Fault diagnosis method for train transmission system

Abstract

The invention relates to the technical field of train fault diagnosis, in particular to a fault diagnosis method for a train transmission system, and overcomes the defects that weak features are difficult to extract and fault features are difficult to effectively distinguish in the prior art. In combination with the strong feature extraction capability of a deep network model constructed by an InceptionResNetV2 network, local and global features are captured by using a parallel multi-branch structure and a residual adaptive fusion mechanism, and through multi-scale perception and residual cross-layer connection, the capture capability of the super-huge bearing weak fault time sequence-frequency domain features can be significantly enhanced, and the fault detection accuracy is improved. And finally, pre-training a model by using source domain data, extracting parameterized knowledge of the model, and migrating to a target domain to realize parameter fine tuning and feature adaptation, so that the generalization performance and convergence speed of fault diagnosis can be improved, different fault categories can be effectively separated, and feature aggregation of the same category can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of train fault diagnosis, and in particular to a fault diagnosis method for a train transmission system. Background Art

[0002] The train drive system is a core component of modern rail transit vehicles. Its primary function is to convert electrical energy into mechanical energy to propel the train. The train drive system not only plays a critical role in power transmission but also directly impacts the train's operational efficiency, safety, and reliability. During the operational service phase, which represents the longest window of opportunity in the train's lifecycle, any minor fault or hidden danger, if not promptly diagnosed and effectively addressed, could lead to a major accident or even catastrophic consequences.

[0003] The full development of deep learning has made it a powerful tool for completing fault diagnosis technology. However, with the continuous development of train technology and the increasing complexity of the operating environment, fault diagnosis faces many challenges. The main difficulties encountered in the current methods of train transmission system fault diagnosis include complex data samples, large data volumes, diverse fault modes, and the existence of weak features that are difficult to extract. In addition, the operating state of the train transmission system is complex, and the data it generates often has nonlinear and non-stationary characteristics. Traditional linear methods (such as FFT and PCA) have difficulty in effectively distinguishing fault characteristics. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to overcome the shortcomings of the existing technology that weak features are difficult to extract and fault features are difficult to effectively distinguish, and to provide a fault diagnosis method for a train transmission system that can improve the model's perception of weak time-frequency features and effectively distinguish different fault categories.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for diagnosing a fault of a train transmission system is provided, comprising the following steps:

[0007] S1: Collect multi-modal signal data of the train transmission system under the same fault state in different working conditions and under different fault states in the same working condition.

[0008] S2: Utilize discrete short-time Fourier transform to convert the multimodal signals into time-frequency images.

[0009] S3: Stack multiple InceptionResNetV2 modules to construct a deep network model, and then use the time-frequency image to pre-train the deep network model; when pre-training the deep network model, use the InceptionResNetV2 modules at different levels to extract different features of the time-frequency image from local to global, and then splice the feature maps at the same time in different convolution channels to achieve multimodal fusion, and then add the spliced ​​map and the input time-frequency image through residual connection and output. The InceptionResNetV2 module is a deep convolutional neural network that combines the multi-scale feature extraction capability of the Inception module and the gradient optimization characteristics of residual connections. The deep network model constructed by stacking multiple InceptionResNetV2 modules can significantly enhance the ability to capture the time-frequency domain features of weak faults of extra-large bearings through multi-scale perception and residual cross-layer connections.

[0010] S4: The weight parameters of the pre-trained deep network model are used as the initialization parameters of the target domain training model, and the weight parameters are migrated to the target domain for parameter fine-tuning and feature adaptation training; the training is terminated after convergence on the set task to obtain a fault prediction model.

[0011] S5: Input a new multimodal signal to the fault prediction model, and the fault prediction model outputs a fault classification.

[0012] In the above technical solution, the multimodal signal is encoded into a time-frequency image through the discrete short-time Fourier transform (D-STFT). Combined with the powerful feature extraction capability of the deep network model constructed by the InceptionResNetV2 network, its parallel multi-branch structure and residual adaptive fusion mechanism are used to capture local and global features. Through multi-scale perception and residual cross-layer connection, the ability to capture the time-frequency domain features of weak faults of extra-large bearings can be significantly enhanced. Finally, the source domain data is used to pre-train the model, and its parameterized knowledge is extracted, which is then transferred to the target domain to achieve parameter fine-tuning and feature adaptation. Ultimately, the generalization performance and convergence speed of fault diagnosis can be improved, and different fault categories can be effectively separated and similar features can be aggregated, which is significantly better than the diagnosis method of traditional network models.

[0013] Preferably, in step S1, the multimodal signal includes a three-directional vibration signal and a three-phase current signal.

[0014] Preferably, in the discrete short-time Fourier transform of step S2, the constraint relationship between the window function length L, the number of frequency sampling points N and the time sampling interval R is N≥L≥R. The short-time Fourier transform signal is mapped to the time-frequency domain, and the time and frequency information of the signal is retained through this mapping method. The definition of Fourier transform is:

[0015]

[0016] The classical Fourier transform can only reflect the overall characteristics of the signal, and requires the signal to meet the stationary condition. From the definition of Fourier transform (1), it can be seen that in order to study the spectral characteristics of time-varying signals, all information in the time domain must be obtained. Time-frequency analysis methods, such as short-time Fourier transform (STFT) and Gabor transform (Gabor Transform), as well as Wigner-Ville distribution, can provide a joint representation of the signal in time and frequency, thereby more accurately analyzing the signal characteristics. The short-time Fourier transform (STFT) maps the signal to the time-frequency domain, and this mapping method retains the time and frequency information of the signal. The short-time Fourier transform of the signal is defined as:

[0017]

[0018] Where w[m] is the window function. For the discrete short-time Fourier transform, the value of m in equation (2) must satisfy w[m]≠0 (0≤m≤L-1). Outside [0,L-1], w[m]=0. Thus, equation (2) can be expressed as:

[0019]

[0020] If the STFT z (n,ω) at N equally spaced frequencies ω f =2πf / N, and N≥f, then the short-time Fourier transform after sampling can be expressed as:

[0021] STFT(n,f)=STFT(n,2πf / N) (4)

[0022] is equivalent to:

[0023]

[0024] 0≤f≤N-1 (5)

[0025] STFT z (n,f) is the discrete Fourier transform of the windowed sequence z[n+m]w[m]. Using the inverse discrete Fourier transform, we can get:

[0026]

[0027] 0≤m≤L-1 (6)

[0028] Since w[m]≠0 (0≤m≤L - 1), the following equation can be obtained:

[0029]

[0030] In Equation (7), the range of m is: 0≤m≤L - 1.

[0031] Restore the time series values within the interval from n to n + L - 1. Because w[m]≠0 (0≤m≤L - 1), Equations (4) and (5) are equivalent to sampling f in Equation (2). Sampling STFT(n, f) at time n enables the reconstruction of z[n] within the range -∞ < n < ∞. Specifically, using Equation (7), the signal can be reconstructed within the interval n0≤n≤n0 + L - 1 through STFT(n0, f), and also within the interval n0 + L≤n≤n0 + 2L - 1 through STFT(n0 + L, f), and so on. Therefore, the short-time Fourier transform sampled in the frequency and time dimensions can completely reconstruct z[n]. For the window function w[m]≠0 (0≤m≤L - 1), using the definitions in Equations (8) and (9) can be expressed more clearly as:

[0032] STFT(rR, f) = STFT(rR, 2πf / N) (8)

[0033] Equivalent to:

[0034]

[0035] In the formula, both r and f are integers, and the ranges are: -∞≤r≤∞, 0≤f≤N - 1.

[0036] Equations (8) and (9) involve the following integer parameters: the window function length L, the number of frequency sampling points N, and the time sampling interval R. Arbitrarily selecting these parameters cannot completely reconstruct the signal. Selecting L≤N can ensure the reconstruction of the windowed signal segment through the block transform STFT(n, f). If R < L, there will be overlap between signal segments; if R > L, some signal samples will not be available, and thus the original signal cannot be reconstructed through STFT(n, f). Therefore, in the discrete short-time Fourier transform, the sampling parameters satisfy N≥L≥R.

[0037] Preferably, in the step S3, the T-SNE (t-distributed stochastic neighborembedding) algorithm is used to visualize the features extracted by the InceptionResNetV2 module, and the confusion matrix is used to further visualize each fault classification result.

[0038] Preferably, in step S3, the InceptionResNetV2 module contains four independent convolution paths, wherein branch 1 uses 1×1 convolution for channel compression; branch 2 extracts local features through 1×1 convolution followed by 3×3 convolution; branch 3 expands the receptive field by cascading 1×1 convolution and two layers of 3×3 convolution; branch 4 uses 3×3 maximum pooling combined with 1×1 convolution to retain spatial information; to improve training stability, batch normalization and maximum pooling layers are introduced after each convolution layer.

[0039] Preferably, in step S3, each branch feature map is fused by channel concatenation and added to the input feature map by residual connection. The expression of the output of the InceptionResNetV2 module is:

[0040]

[0041] Where x is the input feature map and H(x) is the module output.

[0042] Preferably, in step S3, the batch normalization formula is:

[0043]

[0044] μ B and are the mean and variance of the current batch, γ and β are learnable parameters, and ò is a small constant. The normalized features are activated by the rectifier linear unit (ReLU), which is defined as:

[0045] ReLU(x)=max(0,x).

[0046] Preferably, before step S4, a high-level module and a low-level module in the InceptionResNetV2 module are determined, wherein the features extracted by the high-level module include semantic information, and the features extracted by the low-level module include detail texture.

[0047] Preferably, in step S4, the specific steps of parameter fine-tuning and feature adaptation training are: freezing the high-level module parameters of the deep network model to retain the general feature representation, unfreezing the low-level module parameters and fine-tuning using part of the labeled data of the target domain, optimizing the domain invariance of time-frequency feature extraction through back propagation, and using a dynamic learning rate decay strategy to balance knowledge transfer and target domain adaptation to avoid overfitting.

[0048] Preferably, in step S5, the deep network model compresses the features of the original input through global average pooling, and then the last fully connected layer outputs the fault type and converts it into probability through Softmax, and finally selects the fault category with the highest probability as the fault category output.

[0049] Compared with the existing technology, the beneficial effects of the present invention are: multimodal signals are encoded into time-frequency images through discrete short-time Fourier transform, and the powerful feature extraction capability of the deep network model constructed by the InceptionResNetV2 network is combined, and its parallel multi-branch structure and residual adaptive fusion mechanism are used to capture local and global features. Through multi-scale perception and residual cross-layer connection, the ability to capture the time series-frequency domain features of weak faults of extra-large bearings can be significantly enhanced. Finally, the source domain data is used to pre-train the model, and its parameterized knowledge is extracted, and then migrated to the target domain to achieve parameter fine-tuning and feature adaptation. Ultimately, the generalization performance and convergence speed of fault diagnosis can be improved, and different fault categories can be effectively separated and similar features can be aggregated, which is significantly better than the diagnostic method of the traditional network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 1. A flow chart of a method for diagnosing a fault of a train transmission system;

[0051] Figure 2 This is a simplified structural diagram of the train transmission system used in the experiment;

[0052] Figure 3 This is the relationship diagram of the data set in the experiment;

[0053] Figure 4 Patchwork graph for full-channel D-STFT encoding;

[0054] Figure 5 Patchwork diagram for full channel CWT encoding;

[0055] Figure 6 Encode the patchwork graph for the full channel GADF;

[0056] Figure 7 Coding patchwork diagram for full channel GASF;

[0057] Figure 8 This is a comparison chart of feature extraction levels;

[0058] Figure 9 This is one of the confusion matrix diagrams for the D-STFT-InceptionResnetV2 model on the test set classification;

[0059] Figure 10 Another confusion matrix diagram for the D-STFT-InceptionResnetV2 model on the test set classification;

[0060] Figure 11 This is the initial feature distribution graph after model feature distribution visualization;

[0061] Figure 12 The InceptionResnetV2 feature distribution diagram after model feature distribution visualization;

[0062] Figure 13 The InceptionV3 feature distribution diagram after model feature distribution visualization;

[0063] Figure 14 The Resnet101 feature distribution diagram after visualization of the model feature distribution;

[0064] Figure 15 Xception feature distribution diagram after visualization of model feature distribution;

[0065] Figure 16 This is the InceptionResNetV2-parallel feature distribution diagram after visualizing the model feature distribution.

[0066] In the accompanying drawings: 1- traction motor 1; 2- driving gear box; 3- left axle box; 4- right axle box. DETAILED DESCRIPTION

[0067] The present invention is further described below with reference to specific embodiments. The accompanying drawings are for illustrative purposes only and are schematic, not actual, representations. They should not be construed as limiting this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted from the drawings.

[0068] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0069] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0070] Example 1

[0071] like Figure 1 A first embodiment of a method for diagnosing a fault of a train transmission system is shown, which includes the following steps:

[0072] S1: Collect multi-modal signal data of the train transmission system under the same fault state in different working conditions and under different fault states in the same working condition.

[0073] S2: Use discrete short-time Fourier transform to convert multimodal signals into time-frequency images.

[0074] S3: Multiple InceptionResNetV2 modules are stacked to construct a deep network model, which is then pre-trained using time-frequency images. During pre-training, InceptionResNetV2 modules at different levels are used to extract local and global features from the time-frequency images. Feature maps from different convolutional channels at the same moment are then concatenated to achieve multimodal fusion. The concatenated image is then added to the input time-frequency image using residual connections for output. The InceptionResNetV2 module is a deep convolutional neural network that combines the multi-scale feature extraction capabilities of the Inception module with the gradient optimization properties of residual connections. The deep network model constructed by stacking multiple InceptionResNetV2 modules, through multi-scale perception and residual cross-layer connections, can significantly enhance the ability to capture the time-frequency domain features of subtle faults in extra-large bearings.

[0075] S4: Use the weight parameters of the pre-trained deep network model as the initialization parameters of the target domain training model, and migrate the weight parameters to the target domain for parameter fine-tuning and feature adaptation training; end the training after convergence on the set task to obtain a fault prediction model.

[0076] S5: Input the new multimodal signal into the fault prediction model, and the fault prediction model outputs the fault classification.

[0077] Furthermore, in step S1 , the multimodal signal includes a three-directional vibration signal and a three-phase current signal.

[0078] Furthermore, in the discrete short-time Fourier transform of step S2, the constraint relationship between the window function length L, the number of frequency sampling points N, and the time sampling interval R is N≥L≥R. The short-time Fourier transform signal is mapped to the time-frequency domain, and this mapping method preserves the time and frequency information of the signal. The definition of the Fourier transform is:

[0079]

[0080] The classical Fourier transform can only reflect the overall characteristics of the signal, and requires the signal to meet the stationary condition. From the definition of Fourier transform (10), it can be seen that in order to study the spectral characteristics of time-varying signals, all information in the time domain must be obtained. Time-frequency analysis methods, such as short-time Fourier transform (STFT) and Gabor transform (Gabor Transform), as well as Wigner-Ville distribution, can provide a joint representation of the signal in time and frequency, thereby more accurately analyzing the signal characteristics. The short-time Fourier transform (STFT) maps the signal to the time-frequency domain, and this mapping method retains the time and frequency information of the signal. The short-time Fourier transform of a signal is defined as:

[0081]

[0082] Where w[m] is the window function. For the discrete short-time Fourier transform, the value of m in equation (2) must satisfy w[m]≠0 (0≤m≤L-1). Outside [0,L-1], w[m]=0. Thus, equation (2) can be expressed as:

[0083]

[0084] If the STFT z (n,ω) at N equally spaced frequencies ω f =2πf / N, and N≥f, then the short-time Fourier transform after sampling can be expressed as:

[0085] STFT(n,f)=STFT(n,2πf / N) (13)

[0086] is equivalent to:

[0087]

[0088] 0≤f≤N-1 (14)

[0089] STFTz (n, f) is the discrete Fourier transform of the windowed sequence z[n + m]w[m]. Using the inverse discrete Fourier transform, we can obtain:

[0090]

[0091] 0 ≤ m ≤ L - 1 (15)

[0092] Since w[m] ≠ 0 (0 ≤ m ≤ L - 1), the following equation can be obtained:

[0093]

[0094] The range of m in Equation (7) is: 0 ≤ m ≤ L - 1.

[0095] Restore the time series values in the interval from n to n + L - 1. Because w[m] ≠ 0 (0 ≤ m ≤ L - 1), Equations (4) and (5) are equivalent to sampling f in Equation (2). Sampling the time n of STFT(n, f) can reconstruct z[n] within the range of -∞ < n < ∞. Specifically, using Equation (7), the signal can be reconstructed within the interval n0 ≤ n ≤ n0 + L - 1 through STFT(n0, f), and can also be reconstructed within the interval n0 + L ≤ n ≤ n0 + 2L - 1 through STFT(n0 + L, f), and so on. Therefore, the short-time Fourier transform sampled in the frequency and time dimensions can completely reconstruct z[n]. For the window function w[m] ≠ 0 (0 ≤ m ≤ L - 1), using the definitions of Equations (8) and (9) can represent it more clearly:

[0096] STFT(rR, f) = STFT(rR, 2πf / N) (17)

[0097] Equivalent to:

[0098]

[0099] In the formula, both r and f are integers, and the ranges are: -∞ ≤ r ≤ ∞, 0 ≤ f ≤ N - 1.

[0100] Equations (8) and (9) involve the following integer parameters: the window function length L, the number of frequency sampling points N, and the time sampling interval R. Not any arbitrary selection of these parameters can completely reconstruct the signal. Selecting L ≤ N can ensure the reconstruction of the windowed signal segment through the block transform STFT(n, f). If R < L, there will be overlap between signal segments; if R > L, some signal samples will not be used, and thus the original signal cannot be reconstructed through STFT(n, f). Therefore, in the discrete short-time Fourier transform, the sampling parameters satisfy N ≥ L ≥ R.

[0101] Effect verification of this embodiment:

[0102] 1. Data Collection

[0103] The test verification was carried out using the subway bogie fault simulation test bench of the National Key Laboratory of Advanced Rail Transit Autonomous Operation. Figure 2 As shown, it includes a traction motor 1, a drive gear box 2, a left axle box 3 and a right axle box 4.

[0104] The test bench used in this embodiment is based on a real subway bogie and is constructed at a scale of 1:2, which can obtain multi-source sensor signals under different health states. A three-phase asynchronous AC motor drives a single transmission chain, which includes a traction motor (bearing model SKF6205-2RSH), a reduction gearbox (helical gear, driving gear 16 teeth / driven gear 107 teeth, support bearing model Harbin Bearing Factory 32305) and an axle box (bearing model Harbin Bearing Factory 352213) to form the mechanical transmission system of the test bench. The hydraulic loading system of the simulation test bench reproduces the train running load through dynamic load technology, and can output ±10kN cyclic load laterally to simulate track vibration. It is equipped with a variable frequency motor to achieve three-speed speed adjustment of 20 / 40 / 60Hz.

[0105] The test bench has 21 sensor channels with a sampling frequency of 64kHz. Signal types include three-dimensional vibration and three-phase current signals. The three-dimensional vibration signals are collected by 18 acceleration sensors (CH1-CH18) deployed on the motor's drive and non-drive ends, the gearbox input and output shafts, and the left and right axlebox end caps. The three-phase current signals are collected by sensors located on the motor cables (CH19-CH21). Table 1 provides a comparison of the sensor channel definitions.

[0106] Table 1 Sensor channel allocation

[0107]

[0108] As shown in Table 2, the dataset includes nine operating conditions. Different motor speeds simulate different train speeds, and different lateral loads simulate straight-line and cornering. A positive lateral load indicates loading toward the motor side of the drive train under test, while a negative lateral load indicates loading toward the gearbox side of the drive train under test.

[0109] Table 2 Overview of Working Condition Settings

[0110]

[0111] As summarized in Table 3, the health states of the components in the dataset (motor, gearbox drive gear and support bearing, and axlebox bearing) include normal and 16 fault conditions. For each health state, three samples are retained, and each sample contains 640,000 sampling points per channel.

[0112] Tab.3 Health state in the datasets

[0113]

[0114]

[0115] The labels and corresponding faults for the training and validation sets are shown in Table 4. It can be seen that the health status of the axle box motor end (right) is normal, which means that RA0 is normal. Therefore, the corresponding channel can be discarded when considering the problem.

[0116] Table 4 Fault code list

[0117]

[0118] The working conditions of the training set and the test set are unknown, and the working conditions of the test set are different from the previous two. The test set is also generated by a simulation test bench. The test set has 102 samples, and each channel of each sample contains 64,000 sampling points. The purpose of the above processing method is to increase the difficulty of feature extraction and thus increase the difficulty of classification, making the performance difference between models more significant. The relationship between the data sets is as follows Figure 3 As shown in Figure 1, in Set A, the training set and validation set are the same batch of data with the same operating conditions. Set B is an independent data set with different operating conditions from Set A.

[0119] 2. Image Conversion and Feature Extraction

[0120] Taking a sample with a health status of TYPE0 as an example, according to the sampling sequence, the 1024 sample points of each channel are encoded into a D-STFT image with a resolution of 625×625. Figure 4 As shown in FIG, the D-STFT images of 18 channels (excluding the three redundant channels ch19-ch21) are spliced ​​from left to right into a combined image with a resolution of 3750×1875.

[0121] from Figure 4It's not difficult to see significant differences between the images of different signal types. The first row shows the three-axis acceleration values ​​of the traction motor (corresponding to ch1-ch6), the first three images in the second row show the three-phase current values ​​(corresponding to ch7-ch9), and the last three images in the second row and the third row show the three-axis acceleration values ​​of the gearbox and axle box (corresponding to ch10-ch16). Using these images as input data for the fault diagnosis model provides the model with more diverse and richer feature information, further improving the model's accuracy in fault recognition, significantly enhancing its diagnostic capabilities, and better meeting the precise detection needs of practical applications.

[0122] Similarly, in order to verify that D-STFT images can better extract the characteristics of the dataset, the dataset is encoded into CWT, GADF, and GASF under the same conditions. Then the images obtained by these four methods are put into the same model to compare the results. Figures 5 to 7 shown.

[0123] The experiment of this embodiment adopts Deep Learning Toolbox deep learning framework, and the configuration used is: Windows11 system, Intel (R) Core (TM) i7-14650HX CPU and RTX4060 graphics card, running memory 32G. In order to ensure the reliability and stability of the experimental results and avoid the accidental errors that may occur in a single test, this embodiment adopts a repeated measurement method: each group of experimental conditions is tested independently 4 times, and the arithmetic mean is taken as the final result after eliminating obvious outliers, and the standard deviation is calculated to evaluate the degree of data dispersion. All models use the SGDM optimization algorithm and ReLU activation function, and the loss function is cross-entropy loss (Cross-Entropy Loss). The learning rate strategy adopts piecewise constant decay: the initial learning rate is 0.01, and the learning rate is multiplied by 0.1 every 1 epoch. Set A is divided into a training set of 0.85 and a validation set of 0.15. Set B consists of 102 test sets of different working conditions that require judgment. The training set is used to train the model, the validation set is used to evaluate model performance during training and help select model hyperparameters, and the test set is used to ultimately evaluate the generalization ability of the trained model. Each training batch size is 32, and a total of 20 iterations are performed.

[0124] like Figures 4 to 7 As shown in the figure, a preliminary visual observation reveals significant differences between the three-phase current values ​​and the three-axis acceleration values. Except for the CWT-encoded image, the three-phase current values ​​of different channels in all other images exhibit varying degrees of difference. The most obvious differences in the three-axis acceleration values ​​between different channels are in the D-STFT-encoded image. Visual evaluation shows that the D-STFT-encoded image has the most distinct features, outperforming images encoded using other methods.

[0125] In order to further accurately evaluate the level of feature extraction of the four types of images, the D-STFT, GASF, GADF, and CWT images were put into the InceptionResNetV2 model respectively. The performance in the validation set and the test set is as follows: Figure 8 shown.

[0126] according to Figure 8 The data shows that the D-STFT method exhibits the highest classification accuracy and the lowest loss value on both the validation set and the test set, which shows that the algorithm has a significant advantage in feature representation capabilities. This shows that the D-STFT method can more effectively extract and encode key feature information in the image. Its multi-dimensional time-frequency features are not only more distinguishable, but also provide the model with richer discrimination basis. Through comparative analysis with other feature encoding methods, it can be seen that D-STFT not only shows the best performance during training, but also significantly outperforms other methods in model generalization ability, showing significant advantages in the model training and optimization process.

[0127] 3. Model Effect Analysis

[0128] In order to verify the effectiveness of D-STFT-InceptionResNetV2, multiple groups of control variable experiments were conducted and the results of different experiments were compared.

[0129] We tested dataset A (D-STFT encoding) with several classic and highly effective neural networks: AlexNet, VGG19, GoogleNet, ResNet-18, and SqueezeNet. We found that none of these networks converged effectively on this dataset. This reflects the complex features of this dataset, which cannot be effectively learned by typical deep neural networks.

[0130] The main InceptionResNetV2 network architecture was spliced ​​into a parallel architecture to create the InceptionResNetV2-parallel network. Dataset A was simultaneously fed into InceptionResNetV2, Inception-V3, Resnet101, InceptionResNetV2-parallel, and Xception. Dataset B (the test set) was then fed into the trained model. The performance of the five different models on the validation and test sets is shown in Table 5.

[0131] Table 5 The model's performance on the validation and test sets Unit: %

[0132]

[0133] Table 5 shows that the InceptionResNetV2 model achieved the best performance on both the training and test sets. On the validation set, this model achieved an accuracy of 97.87%, surpassing Inception-V3's 94.53%, Resnet101's 93.82%, and Xception's 93.0% by 3.34%, 4.05%, and 4.89%, respectively. On the test set, this model achieved an accuracy of 90.20%, still 2.49% higher than the next-best model, Inception-V3's 87.71%, and significantly ahead of Resnet101's 85.69% and Xception's 85.19%, with performance advantages ranging from 4.51% to 5.01%. Notably, the InceptionResNetV2-parallel model performed significantly worse, achieving the lowest accuracy on both the validation set (86.73%) and the test set (80.43%). The experimental results fully demonstrate that InceptionResNetV2 has significant advantages in complex feature learning and model generalization capabilities. The test set has 102 classification data sets, each with 62 D-STFT encoded images. The performance of the InceptionResNetV2 model on the test set is shown in the confusion matrix. Figure 9 and Figure 10 shown. Figure 9 is the result of classifying all images, Figure 10 The results are for 102 datasets to be classified. This shows that the D-STFT-InceptionResNetV2 model can maintain a relatively good classification accuracy rate on test sets that are different from the training and validation sets.

[0134] This embodiment uses T-SNE nonlinear dimensionality reduction technology to perform visual analysis on the deep features extracted from the output layer. Figures 11 to 16 As shown, Figure 11 The original feature distribution of the target domain shows obvious category confusion, and the feature points of each fault type overlap in disorder; Figures 12 to 16 The figure below shows the distribution of the last layer features extracted from each model. This shows that InceptionResnetV2 can effectively distinguish different fault modes in a mixed state and accurately cluster similar faults in specific areas. Although a small number of fault samples are misclassified as other categories, it achieves high fault clustering accuracy overall. The other models have weaker fault aggregation capabilities, especially InceptionResnetV2-parallel, where some samples overlap and are not fully separated.

[0135] The above experiments show that the D-STFT-InceptionResnetV2 model demonstrates excellent adaptive feature learning capabilities. This model can not only extract weak features to effectively distinguish different fault types, but also accurately aggregate the features of similar faults.

[0136] Example 2

[0137] This embodiment provides further supplementary explanations on steps S3 and S4 based on embodiment 1.

[0138] Furthermore, in step S3, the features extracted by the InceptionResNetV2 module are visualized using the T-SNE (t-distributed stochastic neighbor embedding) algorithm, and each fault classification result is further visualized using the confusion matrix.

[0139] Furthermore, in step S3, the InceptionResNetV2 module contains four independent convolution paths, among which branch 1 uses 1×1 convolution for channel compression; branch 2 extracts local features through 1×1 convolution followed by 3×3 convolution; branch 3 expands the receptive field by cascading 1×1 convolution and two layers of 3×3 convolution; branch 4 combines 3×3 maximum pooling with 1×1 convolution to retain spatial information; the scaling factor of the residual adaptive fusion mechanism in this model is between 0.1 and 0.2 to enhance cross-layer feature interaction; 17 groups of repeated InceptionResNetV2 modules are stacked to achieve deep feature optimization; to improve training stability, batch normalization and maximum pooling layers are introduced after each convolution layer.

[0140] Furthermore, in step S3, each branch feature map is fused through channel concatenation and added to the input feature map through residual connection. The expression of the InceptionResNetV2 module output is:

[0141]

[0142] Where x is the input feature map, H ( x ) is the module output.

[0143] Furthermore, in step S3, the formula for batch normalization is:

[0144]

[0145] μ B and are the mean and variance of the current batch, γ and β are learnable parameters, and ò is a small constant. The normalized features are activated by the rectifier linear unit (ReLU), which is defined as:

[0146] ReLU(x)=max(0,x).

[0147] Furthermore, before step S4, a high-level module and a low-level module in the InceptionResNetV2 module are determined, wherein the features extracted by the high-level module include semantic information, and the features extracted by the low-level module include detail texture.

[0148] Furthermore, in step S4, the specific steps of parameter fine-tuning and feature adaptation training are to freeze the high-level module parameters of the deep network model to retain the common feature representation, unfreeze the low-level module parameters and use the partial labeled data of the target domain for fine-tuning, optimize the domain invariance of time-frequency feature extraction through back propagation, and use a dynamic learning rate decay strategy to balance knowledge transfer and target domain adaptation to avoid overfitting.

[0149] Example 3

[0150] This embodiment further supplements step S5 based on Example 2. In step S5, the deep network model compresses the features of the original input through global average pooling. The final fully connected layer outputs the fault type and converts it into a probability through Softmax. Finally, the fault category with the highest probability is selected as the output fault category.

[0151] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing a fault in a train transmission system, characterized in that: The following steps are involved: S1: Collect multi-modal signal data of the train transmission system under the same fault state in different working conditions and under different fault states in the same working condition; S2: using discrete short-time Fourier transform to convert the multimodal signals into time-frequency images; S3: stacking multiple InceptionResNetV2 modules to construct a deep network model, and then using the time-frequency image to pre-train the deep network model; when pre-training the deep network model, using the InceptionResNetV2 modules at different levels to extract different features of the time-frequency image from local to global, and then splicing the feature maps at the same time in different convolution channels, and then adding the spliced ​​map to the input time-frequency image through a residual connection and outputting the result; S4: Using the weight parameters of the pre-trained deep network model as initialization parameters for the target domain training model, migrating the weight parameters to the target domain for parameter fine-tuning and feature adaptation training; ending the training after convergence on the set task to obtain a fault prediction model; S5: Input a new multimodal signal to the fault prediction model, and the fault prediction model outputs a fault classification.

2. A fault diagnosis method for a train transmission system according to claim 1, characterized in that: In step S1 , the multimodal signal includes a three-directional vibration signal and a three-phase current signal.

3. A fault diagnosis method for a train transmission system according to claim 1, characterized in that: In the discrete short-time Fourier transform of step S2, the constraint relationship among the window function length L, the number of frequency sampling points N and the time sampling interval R is N≥L≥R.

4. A method for diagnosing a fault in a train transmission system according to claim 1, characterized in that: In step S3, the features extracted by the InceptionResNetV2 module are visualized using the T-SNE algorithm, and each fault classification result is further visualized using a confusion matrix.

5. The method for diagnosing a fault of a train transmission system according to claim 1, wherein: In step S3, the InceptionResNetV2 module includes four independent convolution paths, wherein branch 1 uses 1×1 convolution for channel compression; branch 2 extracts local features through 1×1 convolution followed by 3×3 convolution; Branch 3 expands the receptive field by cascading 1×1 convolution and two layers of 3×3 convolution; branch 4 combines 3×3 maximum pooling with 1×1 convolution to retain spatial information; batch normalization and maximum pooling layers are introduced after each convolution layer.

6. A method for diagnosing a fault in a train transmission system according to claim 1, characterized in that: In step S3, the expression output by the InceptionResNetV2 module is: Where x is the input feature map and H(x) is the module output.

7. A method for diagnosing a fault in a train transmission system according to claim 1, characterized in that: In step S3, the batch normalization formula is: μ B and are the mean and variance of the current batch, γ and β are learnable parameters, and ò is a small constant. The normalized features are activated by a linear rectifier unit and are defined as: ReLU(x)=max(0,x).

8. A fault diagnosis method for a train transmission system according to claim 1, characterized in that: Before step S4, a high-level module and a low-level module in the InceptionResNetV2 module are determined, wherein the features extracted by the high-level module include semantic information, and the features extracted by the low-level module include detail texture.

9. A method for diagnosing a fault in a train transmission system according to claim 5, characterized in that: In step S4, the specific steps of parameter fine-tuning and feature adaptation training are: freezing the high-level module parameters of the deep network model to retain the general feature representation, unfreezing the low-level module parameters and fine-tuning using part of the labeled data of the target domain, optimizing the domain invariance of time-frequency feature extraction through back propagation, and using a dynamic learning rate decay strategy to balance knowledge transfer and target domain adaptation.

10. A fault diagnosis method for a train transmission system according to any one of claims 1 to 9, characterized in that: In step S5, the deep network model compresses the features of the original input through global average pooling, and then the last fully connected layer outputs the fault type and converts it into probability through Softmax, and finally selects the fault category with the highest probability as the fault category output.

Images