Rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model

Abstract

The invention discloses a rolling bearing fault diagnosis method based on an SE-ResNeXt and Informer hybrid network model, and relates to the field of industrial process control, spatial features of time series data are extracted through a plurality of convolution modules in the SE-ResNeXt, an SE module is added for adaptively adjusting the response intensity of each channel, and a fault diagnosis result is obtained. The design of a residual network is combined, so that the feature extraction process is more efficient, and fault diagnosis analysis is performed through an Informer model. In the training process of the model, cross entropy is adopted as a loss function, and optimization is carried out through an Adam optimizer, so that the prediction accuracy and stability are improved. Experimental results show that the method shows excellent fault diagnosis capability on a public data set CWRU, has good robustness and prediction precision, is significantly superior to a traditional deep learning model, and provides a new thought for state monitoring and fault prediction under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the field of industrial process control, and in particular to a rolling bearing fault diagnosis method based on a SE-ResNeXt and Informer hybrid network model. Background Art

[0002] As an important component of rotating machinery, the operating state of rolling bearings directly affects the reliability and life of the equipment. However, complex working conditions and high-load operation make rolling bearings prone to failure. Traditional fault diagnosis methods rely on signal processing and expert experience, which are difficult to deal with non-stationary signals or complex working conditions, and the feature extraction and classification effects are limited. The introduction of deep learning technology provides a new solution for fault diagnosis. Convolutional neural networks (CNNs) perform well in local feature extraction, but their pooling operations are prone to information loss and insufficient global feature modeling; although recurrent neural networks (RNNs) are suitable for time series modeling, they have problems with gradient vanishing and low computational efficiency when processing long sequences. Therefore, there is an urgent need for a prediction model that combines local feature extraction and global dependency modeling capabilities, which can efficiently extract multi-scale features under complex working conditions and has strong global semantic modeling capabilities to improve the accuracy and robustness of fault pattern recognition. Summary of the invention

[0003] The present invention provides a rolling bearing fault diagnosis method based on a SE-ResNeXt and Informer hybrid network model. By introducing the SE-ResNeXt network structure to process the time series data of the device, it can more efficiently fuse multi-branch features, fuse the SE-ResNeXt and Informer hybrid network model, combine local feature extraction with global dependency modeling, and thus realize the prediction of the fault category of the rolling bearing. The present invention is achieved through the following technical solutions: A rolling bearing fault diagnosis method based on a SE-ResNeXt and Informer hybrid network model comprises the following steps: 1. Data collection and preprocessing: Rolling bearing fault data set, each data file in the data contains DE (drive end accelerometer data), FE (fan end accelerometer data), BA (basic accelerometer data), time (time series data) and speed. The data is resampled to a uniform length and delayed sampling is used for data enhancement.

[0004] Perform fast Fourier transform (FFT) on the input one-dimensional time domain signal (length 2048) to generate a frequency domain signal. Use a sliding window to split the original time domain signal with a window length of 256 and a step size of 128 to obtain 16 sub-signal windows. Perform FFT on each sub-signal and extract the amplitude spectrum. Combine the amplitude spectra of all windows into a two-dimensional spectrum image, and finally generate a 16×256 spectrum matrix.

[0005] A 3x3 convolutional layer is used to extract preliminary features as the input of the SE-ResNeXt network.

[0006] 2. Model design: Construct a hybrid network model based on SE-ResNeXt and Informer. The model mainly includes the following modules: SE-ResNeXt feature extraction module: Based on the group convolution structure of the ResNeXt network, each convolution block uses a combination of 1×1 convolution (dimensionality reduction), 3×3 convolution (feature extraction), and 1×1 convolution (dimensionality recovery) to extract spatial features of time series data. It can effectively extract and enhance key features in time series data. Combined with the design of the residual network, the feature extraction process is more efficient; Informer time series modeling module: It is good at capturing long-distance dependencies in time series data and is suitable for time series data analysis. It introduces the sparse attention mechanism (ProbSparse Self-Attention) to reduce the computational complexity of traditional self-attention. It uses the attention distillation mechanism to downsample multi-head attention to reduce spatial complexity.

[0007] Output classification module: Use a fully connected layer to map the classification label (class token) output by the encoder to the number of fault categories num_classes for classification. Apply the Softmax function to this output to obtain the probability distribution of the fault category corresponding to each sample.

[0008] 3. Model training: The preprocessed data is input into the designed hybrid model, and the cross entropy is used as the loss function, which is defined as:

[0009] in, is the true label of the i-th sample, is the predicted value of the ith sample.

[0010] The model parameters were optimized using the Adam optimizer, with an initial learning rate of 0.001, which was dynamically adjusted using a learning rate decay strategy. The batch size was set to 32. During the model training process, the loss function was reduced through multiple iterations until the loss function converged, ensuring that the model could stably and accurately diagnose the rolling bearing fault category. The data was divided into a training set (70%), a validation set (15%), and a test set (15%). The training was repeated for 100 epochs, and the accuracy of the validation set was evaluated after each epoch, and the best performing model was saved.

[0011] 4. Fault category diagnosis and output: The trained model is used to predict the sensor data. By extracting the spatial and temporal features in the bearing vibration signal, the rolling bearing fault category can be accurately predicted. The prediction results are verified by a variety of evaluation indicators (such as average loss (AvgLoss), average recognition accuracy (AvgAcc)) to ensure the stability and accuracy of the model performance.

[0012] Preferred embodiment: ① The ResNeXt Block in SE-ResNeXt uses grouped convolution within each residual block to enhance the representation capability of the network without increasing the depth or width of the neural network. Grouped convolution improves computational efficiency and parameter sharing by splitting the convolution operation into multiple smaller groups. Each grouped convolution has its own independent weight, and the outputs of these groups are aggregated to form the final output feature map.

[0013] ②The SE module in SE-ResNeXt is divided into two parts: squeeze compression and excitation excitation. It aims to enhance the feature representation ability of the network by adaptively adjusting the weights of the channels. Squeeze compresses the spatial information of each channel by performing global average pooling on each channel, compresses the feature space of each channel into a scalar, and extracts the global information of each channel; excitation performs an attention mechanism on the channel dimension and generates a weight coefficient for each channel through a fully connected layer. This weight coefficient is used to dynamically adjust the importance of each channel, thereby improving the expressiveness of the model.

[0014] ③The Informer module proposes the ProbSparse autocorrelation mechanism, which successfully improves the prediction ability in the long sequence time series prediction (LSTF) problem; it proposes the autocorrelation distillation operation, which strengthens the focus on important features on the stacked layer and significantly reduces the spatial complexity, so that the model can handle long sequence inputs. ④The model's loss function uses cross entropy to ensure stability during training and improve the accuracy of prediction results.

[0015] Compared with the prior art, the present invention has the following technical advantages: ① The accuracy and robustness of bearing fault category prediction have been improved, and the spatial and temporal features of rolling bearing vibration signals can be effectively extracted; ② By combining SE-ResNeXt and Informer, the spatial and temporal features in the rolling bearing vibration signal can be effectively extracted. Especially when processing long time series data, the self-attention mechanism of Informer can model the global temporal dependency, thereby significantly improving the accuracy of fault diagnosis; ③The sparse attention mechanism is adopted, which can significantly reduce the computational complexity when processing long time series data while retaining powerful global dependency modeling capabilities.

[0016] ④ The experimental results show that the present invention can demonstrate strong generalization ability on data sets of different working conditions and different fault types, avoids the occurrence of overfitting, and has better adaptability and stability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of the overall model network structure of the present invention; Figure 2 This is a schematic diagram of converting time domain data into frequency domain data through FFT in the present invention; Figure 3 This is the structure diagram of SE-ResNeXt of the present invention; Figure 4 Schematic diagram of group convolution of the present invention; Figure 5 This is the structural diagram of the SE module of the present invention; Figure 6 This is the structural diagram of the ProbSparse Self-Attention module of the present invention; Figure 7 This is a schematic diagram of a fully connected network of the present invention; Figure 8 This is a diagram of the prediction results of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below in conjunction with the accompanying drawings, which are intended to explain the present invention rather than to limit it.

[0019] See also Figure 1-8 , a rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model, comprising the following steps: 1. Data Acquisition and Preprocessing (Step S1) S1.1 In the process of rolling bearing fault category prediction, the rolling bearing fault dataset is first resampled to a uniform length and delayed sampling is used for data enhancement.

[0020] S1.2 performs fast Fourier transform (FFT) on the input one-dimensional time domain signal (length 2048) to generate a frequency domain signal. Use a sliding window to split the original time domain signal with a window length of 256 and a step size of 128 to obtain 16 sub-signal windows. Perform FFT on each sub-signal and extract the amplitude spectrum. Combine the amplitude spectra of all windows into a two-dimensional spectrum image, and finally generate a 16×256 spectrum matrix. The generated spectrum is shown in the figure below. Figure 2 shown.

[0021] A 3x3 convolutional layer is used to extract preliminary features, and the processed data is then split into training and test sets for subsequent model training and verification.

[0022] 2. Feature extraction module (step S2) The SE-ResNeXt network is used to extract spatial features of time series data, such as Figure 3 As shown. SE-ResNeXt consists of two main parts: S2.1 ResNeXt part: Grouped convolution is used in each residual block to enhance the representation ability of the network. Each convolution layer extracts local features through a combination of 1x1 convolution and 3x3 convolution to extract information of different scales. Grouped convolution is as follows Figure 4 shown.

[0023] S2.2SE part: The SE module is divided into two parts: squeeze compression and excitation stimulation, which aims to enhance the feature representation ability of the network by adaptively adjusting the weights of the channels. Figure 5 shown.

[0024] The weight coefficient of each channel is generated through the fully connected layer and mapped to the range of [0, 1] through the Sigmoid function to adjust the feature response of each channel.

[0025] S2.3 Feature fusion part: The weights learned by the SE module are multiplied element-by-element with the output of the convolutional layer to adjust the feature value of each channel. The weighted channel features are added to the input features to form a residual connection.

[0026] 3. Timing modeling module (step S3) The time features processed by SE-ResNeXt are input into the Informer module. In order to effectively capture the interdependence of long sequence input and output, focus on key features and reduce computational complexity, the present invention adopts the Informer module. The Informer module uses the ProbSparse autocorrelation mechanism to significantly improve the prediction ability on the LSTF problem; through the autocorrelation distillation operation, it strengthens the focus on important features and greatly reduces the space complexity.

[0027] The S3.1 embedding layer maps low-dimensional data to a high-dimensional feature space through linear projection, which enhances the model's ability to express features. It introduces a learnable global vector classification token (Classtoken), adds the ClassToken to the first position of the sequence, and updates the sequence length. It introduces position information to capture the temporal relationship and adds the position information of each time step to the sequence.

[0028] The S3.2 encoder consists of a 3-layer stacked structure to gradually extract deep features. ProbSparseSelf-Attention is used in each layer, such as Figure 6 As shown in the figure. The sparse attention mechanism is used to reduce computational complexity and focus on extracting key time-step features. Each layer of the encoder halves the length of the sequence, uses max pooling to downsample the sequence, and extracts key features. The first time step, which is the output corresponding to the class token, is extracted from the encoder output.

[0029] 4. Classification module (step S4) The feature data processed by Informer is input into the classification layer, which is composed of a fully connected network, such as Figure 7 As shown in the figure, the classification token output by the encoder is mapped to the number of fault categories (num_classes) through the fully connected layer and classified. The Softmax function is applied to the output to obtain the probability distribution of each category, thereby realizing fault diagnosis.

[0030] 5. Model training and optimization (step S5) The preprocessed data is input into the hybrid network model for training, and the cross-entropy loss is used as the loss function. The formula of the loss function is as follows:

[0031] in, represents the i-th element of the true label, It represents the model prediction The probability of belonging to the i-th category.

[0032] S5.1 Parameter Optimization The model is optimized using the Adam optimizer. The learning rate is set to 0.001 and the learning rate decay strategy is used to dynamically adjust the learning rate.

[0033] 6. Working phase (step S6) During the operation of the equipment, the real-time collected data is input into the trained hybrid network model through the preprocessing module for prediction.

[0034] S6.1 Data input and processing The model combines the SE-ResNeXt feature extraction module, the Informer time series modeling module and the classification module to analyze the equipment operation data and output the fault category prediction value of the equipment.

[0035] 7. Model evaluation (step S7) By comparing the predicted values ​​with the actual collected data, the performance of the model is evaluated using indicators such as average loss (AvgLoss) and average recognition accuracy (AvgAcc) to further optimize the model.

[0036] S7.1 Model performance evaluation The performance indicator formulas of AvgLoss and AvgAcc are as follows:

[0037]

[0038] Where N is the total number of samples, n is the number of fault categories, is the true label of the i-th sample, is the predicted value of the ith sample, is the loss function of the i-th sample, is the indicator function, , the value is 1, otherwise it is 0.

[0039] These evaluation results are used to retrain and adjust parameters of the model to improve the accuracy and stability of its predictions.

[0040] 8. Overview of the complete process (step S8) The operation process of the rolling bearing fault diagnosis method based on the SE-ResNeXt and Informer hybrid network model is as follows: S8.1 Startup phase: Initialize the equipment and model, and start data collection.

[0041] S8.2 Data collection and preprocessing: Collect equipment operation data in real time and perform resampling processing.

[0042] S8.3 Model prediction: The preprocessed data passes through the SE-ResNeXt feature extraction module, the Informer time series modeling module and the classification module in turn, and finally obtains the predicted bearing fault category.

[0043] S8.4 Real-time monitoring and adjustment: During the operation of the equipment, the bearing fault category is predicted in real time, the equipment status is monitored, and the accuracy of the prediction is ensured.

[0044] S8.5 Shutdown and Evaluation: After the equipment is shut down, evaluate the model performance and make optimization adjustments.

[0045] The present invention relates to a rolling bearing fault diagnosis method based on a SE-ResNeXt and Informer hybrid network model, and in particular to a technology for rolling bearing fault diagnosis. Combining the Squeeze-and-Excitation (SE) module and the ResNeXt network structure, as well as the Informer model, an efficient bearing fault category diagnosis method is proposed. Through the SE-ResNeXt network, the spatial features and temporal features in the rolling bearing vibration signal can be extracted more efficiently, and the long temporal dependencies can be effectively captured through the self-attention mechanism of the Informer, thereby achieving high-precision classification of different types of faults. The method aims to accurately identify and classify different types of faults through deep learning analysis of rolling bearing vibration signals, provide support for predictive maintenance of industrial equipment, and is particularly suitable for the field of bearing condition monitoring and fault diagnosis based on sensor data.

[0046] In addition, the invention uses cross-entropy loss as the loss function during model training, and optimizes model parameters through the Adam optimizer, so that the model can stably and accurately predict the bearing fault category. By combining the advantages of SE-ResNeXt and Informer, it effectively addresses the challenge of insufficient processing of time-varying and nonlinear characteristics of time series data in bearing fault category diagnosis, and significantly improves the prediction accuracy and stability of bearing fault categories.

[0047] Through this multi-module combined deep learning model, when processing large-scale data and coping with complex industrial production environments, it has high prediction accuracy and robustness, can achieve fast and accurate prediction of rolling bearing fault categories, and effectively ensure the quality stability of mechanical processing. The above content is only to illustrate the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model, characterized in that: The following steps are involved: Data collection and preprocessing: rolling bearing fault data set, the data includes ten types of bearing faults, each data file contains DE (drive end accelerometer data), FE (fan end accelerometer data), BA (basic accelerometer data), time (time series data) and speed; resample the data to a uniform length; The input one-dimensional time domain signal is converted into a frequency domain signal through fast Fourier transform (FFT), and the original signal is segmented using a sliding window to generate a spectrum image; Feature extraction module: Extract spatial features from spectrum images based on the SE-ResNeXt network structure, enhance key channel features with the Squeeze-and-Excitation (SE) module, and extract local features through grouped convolution to extract feature information at different scales; Time series modeling module: Based on the Informer network, it uses the self-attention mechanism and sparse attention (ProbSparseSelf-Attention) to reduce the computational complexity while capturing long-term time series dependency features; Output classification module: classifies fault categories through the fully connected layer and Softmax function.

2. A rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The data is converted from time series data to frequency domain signals via Fast Fourier Transform (FFT): The original time domain signal is divided into multiple windows using a sliding window. FFT is performed on each sub-signal, and the amplitude spectrum is extracted. The amplitude spectra of all windows are merged to generate a two-dimensional spectrum image as the input of SE-ResNeXt.

3. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1 is characterized in that: The data is transferred in the feature extraction module SE-ResNeXt through the following process: Grouped convolution: The preprocessed data enters the grouped convolution layer, where each convolution block contains 1×1 convolution (for dimensionality reduction), 3×3 convolution (for feature extraction), and 1×1 convolution (for dimension recovery) to extract local features; SE module: A Squeeze-and-Excitation (SE) module is added to each convolution block to perform global average pooling on each channel, extract the global information of each channel, generate the weight coefficient of each channel through the fully connected layer, and adjust the feature response of each channel through the Sigmoid function, thereby improving the expressiveness of the model; Feature fusion: The weights learned by the SE module are multiplied element-by-element with the output of the convolutional layer to adjust the feature values ​​of each channel; the weighted channel features are added to the input features to form a residual connection.

4. A rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 3, characterized in that: The transfer function of the data in the ResNeXt module is as follows: Grouped convolution of ResNeXt module: ； Where g is the number of groups, each group convolution is calculated independently, and the output feature map of all group convolutions is obtained after splicing operation; The output channel is represented by C, and the global average pooling result of each channel is: ； in is the eigenvalue of the i-th and j-th positions on channel c, and H and W are the height and width of the feature map.

5. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 3 is characterized in that: The transfer function of the data in the SE module is as follows: After two fully connected layers, the channel weights are calculated to determine the number of output channels: ； The weighting process can be expressed as: ； final Output characteristics: 。 6. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 3 is characterized in that: The data adjustment method after feature fusion is as follows: The Flatten operation combines the time dimension H and the frequency dimension W of each sample, flattens each matrix into a vector, and each channel becomes a vector; Use a fully connected layer to map the flattened features to the output layer; The output features of the SE-ResNeXt module will be used as the input of the Informer module.

7. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1 is characterized in that: The features of the SE-ResNeXt feature extraction module are input into the Informer time series modeling module. The specific process is as follows: The input SE-ResNeXt features are linearly projected into a high-dimensional feature space through the embedding layer, and the class token and position information encoding are added to form an input sequence; Use ProbSparse Self-Attention to model long-term dependencies of input features; The data is downsampled through the attention distillation mechanism, and the sequence length is halved layer by layer, which reduces the space complexity and memory usage; The encoder consists of 3 layers of stacked ProbSparse Self-Attention modules, which perform sparse attention modeling on the input features layer by layer.

8. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1 is characterized in that: The ProbSparse Self-Attention method in the Informer module is as follows: ProbSparse Self-Attention is an improved attention mechanism proposed in the Informer model. It focuses on key points through a sparse strategy and reduces computational complexity. The core idea is to select a small number of query vectors with large amounts of information for calculation. The importance score of each query is measured by the norm of the dot product result between it and all keys.

9. A rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 8, characterized in that: The transfer function formula of ProbSparse Self-Attention is as follows: The importance score of the query vector is calculated using the following formula: ； in Represents the p-norm, the commonly used p value is 2, that is, the L2 norm; Represents the dot product calculation between the i-th query (Query) and the j-th key (Key); According to the importance score of each query, select the top U queries: ； Here , represents the query vector set after sparse selection and satisfies U L; Selected query Corresponding to all keys = Sum = Keep the original dimension [L, d] and still interact with the global data; Only for filtered queries and all , Perform attention calculation: ； in, is the query after sparse selection, the attention score matrix A ; The computational complexity is from Reduce to , where only × where U represents the number of sparse query vectors and L is the sequence length.

10. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The sparse threshold of ProbSparse Self-Attention is calculated as follows: The sparse attention mechanism is used to reduce computational complexity and focus on feature extraction at key time steps; the number of attention heads is set to 8, and the sparse threshold is: ； Here the sparse factor ( ) is set to 10, L is 257, then m=25.

11. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1 is characterized in that: The data processed by the informer module is passed to the classification layer. The specific steps are as follows: Extract the first time step from the encoder output, which is the output corresponding to the class token; Use a fully connected layer to map the class token output by the encoder to the number of fault categories num_classes and perform classification; Applying a Softmax operation to this output yields a probability distribution for each class.

12. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The output of Informer is used as the final diagnostic output for rolling bearing fault category diagnosis, and cross entropy is used as the loss function, the formula is: ； in, is the true category label of the sample, is the predicted class probability of the sample.

13. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The activation function is formulated as: The ReLU activation function is used in the convolutional layer and feed-forward network, and the formula is: ； The Sigmoid activation function is used in the SE module to generate the channel weights. The formula is: ； In the final classification layer, the Softmax activation function is used for classification, and the formula is: 。 14. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The model uses the Adam optimizer, the learning rate is set to 0.001, the batch size is 32 during training, the learning rate decay strategy is used to dynamically adjust the learning rate, and the number of iterations is 100 to ensure stable convergence of the model.

15. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: The evaluation indicators of the model during the training process include average loss (AvgLoss) and average recognition accuracy (AvgAcc), which are used to evaluate the prediction effect of the model.

16. The rolling bearing fault diagnosis method based on SE-ResNeXt and Informer hybrid network model according to claim 1, characterized in that: After the model training is completed, the data is predicted and the predicted values ​​are compared and evaluated with the actual values.

Images