
[0062] ,in The path weight is selected adaptively based on the operating conditions.
[0063] The measured features after noise reduction were obtained through semi-soft thresholding. This effectively reduces noise interference and improves the accuracy of feature extraction.
[0064] The semi-soft threshold noise reduction process is as follows:
[0065]
[0066]
[0067] In the formula, It is a two-channel attention mechanism; For adaptive average pooling; For feature extraction operations; This is a binary masking operation; The positive values are taken for the weighted measured samples; This represents the original weighted measured sample signal; These are learnable compensation coefficients that adaptively optimize during training.
[0068] Step S43: Transfer the measured features Simulation features Input the virtual-real feature fusion module PF to obtain the fused features. ;
[0069] Optionally, virtual-real feature fusion includes reconstructing simulation features. The reconstructed simulation features Compared with measured features To integrate.
[0070] The virtual-real feature fusion network model operates as follows:
[0071]
[0072]
[0073]
[0074] In the formula, For the first Class-weighted simulation sample set; The total number of samples corresponding to class c; For simulation features The class prototype vector; This is a temperature coefficient, adjusted based on the magnitude of the difference between the fused virtual and real samples; the greater the difference, the better. The larger the value; For the number of running status categories, Let i be the simulation feature of the i-th sample. To reconstruct simulation features; For measured features, This is the fusion coefficient, which is adjusted according to the operating conditions and environment.
[0075] Reconstructing simulation features using class prototype vectors Compared with measured features The fusion process enhances the detection of subtle faults in the measured features.
[0076] Step S5: Merge the features The input classification module is based on the cross-entropy loss function. The diagnostic model is then optimized and trained.
[0077] The cross-entropy loss function for:
[0078]
[0079] In the formula, Number of fault categories; For the first Similar to real-world labels; The model predicts that the sample belongs to the first... The probability of a class.
[0080] Step S6: Repeat steps S4 and S5 to train the diagnostic model. When the number of training rounds E is reached, save the trained diagnostic model.
[0081] Preferably, the number of training rounds E is adaptively set according to the model convergence or diagnostic accuracy requirements.
[0082] S7: Input the dataset of the bogie gearbox of the train under test into the trained diagnostic model for processing to obtain the fault diagnosis results.
[0083] Working process of this invention:
[0084] The present invention provides a method for diagnosing gear faults in a rail train gearbox based on the fusion of virtual and real features, which is illustrated in the following example.
[0085] I. Data Acquisition:
[0086] The train bogie operates under the following conditions: motor speed 3000 r / h, vertical load 1200 N. Accelerometers mounted on the bogie gearbox collect vibration acceleration signals from the gearbox under different gear operating conditions. The gears are categorized into five operating states: normal, cracked, broken tooth, worn, and missing tooth. Rotation at the faulty gear location triggers periodic abnormal impacts, forming a periodic fault characteristic frequency, which in turn affects other components, generating harmonics.
[0087] II. Simulation Data Acquisition:
[0088] A 1:1 three-dimensional model of the train bogie gearbox was constructed using 3D modeling software to obtain the x_t file. The x_t file was then imported into dynamic simulation software and the model excitation settings were configured to obtain the bin file. Faults and defects in the health status of the train bogie gearbox were constructed on the generated model, and fault data of the train bogie gearbox were simulated.
[0089] The three-dimensional model structure includes input and output shafts, rolling bearings, large and small gears, and gearbox housing; the model excitation settings include constraint settings, contact settings, load settings, drive settings, sensor settings, and signal acquisition parameter settings; the fault modeling includes fault type, fault size, and fault location.
[0090] Based on the actual fault type, the same type of fault is constructed in the dynamic model, and the same sensor sampling points and sampling frequency are set to simulate various faults and obtain simulation data.
[0091] III. Data Preprocessing:
[0092] Vibration data of train bogie gearboxes under various operating conditions were sampled using a sliding window. The sliding window length and sliding step size were set to 2048, and five different noise conditions were set: 4dB, 2dB, 0dB, -2dB, and -4dB. Subsequently, a Fast Fourier Transform was performed on each sample, and then normalization was applied.
[0093] IV. Diagnostic Model Construction:
[0094] A diagnostic model was built based on the PyTorch deep learning framework. The experimental setup included a 64-bit Windows 10 operating system, an Intel Xeon Silver 4210R CPU at 2.4GHz, and an NVIDIA GeForce RTX 3090 GPU. Hyperparameters of the virtual-real feature fusion network model FN were set, including kernel size, kernel stride, number of neurons in fully connected layers, threshold compensation coefficient, and similarity coefficient. The main parameters of the network structure in this embodiment are shown in Table 1.
[0095] Table 1 Network Structure Parameter Settings
[0096]
[0097] V. Model Training:
[0098] The network model was optimized using the Adam optimizer with a learning rate of 0.01 and a batch size of 32, and trained for 300 epochs. Each training sample was input into the model for training.
[0099] VI. Fault Classification:
[0100] Input the dataset of the bogie gearbox of the train under test into the trained diagnostic model to obtain the fault diagnosis results.
[0101] This invention proposes a fault diagnosis method for railway train gearboxes based on virtual-real feature fusion. It combines the actual operating state and dynamic simulation model of the train bogie gearbox, effectively integrating simulation features into measured features through virtual-real feature fusion, thus enhancing weak fault signals in the measured features. During the fusion process, measured noise is suppressed to reduce interference. The weighted signal processing makes the signals more distinguishable across different categories, and high-frequency noise is significantly suppressed. (See appendix) Figure 3 The features fused by the method proposed in this invention exhibit better aggregation and clearer classification boundaries compared to WDCNN, ResNet, EWSNet, WGAN-GP, and HAMFD. (See appendix.) Figure 4 The diagnostic accuracy and stability of the method of the present invention are significantly higher than those of the comparative methods under different noise backgrounds. Even in high noise environment, the method proposed in the present invention still maintains the highest accuracy.
[0102] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should be covered within the protection scope of the present invention.
Claims
1. A method for diagnosing gear faults in rail train gearboxes based on the fusion of virtual and real features, characterized in that... include: Step S1: Collect measured samples of vibration acceleration signals from the train bogie gearbox under various gear operating conditions. Extract and label the measured samples to obtain the original measured sample set. ; Step S2: Construct a gearbox dynamic model DM based on the actual train bogie gearbox. Simulate the gearbox using the DM to obtain vibration acceleration signal simulation samples of the gearbox under various gear operating states. Extract and label the simulation samples to obtain the original simulation sample set. ; Step S3: For the original measured sample set Original simulation sample set Signal preprocessing is performed to obtain the processed measured sample set. and the processed simulation sample set ; Step S4: Measure the sample set Simulation sample set The virtual-real feature fusion network model FN is input to perform virtual-real feature fusion. The virtual-real feature fusion network model FN includes a weighted convolutional unit C, a feature extraction module ME, and a virtual-real feature fusion module PF. The process of virtual-real feature fusion network model includes: Step S41: Measure the sample set Simulation sample set Input a weighted convolutional unit C, perform a convolution operation between the input sample set and the weights of the convolutional kernel of unit C, and perform weighted processing on the sample set to obtain a weighted measured sample set. and weighted simulation sample set ; Step S42: Process the weighted measured sample set using the feature extraction module ME. Weighted simulation sample set Feature extraction was performed separately to obtain the measured features. Simulation features ; Step S43: Measure the features Simulation features Input the virtual-real feature fusion module PF to obtain the fused features. ; Step S5: Fuse features The input classification module is based on the cross-entropy loss function. The diagnostic model is then optimized and trained. Step S6: Repeat steps S4 and S5 to train the diagnostic model. When the number of training rounds E is reached, save the trained diagnostic model. S7: Input the dataset of the bogie gearbox of the train under test into the trained diagnostic model for processing to obtain the fault diagnosis results.
2. The method for fault diagnosis of railway train gearbox based on virtual-real feature fusion as described in claim 1, wherein the construction of the gearbox dynamic simulation model DM includes three-dimensional model establishment, model excitation, and fault modeling.
3. In the method for diagnosing gear faults in a rail train gearbox based on virtual-real feature fusion as described in claim 2, the kernel weights of the weighted convolutional unit C are: In the formula, These are the scaled and translated discrete time axis coordinates. t is the initial offset of the discrete time axis coordinates, a is the width scaling factor, and b is the position translation amount, which is updated as the model is trained. For scaling control factors, a ≠ ; The amplitude of the sample signal; The convolution kernel weight stationarity coefficient, , The characteristic frequency of gear faults; These are the kernel weights.
4. The method for fault diagnosis of railway train gearbox based on virtual-real feature fusion as described in claim 3, wherein the feature extraction module ME includes a semi-soft threshold noise reduction feature extraction submodule NR and a multi-scale feature extraction submodule PE, which weights the simulation sample set. The simulation features are extracted by the PE submodule of the multi-scale feature extraction module. The weighted measured sample set Simultaneously, the semi-soft threshold denoising module NR and the multi-scale feature extraction module submodule PE are input to extract the measured features after denoising. and multi-scale measured features The measured features are obtained by adding the results after assigning path weights: in This represents the path weight.
5. The method for fault diagnosis of rail train gearboxes based on virtual-real feature fusion as described in claim 4, wherein the semi-soft threshold noise reduction processing is as follows: In the formula, It is a two-channel attention mechanism; For adaptive average pooling; For feature extraction operations; This is a binary masking operation; The positive values are taken for the weighted measured samples; This represents the original weighted measured sample signal; These are learnable compensation coefficients that adaptively optimize during training.
6. The method for fault diagnosis of railway train gearbox based on virtual-real feature fusion as described in claim 3, wherein the virtual-real feature fusion operation is as follows: In the formula, For the first Class-weighted simulation sample set; The total number of samples corresponding to class c; For simulation features The class prototype vector; This is a temperature coefficient, adjusted based on the magnitude of the difference between the fused virtual and real samples; the greater the difference, the better. The larger the value; For the number of running status categories, For the simulation features of the i-th sample, To reconstruct simulation features; For measured features, This is the fusion coefficient.
7. The method for diagnosing gear faults in a rail train gearbox based on virtual-real feature fusion as described in claim 6, characterized in that, Cross-entropy loss function for: In the formula, Number of fault categories; For the first Similar to real-world labels; The model predicts that the sample belongs to the first... The probability of a class.