Traction motor bearing failure mode identification method based on CNN-GRU-RC-Transform hybrid neural network
By constructing a CNN-GRU-RC-Transformer hybrid neural network, the limitations of a single model in the recognition of the failure mode of the traction motor bearing are solved, efficient and accurate failure mode recognition is achieved, and the operation reliability and recognition accuracy of the traction motor are improved.
Patent Information
- Application Number
- CN202510680300.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-26
- Publication Date
- 2025-08-26
AI Technical Summary
In the prior art, a single neural network model has limitations in identifying failure modes of traction motor bearings, and it is difficult to efficiently and accurately identify multiple failure modes, and it relies on artificial feature extraction efficiency and is susceptible to subjective factors.
Using CNN-GRU-RC-Transformer hybrid neural network, by building data sets, training models, and combining cross-entropy loss function optimization, it combines the advantages of CNN, GRU and Transformer to achieve feature extraction, timing processing and gradient mitigation, and improves recognition accuracy.
High-precision recognition of the bearing failure mode of the traction motor is realized, which improves operating reliability and reduces maintenance costs. The recognition accuracy of the model in different data sets reaches more than 99.5%, showing good generalization performance.
Smart Images
Figure CN120541526A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of traction motor fault diagnosis, and in particular to a traction motor bearing failure mode recognition method based on a CNN-GRU-RC-Transformer hybrid neural network. Background Art
[0002] Traction motors, as critical power equipment in rail transit and other fields, have a significant impact on the safety and stability of the entire system through their operational reliability. Bearings, crucial components of traction motors, are prone to various failure modes, including inner race failure, ball bearing failure, and outer race failure. Bearing failure can damage the motor and even lead to serious safety incidents.
[0003] Traditional methods for traction motor bearing failure mode recognition rely primarily on manual feature extraction. This approach is not only inefficient, but also requires high operator expertise and is susceptible to subjective factors. With the development of deep learning technology, several single neural network models, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), their variants long short-term memory (LSTM) and gated recurrent units (GRUs), and the Transformer model, have begun to be applied to bearing failure mode recognition. However, these single models often have limitations when processing complex bearing vibration signals. For example, CNNs excel at extracting local features but are less capable of capturing long-sequence temporal dependencies. While GRUs can process time series data, they lack global features. Transformers have advantages in processing long-sequence data, but they are computationally complex and prone to overfitting when using small sample sizes. Therefore, developing a method that combines the strengths of multiple models to accurately and efficiently identify traction motor bearing failure modes is of great practical significance. Summary of the Invention
[0004] The purpose of the present invention is to provide a traction motor bearing failure mode recognition method based on a CNN-GRU-RC-Transformer hybrid neural network to address the limitations of a single model in the prior art in identifying traction motor bearing failure modes, achieve more accurate and efficient failure mode recognition, improve the reliability of traction motor operation, and reduce maintenance costs.
[0005] To achieve the above object, the technical solution adopted by the present invention is:
[0006] A traction motor bearing failure mode recognition method based on a CNN-GRU-RC-Transformer hybrid neural network is characterized by comprising the following steps:
[0007] S1. Constructing the dataset: The original dataset is formed through the data augmentation method of overlapping sampling; the original dataset is divided into non-overlapping training set and test set through random stratified sampling, and the sample distribution is ensured to be balanced.
[0008] S2. Training the CNN-GRU-RC-Transformer model: Build a CNN-GRU-RC-Transformer model and train it using the training set data. The model acquires sample features through forward propagation, quantifies the difference between the predicted value and the true value using a loss function, and uses a backpropagation algorithm to adjust parameters, continuously optimizing network performance until the set indicators are met. Finally, the trained model is output and saved. The cross-entropy loss function is selected as the loss function. The cross-entropy loss function expression is as follows:
[0009] (1)
[0010] S3, failure mode recognition and result evaluation: Input the test set data into the model obtained in S2 to perform failure mode recognition test, and use evaluation indicators such as failure mode recognition accuracy and confusion matrix to evaluate the failure mode recognition effect of the model.
[0011] Furthermore, the CNN-GRU-RC-Transformer model described in S2 includes the following steps:
[0012] S21, CNN feature extraction, including 3 convolutional layers, each convolutional layer has a convolution kernel size of 3×3 and a stride of 1. Each convolutional layer is followed by a ReLU activation function to enhance nonlinear capabilities; the convolutional layer is followed by a pooling layer, which uses a 2×2 maximum pooling method to downsample the features obtained by the convolutional layer to reduce the feature dimension and improve computational efficiency; a dropout layer is added after the CNN feature extraction process, and the dropout layer ratio is set to 0.3 to improve the generalization ability of the model;
[0013] S22, GRU time series processing, includes a three-layer GRU structure, which can further abstract and refine information based on the initial feature extraction, capturing long-term dependencies and more complex dependency patterns; a dropout layer is added after the GRU time series processing, with the dropout layer ratio set to 0.3 to improve the generalization ability of the model;
[0014] S23, RC residual connection, the output after GRU processing is added to the original input processed by residual_connection at each time step of the time series. By constructing residual connections, the gradient disappearance or gradient explosion phenomenon that occurs during deep network training can be effectively alleviated;
[0015] S24, Transformer encoding, contains 3 layers of Encoder. Each Encoder layer contains an 8-head self-attention mechanism and a 512-dimensional feedforward neural network, which is responsible for mapping the input sequence to a hidden layer representation. The EncoderLayer structure helps the model better capture the characteristics of the input sequence.
[0016] S25, classification output, outputs the identification result of the bearing failure mode through the fully connected layer fc.
[0017] Compared with the existing technology, the present invention has the following advantages: it innovatively integrates multiple advanced neural network structures, including CNN, GRU, RC, and Transformer, fully leveraging the advantages of each model. CNN can automatically learn local features, GRU effectively handles long-term and short-term dependencies in time series data, RC alleviates the gradient vanishing problem, and Transformer captures global dependencies in the sequence, thereby achieving high-precision identification of traction motor bearing failure modes. Experimental verification using different data sets has demonstrated that the model constructed by the present invention has good generalization performance, can adapt to traction motor bearing vibration signal data from different sources, and accurately identify failure modes. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 This is a flow chart of a method for identifying failure modes of traction motor bearings based on a CNN-GRU-RC-Transformer hybrid neural network according to the present invention;
[0019] Figure 2 This is a network structure diagram of a traction motor bearing failure mode recognition method based on a CNN-GRU-RC-Transformer hybrid neural network according to the present invention;
[0020] Figure 3 The failure recognition accuracy of the CNN-GRU-RC-Transformer model based on the bearing dataset of the Western Reserve University is shown in the figure.
[0021] Figure 4 The failure recognition accuracy of the CNN-GRU-RC-Transformer model based on the present invention on the same working condition data set of a rail vehicle manufacturer's test bench;
[0022] Figure 5 This is the failure mode recognition accuracy of the present invention based on the same data set of bearing vibration of Case Western Reserve University and different model test sets;
[0023] Figure 6 This is the failure mode recognition accuracy of the present invention based on a test set of different models under the same working condition on a test bench of a certain rail vehicle manufacturer;
[0024] Figure 7 This is a failure evaluation index obtained by the present invention based on a comparative experiment of a bearing data set from Case Western Reserve University;
[0025] Figure 8 This is a failure evaluation index obtained by the present invention based on a comparative experiment of a data set of the same operating condition on a test bench of a certain rail vehicle manufacturer. DETAILED DESCRIPTION
[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0027] 1. Data acquisition and preprocessing
[0028] Data source:
[0029] The vibration signal data of traction motor bearings was collected from the open source vibration signal dataset of the Case Western Reserve University Bearing Center and the traction motor test bench of a rail vehicle manufacturer.
[0030] (1) For the Case Western Reserve University dataset, we selected the motor drive end bearing data. This location is affected by the motor rotor rotation and the excitation of the transmission system, and can provide rich bearing failure information. Bearing failure types include inner race failure, ball failure, and outer race failure. Each failure type is further divided into three failure diameter sizes: 0.007 inches, 0.014 inches, and 0.021 inches. According to the bearing failure location and bearing failure diameter size, a total of 10 data types are divided.
[0031] (2) For the same working condition data set of a rail vehicle manufacturer's test bench, the normal operation, cage failure, roller failure, inner ring failure, and inner ring and roller combined failure modes of the traction motor bearing are simulated on the test bench, and vibration signals are collected under different speed conditions.
[0032] To construct a dataset (using the Case Western Reserve University dataset as an example): 1024 sampling points are used as a sample, with 512 overlapping sampling points. Each pattern corresponds to 233 samples, for a total of 2330 samples across 10 categories. The 2330 data samples are randomly shuffled, and 80% of them are randomly selected as the training set, with the remaining 20% as the test set. After randomly shuffling the 2330 samples, 80% is selected as the training set, and the remaining 20% is selected as the test set. Table 1 shows the results of constructing the Case Western Reserve University dataset.
[0033] Table 1 Case Western Reserve University dataset construction results
[0034]
[0035] 2. Model construction and training
[0036] Model construction: Build a CNN-GRU-RC-Transformer model according to the structure described in the invention content. Construct the CNN feature extraction submodule in sequence, set up 3 convolution layers and corresponding activation functions, pooling layers, and Dropout layers; construct the GRU time series processing submodule, set up a 3-layer GRU structure and a Dropout layer; construct the RC residual connection submodule to achieve the fusion of the GRU output and the CNN submodule input; construct the Transformer encoding submodule, set up a 3-layer Encoder and its internal self-attention mechanism and feedforward neural network. Finally, connect the fully connected layer of the classification output module.
[0037] Model training: The constructed model is trained using the training set data. During training, the training set is divided into batches, each containing 64 data points. Each iteration trains a batch of data and calculates the failure mode recognition accuracy and cross-entropy loss function. Because experimental results may be haphazard, the model is trained 10 times.
[0038] 3. Model testing and verification
[0039] Case Western Reserve University Dataset Test: The preprocessed Case Western Reserve University bearing data was used as the model input for testing. After 10 repeated experiments, the failure mode recognition accuracy of the test set of all experiments was above 99.6%, and the average accuracy reached 99.8%. One of the experiments was selected for analysis. The failure mode recognition accuracy of the model was basically stable when the model was iterated on the training set to the 25th time, and it was basically stable when it was iterated on the test set to the 35th time, and the overall accuracy exceeded 99%. The experimental results are as follows: Figure 3 As shown in the figure, the training set accuracy curve shows that the model can accurately identify the bearing failure mode under this working condition and has good robustness.
[0040] The same working condition data set of a railway vehicle manufacturer's test bench: The vibration signal data of a railway vehicle manufacturer's test bench under the same working condition (such as a speed of 800r / min) was tested. The model was trained 10 times. The failure mode recognition accuracy of the test set was above 99.5%, and the average accuracy reached 99.8%. The experimental results are as follows: Figure 4 As shown in the figure, the training set accuracy curve shows that the model can accurately identify bearing failure modes under this working condition and has good robustness.
[0041] 4. Comparative experiment
[0042] Ablation experiment: Ablation experiments were conducted on the same dataset of bearing vibration from a bearing test bench at Case Western Reserve University and the same working condition dataset from a test bench of a rail vehicle manufacturer. Single CNN, GRU, and Transformer models were constructed, as well as different combination models such as CNN-Transformer, CNN-GRU-Transformer, etc., and compared with the CNN-GRU-RC-Transformer model of the present invention. The experimental results are shown in the figure. Figure 5 and Figure 6 The experimental results show that the failure mode recognition accuracy of the proposed model is significantly higher than that of other comparison models, proving the necessity of the GRU network structure and the introduction of RC.
[0043] Other comparative experiments: The model of the present invention is compared with deep learning and machine learning algorithms such as LSTM, SVM, XGBoost, and random forest on the open source vibration signal dataset of the Case Western Reserve University Bearing Center and the same working condition dataset of a rail vehicle manufacturer's test bench. Evaluation indicators such as precision, recall rate, and F1 score of each model are calculated. The experimental results are as follows: Figure 7 and Figure 8 The results show that the proposed model outperforms other comparison models in terms of failure recognition accuracy, precision, recall rate, and F1 score, fully demonstrating the superiority of the proposed model in traction motor bearing failure pattern recognition performance.
[0044] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.
Claims
1. A traction motor bearing failure pattern recognition method based on CNN-GRU-RC-Transformer hybrid neural network, characterized in that: The following steps are involved: S1. Constructing the dataset: The original dataset is formed through the data augmentation method of overlapping sampling. The original dataset is divided into non-overlapping training and test sets through random stratified sampling, and the sample distribution is balanced. S2. Training the CNN-GRU-RC-Transformer model: Build a CNN-GRU-RC-Transformer model and train it using the training set data. The model obtains sample features through forward propagation, quantifies the difference between the predicted value and the true value using the loss function, adjusts parameters with the help of the backpropagation algorithm, and continuously optimizes network performance until the set indicators are met. Finally, the trained model is output and saved. The cross-entropy loss function is selected as the loss function. The cross-entropy loss function expression is as follows: （1） S3, failure mode recognition and result evaluation: Input the test set data into the model obtained in S2 to perform failure mode recognition test, and use evaluation indicators such as failure mode recognition accuracy and confusion matrix to evaluate the failure mode recognition effect of the model.
2. The method for identifying failure modes of traction motor bearings based on a CNN-GRU-RC-Transformer hybrid neural network according to claim 1, characterized in that: The CNN-GRU-RC-Transformer model described in S2 consists of the following steps: S21, CNN feature extraction, including 3 convolutional layers, each convolutional layer has a convolution kernel size of 3×3 and a stride of 1. Each convolutional layer is followed by a ReLU activation function to enhance nonlinear capabilities; the convolutional layer is followed by a pooling layer, which uses a 2×2 maximum pooling method to downsample the features obtained by the convolutional layer to reduce the feature dimension and improve computational efficiency; a dropout layer is added after the CNN feature extraction process, and the dropout layer ratio is set to 0.3 to improve the generalization ability of the model; S22, GRU time series processing, includes a three-layer GRU structure, which can further abstract and refine information based on the initial feature extraction, capturing long-term dependencies and more complex dependency patterns; a dropout layer is added after the GRU time series processing, and the dropout layer ratio is set to 0.3 to improve the generalization ability of the model; S23, RC residual connection, the output after GRU processing is added to the original input processed by residual_connection at each time step of the time series. By constructing residual connections, the gradient disappearance or gradient explosion phenomenon that occurs during deep network training can be effectively alleviated; S24, Transformer encoding, consists of 3 layers of encoders. Each encoder layer contains an 8-head self-attention mechanism and a 512-dimensional feedforward neural network, which is responsible for mapping the input sequence into a hidden layer representation. The EncoderLayer structure helps the model better capture the characteristics of the input sequence; S25, classification output, outputs the identification result of the bearing failure mode through the fully connected layer fc.