A gearbox fault diagnosis method based on MCNN and feature fusion ViT

Abstract

This invention relates to the field of gearbox fault diagnosis, and specifically to a gearbox fault diagnosis method based on MCNN and Feature Fusion ViT. This invention utilizes MCNN to capture feature information from data at multiple scales, fully extracting local features of the input information; then, it obtains global features of the fault information through the multi-head attention mechanism in the ViT network, and uses depthwise separable convolution to reduce the number of model parameters. Experimental verification using gearbox data shows that the proposed model has high fault diagnosis accuracy and good generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of gearbox fault diagnosis, and in particular to a gearbox fault diagnosis method based on MCNN and feature fusion ViT. Background Art

[0002] Large extrusion granulators are key equipment in petrochemical enterprises. They are mainly used to convert powdered raw materials such as polyolefins generated in polymer reactors into particles with regular geometric shapes. Due to the complex structure of the box, the variable influencing factors and the various forms of faults, it brings great obstacles to the location of the fault source and the identification of the fault type. The most common gear faults include wear, cracks, fractures and pitting. According to the operating characteristics and fault types of the gearbox, clarifying the mechanism and characteristics of the fault, and reasonably selecting the sensor network layout and fault diagnosis methods are important foundations for realizing fault diagnosis of petrochemical extrusion granulators.

[0003] The current conventional management method is to combine equipment operation status monitoring with manual inspections. However, due to the incompleteness of monitoring data, the lag of manual inspections, and the limited human experience, online diagnosis and prediction of equipment failures have not achieved significant results.

[0004] With the further development of technology, gearbox vibration signals are mostly analyzed to diagnose the operating status, occurrence and type of faults. The main diagnostic methods include dynamic model-based diagnosis and data-driven machine learning technology. Machine learning technology can be further divided into shallow machine learning and deep learning, each of which has unique advantages in dealing with complex fault diagnosis problems.

[0005] In summary, there are mainly three existing technical routes:

[0006] 1. Fault prediction method based on traditional dynamic model. Real-time fault prediction is achieved by analyzing the inherent characteristics of fault signals. A mathematical dynamic model is constructed by combining physical factors such as machining accuracy, transmission error, transmission efficiency, time-varying meshing stiffness, bearing friction, meshing error, etc. to simulate the gear running state and achieve the purpose of fault prediction.

[0007] This method has certain complexity in the modeling process, especially for large units such as extrusion granulators in the petrochemical industry, modeling is particularly difficult, thus limiting its wide application.

[0008] 2. Fault diagnosis methods based on traditional machine learning. Machine learning improves the ability to process and analyze data by continuously training the model's pattern recognition capabilities using labeled samples. It focuses on enabling computer systems to learn from data and optimize themselves. By analyzing large amounts of sample data, patterns and associations are automatically identified, achieving the effect of making corresponding decisions and predictions without explicit programming.

[0009] Among them, shallow machine learning methods require manual extraction of effective data features as input to the machine learning algorithm, so as to identify the fault information contained in the data. Time domain analysis is mainly based on parameter analysis and diagnosis of vibration signals such as amplitude and frequency, but the diagnostic effect of this method is often greatly affected by the operating status of the equipment, especially in the fault state. The non-stationarity and nonlinearity of the signal may lead to inaccurate analysis results. Frequency domain analysis focuses on separating and strengthening the frequency components related to the fault. This method is more intuitive in showing fault characteristics, but its analysis accuracy may be limited for signals with complex frequency components.

[0010] 3. Deep learning is a machine learning method based on multi-level complex artificial neural networks. In essence, it simulates neurons and builds multi-layer neural networks to achieve complex calculations and reasoning processes. In its early development, deep learning-based algorithms fell into a trough because they could not solve linear inseparable problems such as XOR. However, with the introduction of activation functions such as Sigmoid, Tanh, and ReLU, neural networks have achieved delinearization of calculation results, thus possessing nonlinear fitting capabilities and making great progress. In specific tasks, deep learning can automatically analyze and learn data, effectively mining feature representations of various levels and dimensions from data, thereby demonstrating stronger data parsing and understanding capabilities. At present, commonly used technologies in the field of deep learning include deep belief networks (DBNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), and Transformers.

[0011] Although deep learning technology performs well, its performance depends on sufficient and balanced labeled data samples. This requirement has become an important limitation of deep learning applications in the field of fault diagnosis.

[0012] The deep learning model automatically learns feature representation from data through end-to-end learning, and can better adapt to the data features under different fault modes and working conditions in the fault diagnosis task, with stronger generalization and adaptability. However, in the actual working environment, since the gearbox of the extrusion granulator is in a normal state most of the time, the number of fault samples is very scarce. Therefore, the fault diagnosis model must have a strong feature extraction capability under a limited number of samples to complete the fault diagnosis task. Summary of the invention

[0013] In view of the shortcomings of the prior art, the present invention provides a lightweight gearbox fault diagnosis model based on multi-scale convolutional neural network and feature fusion ViT. The present invention uses MCNN to capture the feature information of data from multiple scales and fully extract the local features of the input information; then, the global features of the fault information are obtained through the multi-head attention mechanism in the ViT network, and the deep separable convolution is used to reduce the number of model parameters. Experimental verification is carried out using gearbox data, and the results show that the proposed model has high fault diagnosis accuracy and good generalization ability.

[0014] The technical solution of the present invention is: a gearbox fault diagnosis method based on MCNN and feature fusion ViT, characterized in that it includes the following steps:

[0015] Step 1: Normalize the one-dimensional vibration data of the gearbox collected by the sensor and divide it into a training set and a test set;

[0016] Step 2: Initialize the model parameters, and send the training set data into the model to extract fault features, and use the back propagation optimization algorithm to optimize the model parameters.

[0017] Step 21: Initialize the parameters of the convolutional layer, fully connected layer, multi-head attention mechanism, etc.

[0018] Step 22: Multi-scale feature extraction module:

[0019] The input training set data passes through the multi-scale convolution extraction module. The convolution kernel slides on the input signal to extract local features at different positions. The first layer of convolution: the convolution kernel size is 300×1, which extracts the overall general features of the vibration signal;

[0020] The second convolution layer: the convolution kernel size is 200×1, focusing on the changing characteristics of the vibration signal;

[0021] The third convolution layer: the convolution kernel size is 100×1, focusing on the detailed features of the vibration signal;

[0022] Finally, after three layers of convolution, a feature map sequence is generated;

[0023] Step 23: Lightweight feature fusion ViT:

[0024] The input feature map sequence is divided into multiple subsequences, and each subsequence is linearly projected and mapped to different subspaces; the attention weight is calculated independently in each subspace, and residual connections are introduced before and after each single-head attention mechanism; the multi-head attention output feature vector is obtained by weighted summation;

[0025] Step 24: Deep Separable Multilayer Perceptron D-MLP: Map the input feature vector to a higher dimensional space through a fully connected layer to capture complex features, and use point-by-point convolution to reduce the number of model parameters; Use GELU activation function: Use GELU function to enhance the convergence and stability of the network;

[0026] Step 25: Classification layer: The extracted features are mapped to fault categories through the classification layer, and the fault diagnosis results are output;

[0027] Step 26: Optimize the parameters using the back propagation optimization algorithm;

[0028] Step 3: Determine whether the model iteration batch is greater than the preset value. If it is greater, save the model parameters. If it is less than or equal to it, return to step 2 to continue training.

[0029] Step 4: Send the test set data into the trained model to test the model, and the classification layer outputs the fault diagnosis results.

[0030] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, it is characterized in that: in step 1, 90% of the data is a training set and 10% of the data is a test set.

[0031] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, it is characterized in that: in step 21, the parameters of the convolution layer include the convolution kernel weight and the bias item corresponding to each convolution layer; the parameters of the fully connected layer include the weight matrix and the weight matrix of the fully connected layer; the parameters of the multi-head attention mechanism are the projection weight matrix and the parameters of the optimized residual connection.

[0032] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, the characteristic is that: the specific process of step 23 is: dividing the input feature map sequence into multiple subsequences, and linearly projecting each subsequence, mapping it to different subspaces to generate representations of multiple subsequences; in each subspace, independently calculating the attention weight by formula (1) to formula (4),

[0033] The weighted sum of the values ​​calculated for the attention of each subspace is performed, and the output of the multi-head attention is obtained by formula (5):

[0034] Q=X f W Q (1)

[0035] K=X f W K (2)

[0036] V=X f WV (3)

[0037]

[0038] MultiHead(Q,K,V)=Concat(head1,head2,…,head n )W (5)

[0039] Where: Concat represents the concatenation operation, W represents the weight matrix of the mapping, Q, K, V represent the query, key, and value matrices respectively, X represents the feature matrix, W represents the learnable matrix, and d represents the dimension of the matrix.

[0040] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, it is characterized in that: the specific process of step 24 is: first, the separable multi-layer perceptron D-MLP maps the multi-head attention output feature vector to a higher dimensional space through the first fully connected layer according to matrix multiplication and bias term addition, and then uses the second fully connected layer to map the output of the previous layer again to a lower dimensional space, reducing the dimension of the feature vector, and using deep convolution and point-by-point convolution step calculation to greatly reduce the number of model parameters; finally, the GELU function is used to enhance the convergence and stability of the network, and the formula (6) of the GELU function is:

[0041]

[0042] Among them: x represents the input information of the MLP module; tanh represents the hyperbolic tangent function.

[0043] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, it is characterized in that: in step 25, the classification layer adopts a Softmax classifier to convert the output feature nonlinear score into a probability distribution, and each category corresponds to a probability value.

[0044] According to the gearbox fault diagnosis method based on MCNN and feature fusion ViT as described above, it is characterized in that: the model iteration batch range in step 3 is between 100 and 200 times.

[0045] The beneficial effects of the present invention are: in an environment with few samples, the fault diagnosis task can still be completed, the computational complexity of the model is reduced, and the generalization ability and stability of the model are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a model diagnosis flow chart of the present invention.

[0047] Figure 2 This is the model network structure diagram.

[0048] Figure 3 It is a multi-scale feature extraction module.

[0049] Figure 4 This is the structure diagram of the multi-head attention mechanism.

[0050] Figure 5 This is the structure diagram of D-MLP.

[0051] Figure 6 This is a diagram of the experimental platform installation.

[0052] Glossary: ​​MCNN: (Multi-scale Convolutional Neural Network) Multi-scale Convolutional Neural Network

[0053] ViT: Feature Fusion.

[0054] DSC: (Depthwise separable convolution, DSC) depth-wise separable convolution.

[0055] D-MLP: (Depth-separable multilayer perceptrons) separable multilayer perceptron. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0057] Gearbox gears usually work in a very complex environment. In the process of fault diagnosis, the adequacy of feature extraction will directly affect the diagnostic effect. First, the one-dimensional vibration signal collected by the sensor is normalized and used as the model input; then the designed multi-scale convolution extraction module is used to extract the local features of the fault information and resist noise interference; then the global feature information is extracted through lightweight feature fusion ViT, in which the multi-head attention mechanism fuses the features of different subspaces, and the deep separable multi-layer perceptron (D-MLP) linearly maps the fused features and reduces the number of parameters; finally, the classification head is used to classify the fault information and output the recognition result. The network structure diagram of the present invention is shown in the figure. Figure 2 shown.

[0058] like Figure 1 As shown in the figure, a gearbox fault diagnosis method based on MCNN and feature fusion ViT of the present invention comprises the following steps:

[0059] Step 1: Normalize the one-dimensional vibration data of the gearbox collected by the sensor and divide it into a training set and a test set according to a fixed ratio.

[0060] The training set of the present invention is 70% to 90% of the total data, and the test set is 10% to 30% of the total data. The best invention is to use 90% of the training set and 10% of the test set for division, so that the training resources can be fully utilized under the condition of small samples.

[0061] Step 2: Initialize the model parameters, and send the training set data into the model to extract fault features, and use the back propagation optimization algorithm to optimize the model parameters.

[0062] Step 21: Initialize parameters such as convolutional layer, fully connected layer, and multi-head attention mechanism

[0063] (I) Parameters of the convolutional layer:

[0064] a. Convolution kernel weights: Convolution kernel weights in the multi-scale feature extraction module, including the first layer convolution kernel size of 300 × 1, the second layer convolution kernel size of 200 × 1, and the third layer convolution kernel size of 100 × 1. The weights of these convolution kernels need to be randomly initialized.

[0065] b. Bias term: The bias term corresponding to each convolutional layer is usually initialized to 0 or a small random value.

[0066] (II) Parameters of the fully connected layer:

[0067] a. Weight Matrix: The weight matrix of the fully connected layer in the Deep Separable Multilayer Perceptron (D-MLP), which is used to map the input features to higher or lower dimensions.

[0068] b. Bias term: The bias term corresponding to the fully connected layer, usually initialized to 0 or a small random value.

[0069] (III) Parameters of the multi-head attention mechanism:

[0070] a. Projection weight matrix: The weight matrix used to linearly map the output of the multi-head attention and optimize the parameters of the residual connection.

[0071] Step 22: Multi-scale feature extraction module (feature extraction process 1):

[0072] First, a convolution operation is performed: the input training set data passes through a multi-scale convolution extraction module, and the convolution kernel slides on the input signal to extract local features at different positions. Any of the following convolution layers is a naive CNN, but through the careful design of three layers of convolution, the present invention extracts vibration features of various levels of gearbox vibration as much as possible.

[0073] Among them, the first layer of convolution: the convolution kernel size is 300×1, which extracts the overall general features of the vibration signal, such as vibration amplitude and frequency offset. The second layer of convolution: the convolution kernel size is 200×1, focusing on the changing characteristics in the vibration signal, such as harmonics and resonance phenomena within a specific frequency range. The third layer of convolution: the convolution kernel size is 100×1, focusing on the detailed features of the vibration signal, such as instantaneous vibration value and time series pattern of the vibration signal.

[0074] Finally, after three layers of convolution, the feature map is generated, and the input of "Step 23" has been generated.

[0075] The specific process of this step is as follows: the actual working environment of the gear is harsh and the interference information is complex. Since CNN has the advantages of parameter sharing, sparse connection, translation invariance and hierarchical feature extraction, the present invention designs a multi-scale feature extraction module based on CNN. The module aims to extract local fault feature information. The specific structure is as follows Figure 3 shown.

[0076] The convolution layer in this module extracts local feature information at different positions in the input signal through a sliding window, where the feature of each position is obtained by convolution operation between the convolution kernel and the vibration point in the corresponding window.

[0077] Since the signal processed by the present invention is a one-dimensional vibration signal, the convolution form is a one-dimensional convolution. The size of the first-layer convolution kernel is 300×1. The wide convolution can not only expand the receptive field of the input signal and cover more contextual information of the input signal, but also capture the overall general characteristics of the entire vibration signal, such as vibration amplitude, frequency offset, etc. In addition, the larger the width of the convolution kernel, the smaller the width of its feature map, which can effectively reduce the computational complexity and parameter quantity of the model. The size of the second-layer convolution kernel is 200×1, and the convolution mainly focuses on the changing characteristics in the vibration signal, such as harmonics and resonance phenomena within a specific frequency range. The size of the third-layer convolution kernel is 100×1, and the convolution mainly focuses on the detailed characteristics of the vibration signal, such as instantaneous vibration values, time series patterns of vibration signals, etc. This module uses the working mode of cooperating with convolutions of different scales to not only capture features of different ranges, but also help the network better deal with noise and interference in the data and reduce the impact of errors. After three layers of convolution, a feature map is generated, which includes features such as vibration amplitude, frequency offset, harmonics and resonance phenomena within a specific frequency range, instantaneous vibration value, and time series pattern of vibration signals. These feature map sequences are generated as inputs of step 23.

[0078] Step 23: Lightweight feature fusion ViT (feature extraction process 2):

[0079] In order to capture the feature representation of different spaces, the multi-head attention mechanism combines the features of different subspaces extracted by multiple single-head attention mechanisms. Using the multi-head attention mechanism, the input feature map sequence is divided into multiple subsequences, and each subsequence is linearly projected and mapped to different subspaces. The attention weight is calculated independently in each subspace, and the multi-head attention output feature vector is obtained by weighted summation. At this point, the input of step 24 has been generated.

[0080] In this step, residual connections are introduced before and after each single-head attention mechanism. Residual connections are basic operations that preserve the original feature information and facilitate the transfer and reuse of features.

[0081] The specific process of lightweight feature fusion ViT in this step is as follows: after the multi-scale feature extraction module extracts local features in the gear signal, lightweight feature fusion ViT in this step captures the global dependency of the entire vibration sequence data. This module not only has a very low number of feature parameters, but also fuses features from different layers to increase feature complementarity in different spaces at each layer.

[0082] In order to capture the feature representations of different spaces, the multi-head attention mechanism combines the features of different subspaces extracted by multiple single-head attention mechanisms. First, the input feature map sequence is divided into multiple subsequences, and each subsequence is linearly projected and mapped to different subspaces to generate representations of multiple subsequences. Then, in each subspace, the attention weight is calculated independently, which includes calculating the projections of the query, key, and value, as shown in formulas (1) to (4). Finally, the output of the multi-head attention is obtained by weighted summing the values ​​obtained by the attention calculation of each subspace, as shown in formula (5). Figure 4 It is the structure of the multi-head self-attention mechanism.

[0083] Q=X f W Q (1)

[0084] K=X f W K (2)

[0085] V=X f W V (3)

[0086]

[0087] MultiHead(Q,K,V)=Concat(head1,head2,…,head n )W (5)

[0088] Among them: Concat represents the connection operation, W represents the weight matrix of the mapping. Q, K, V represent the query, key, and value matrices respectively. These matrices are obtained by multiplying the feature matrix X and the learnable matrix W. d represents the dimension of the matrix, X f represents the input matrix of the multi-head attention mechanism, and head1 represents the output matrix of a single head in the multi-head attention mechanism.

[0089] In order to enhance the performance of the model, the module introduces residual connections before and after each single-head attention mechanism, which allows the original input to be passed directly through the network to subsequent layers instead of undergoing multiple nonlinear transformations, helping to retain more original feature information and promote the transfer and reuse of features. In addition, the introduction of residual connections adds the input to the output, which allows the gradient to flow directly to the shallower layers, making it easier to backpropagate, helping to reduce the gradual disappearance of gradients during training and making the network easier to converge.

[0090] Step 24: Depthwise Separable Multilayer Perceptron D-MLP (Feature Extraction Process 3):

[0091] This step focuses on extracting high-dimensional features. First, the input feature vector is mapped to a higher-dimensional space through a fully connected layer to capture the complex nonlinear patterns in the gearbox vibration signal, and the number of model parameters is reduced by point-by-point convolution. Use GELU activation function: The GELU function is used to enhance the convergence and stability of the network and prevent overfitting.

[0092] The specific process of the deep separable multilayer perceptron in this step is as follows: Since each neuron in the MLP is connected to all neurons in the previous layer, the number of parameters grows quadratically, which leads to a huge number of characteristic parameters of the model and increases the computational complexity. To address the above problems, the present invention adopts D-MLP, which consists of two fully connected layers, deep convolution, point-by-point convolution and GELU function, such as Figure 5 shown.

[0093] First, D-MLP maps the multi-head attention output feature vector to a higher-dimensional space through the first fully connected layer according to matrix multiplication and bias addition. This operation expands the dimension of the input feature vector, so that the model can better capture the complex features of the input data. Next, D-MLP uses the second fully connected layer to map the output of the previous layer to a lower-dimensional space again, reducing the dimension of the feature vector, thereby extracting more abstract and useful feature information. Then, D-MLP uses deep convolution and point-by-point convolution step calculation to significantly reduce the number of model parameters. Finally, D-MLP uses the GELU function to enhance the convergence and stability of the network and prevent overfitting. The function is shown in formula (6):

[0094]

[0095] in: x represents the input information of the MLP module; tanh represents the hyperbolic tangent function. The parameters of the GELU function are optimized through back propagation, and the output of the GELU function is a complex feature.

[0096] Step 25: Classification layer (result output process): The extracted features (features extracted in steps 22, 23, and 24) are mapped to fault categories through the classification layer, and the fault diagnosis results are output. The classification layer of this step can use a Softmax classifier to convert the nonlinear scores of the output features from steps 22 to 24 into a probability distribution, where each category corresponds to a probability value, and the sum of the probability values ​​is 1.

[0097] Step 26: Use the back propagation optimization algorithm to optimize the parameters of the convolutional layer, fully connected layer, and multi-head attention mechanism in step 21. The back propagation algorithm optimizes the weights and biases of the neural network by calculating the gradient of the loss function with respect to the network parameters.

[0098] Step 3: Determine whether the model iteration batch is greater than the preset value. If it is, save the model parameters. If it is less than or equal to it, return to step 2 to continue training. Generally, the model iteration batch range is between 100 and 200 times.

[0099] Step 4: Send the test set data to the trained model to test the model, that is, the model (parameters) output in step 3 is used to calculate the test set, and the classification layer outputs the fault diagnosis results.

[0100] The multi-scale convolution extraction module of the present invention respectively extracts the overall features, change features and detail features of the vibration signal through convolution kernels of different sizes (300×1, 200×1, 100×1). This multi-scale convolution design can better capture features of different ranges, enhance the robustness of the model to noise and interference, and reduce the impact of errors. The present invention introduces lightweight feature fusion ViT, utilizes a multi-head attention mechanism to fuse features of different subspaces, and retains the original feature information through residual connections. After the improvement, the amount of feature parameters is reduced, and the model's ability to capture global dependencies is enhanced, thereby improving the performance of the model. The deep separable multi-layer perceptron (D-MLP) of the present invention greatly reduces the amount of model parameters through deep convolution and point-by-point convolution, and enhances the convergence and stability of the network through the GELU function. The present invention reduces the computational complexity of the model while improving the generalization ability and stability of the model.

[0101] The experiment uses the Southeast University gearbox dataset. The test bench is mainly composed of a motor, a motor controller, a planetary gearbox, a reduction gearbox, and a load controller. Figure 6 The transmission system dynamic simulator (DDS) shown in the figure collects a total of 8 channel signals with a sampling frequency of 5120Hz. The gears with different fault states are processed in advance. The speed change can be achieved through the motor controller, and the load change can be achieved through the load controller.

[0102] In the operation of the gearbox, the number of fault samples is uncertain, so the diagnostic model must be able to effectively complete fault identification with more or less samples. However, when the number of samples is too large, the model training time and memory consumption will increase. Therefore, the diagnostic model needs to complete the fault identification work with as few samples as possible. Therefore, in order to verify the fault diagnosis performance of the model of the present invention in a small sample environment, a small sample experiment is designed. And each experiment repeats the sample selection 5 times to reduce the deviation of the random selection of small training sets, and the average of the results of the 5 experiments is taken to ensure the rationality of the fault results. The experimental results are shown in Table 1. It can be seen that with the increase in the number of training samples, the accuracy of the proposed model on the test set gradually increases. When the samples are 100 and 120, the accuracy is 93.24% and 96.72%, and the number of samples increases by 20, and the accuracy increases by 3.48%.

[0103] Finally, to verify the superiority of the proposed model, a comparative experiment was conducted with the comparison model, and the results are shown in Table 3-6. It can be seen that when the number of samples is 200, the accuracy of the comparison model is 90.16%, 91.41%, 96.66% and 97.24%, respectively, while the accuracy of the model of the present invention is 99.90%. This shows that the model of the present invention can still complete the task of fault diagnosis in an environment with few samples.

[0104] Table 1 Fault diagnosis accuracy under different samples

[0105]

[0106] The present invention combines MCNN and feature fusion ViT, extracts data features at multiple scales and combines global information, and reduces the number of parameters through deep separable convolution.

Claims

1. A gearbox fault diagnosis method based on MCNN and feature fusion ViT, characterized in that: The following steps are involved: Step 1: Normalize the one-dimensional vibration data of the gearbox collected by the sensor and divide it into a training set and a test set; Step 2: Initialize the model parameters, and send the training set data into the model to extract fault features, and use the back propagation optimization algorithm to optimize the model parameters; Step 21: Initialize the parameters of the convolutional layer, fully connected layer, and multi-head attention mechanism; Step 22: Multi-scale feature extraction module: The input training set data passes through the multi-scale convolution extraction module. The convolution kernel slides on the input signal to extract local features at different positions. The first layer of convolution: the convolution kernel size is 300×1, which extracts the overall general features of the vibration signal; The second convolution layer: the convolution kernel size is 200×1, focusing on the changing characteristics of the vibration signal; The third convolution layer: the convolution kernel size is 100×1, focusing on the detailed features of the vibration signal; Finally, after three layers of convolution, a feature map sequence is generated; Step 23: Lightweight feature fusion ViT: The input feature map sequence is divided into multiple subsequences, and each subsequence is linearly projected and mapped to different subspaces; The attention weights are calculated independently in each subspace, and residual connections are introduced before and after each single-head attention mechanism; the multi-head attention output feature vector is obtained by weighted summation; Step 24: Deep Separable Multilayer Perceptron D-MLP: Map the input feature vector to a higher dimensional space through a fully connected layer, capture the complex nonlinear patterns in the gearbox vibration signal, and use point-by-point convolution to reduce the number of model parameters; Use GELU activation function: Enhance the convergence and stability of the network through the GELU function; Step 25: Classification layer: The extracted features are mapped to fault categories through the classification layer, and the fault diagnosis results are output; Step 26: Optimize the parameters using the back propagation optimization algorithm; Step 3: Determine whether the model iteration batch is greater than the preset value. If it is greater, save the model parameters. If it is less than or equal to it, return to step 2 to continue training. Step 4: Send the test set data into the trained model to test the model, and the classification layer outputs the fault diagnosis results.

2. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: In step 1, 90% of the data is the training set and 10% of the data is the test set.

3. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: In step 21, the parameters of the convolution layer include the convolution kernel weight and the bias item corresponding to each convolution layer; the parameters of the fully connected layer include the weight matrix and the weight matrix of the fully connected layer; the parameters of the multi-head attention mechanism are the projection weight matrix and the parameters of the optimized residual connection.

4. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: The specific process of step 23 is: dividing the input feature map sequence into multiple subsequences, and linearly projecting each subsequence to map it into different subspaces to generate representations of multiple subsequences; In each subspace, the attention weight is calculated independently by formula (1) to formula (4), The weighted sum of the values ​​calculated for the attention of each subspace is performed, and the output of the multi-head attention is obtained by formula (5): Q=X f W Q (1) K=X f W K (2) V=X f W V (3) MultiHead(Q,K,V)=Concat(head1,head2,…,head n )W (5) Where: Concat represents the connection operation, W represents the weight matrix of the mapping; Q, K, V represent the query, key, and value matrices respectively, X represents the feature matrix, W represents the learnable matrix, and d represents the dimension of the matrix.

5. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: The specific process of step 24 is as follows: first, the separable multi-layer perceptron D-MLP maps the multi-head attention output feature vector to a higher-dimensional space through the first fully connected layer according to matrix multiplication and bias term addition, and then uses the second fully connected layer to map the output of the previous layer to a lower-dimensional space again, reducing the dimension of the feature vector, and using deep convolution and point-by-point convolution step-by-step calculation to greatly reduce the number of model parameters; finally, the GELU function is used to enhance the convergence and stability of the network. The formula (6) of the GELU function is: Among them: x represents the input information of the MLP module; tanh represents the hyperbolic tangent function.

6. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: In step 25, the classification layer uses a Softmax classifier to convert the output feature nonlinear score into a probability distribution, with each category corresponding to a probability value.

7. The gearbox fault diagnosis method based on MCNN and feature fusion ViT according to claim 1 is characterized in that: The model iteration batch range in step 3 is between 100 and 200 times.

Images