Grouping parallel attention anti-noise bearing fault diagnosis method

Abstract

An anti-noise bearing fault diagnosis method based on grouping parallel attention comprises the following steps: 1) acquiring acceleration signals in different bearing states, storing the signals in a data set, preprocessing data and dividing the data set; 2) building an anti-noise bearing fault diagnosis model based on grouping parallel attention, and training the model by using training set data; 3) performing an experiment by using test set data to inspect the performance of the model; and 4) carrying out actual bearing fault diagnosis. By introducing a deep learning method and proposing a grouping parallel attention mechanism, the bearing fault diagnosis method solves the problem that the diagnosis precision is reduced under a complex noise condition in the field of bearing intelligent fault diagnosis, and is high in robustness and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis in mechanical health management, and designs a group-based parallel attention anti-noise bearing fault diagnosis method. Background Art

[0002] In the modern industrial system, rotating machinery and equipment are widely used in various fields, from large-scale production lines in the manufacturing industry to power generation equipment in the energy industry, from vehicles in transportation to aircraft in aerospace, and are almost everywhere. Bearings, as the core components of rotating machinery and equipment, are responsible for supporting and guiding the rotation of the shaft, just like human joints, while bearing loads from all directions. Its performance directly affects the operating stability, accuracy and service life of the entire equipment. Once a bearing fails, due to its suddenness and concealment, it will cause serious losses to the entire equipment and even cause serious production safety accidents. At the same time, since the actual industrial production environment is very complex and there are often a large number of external interference factors, it is urgent to solve the problem of bearing fault diagnosis accuracy under complex external interference.

[0003] In order to effectively improve the fault diagnosis accuracy of complex external interference, a parallel attention anti-noise bearing fault diagnosis method is proposed. The core of the present invention is the group parallel attention mechanism. The whole method is as follows: First, the collected one-dimensional time series signal containing industrial noise is input into a wide convolution kernel for convolution. Studies have shown that wide kernel convolution can effectively suppress high-frequency noise and extract low-frequency global features. Then, the Gram denoising strategy is used to remove the noise anomalies mixed in the signal. Gram denoising can effectively suppress noise interference by using matrix transformation and operation. Then, feature information of different granularities is extracted through a dual-path feature extraction unit containing secondary convolution, and feature information is fused. Through group parallel attention, the group learning strategy can better weaken the secondary effects of noise and better capture the "hidden" information of important channels and key spatial positions. Finally, the final fault classification is completed through a bidirectional gated recurrent unit and a fully connected layer. This method can effectively suppress noise interference, fully tap key feature information, and achieve accurate diagnosis under complex noise conditions. Summary of the invention

[0004] The technical problem to be solved by the present invention is to propose a noise-resistant bearing fault diagnosis method based on group parallel attention, which solves the problems of noise interference, insufficient sample number and decreased accuracy caused by cross-domain diagnosis, and provides a new idea for the research on bearing fault diagnosis under complex noise conditions.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is: a method for bearing fault diagnosis based on group parallel attention and noise resistance, the steps of which are:

[0006] 1) Collect the acceleration signals in different bearing states, save the signals to the dataset, preprocess the data and divide the dataset.

[0007] 1.1) After collecting the bearing acceleration signals in different states, perform preprocessing operations on the data: normalize the data, using the maximum-minimum normalization, and the mathematical expression is as follows:

[0008]

[0009] In the formula: z is the output after normalization, x is the initial value of the data, μ is the mean value of the data, and σ is the standard deviation.

[0010] 1.2) After data normalization, since the number of actual fault samples that can be collected is small, expand the number of samples through data augmentation methods, and use the sliding window method for data augmentation. The sliding window formula is as follows:

[0011]

[0012] In the formula: L is the total length of the signal, s is the step size, and n is the number of samples.

[0013] 1.3) Continue to divide the data according to the ratio of 2:1 for the training set and the test set. The training set is used for model training, and the test set is used for model performance verification.

[0014] 2) Build a noise-resistant bearing fault diagnosis model based on grouped parallel attention, and use the training set data to train the model.

[0015] 2.1) The dual-channel feature extraction unit first inputs the signal into a wide convolutional kernel for convolution, suppresses high-frequency noise interference through wide-kernel convolution, and captures low-frequency global features; then uses the Gram noise reduction strategy to perform preliminary noise reduction on the features after wide-kernel convolution, reduces the outliers in the feature map, transforms the two-dimensional matrix from a high-rank mode to a low-rank mode, and transforms the low-rank two-dimensional matrix into a one-dimensional vector; finally, the quadratic convolutional kernels of different sizes in the dual-channel feature extraction unit enable the model to extract features of different granularities; at the same time, the quadratic convolution sizes in each branch decrease in an arithmetic progression rule to avoid the loss of key feature information of intermediate granularities. The mathematical formula of the quadratic convolution is as follows:

[0016] σ(f(x))=σ((w T w r +b r )(x T w g +b g )+(x⊙x) T w b +b b )

[0017] where σ(.) is the activation function, and w r , w g and w b are weight matrices, and b r , b g and b b are biases.

[0018] In 2.1), the dual-channel feature extraction unit is as follows:

[0019] First, a wide kernel convolution is used to extract the global features of the input signal. At the same time, using a wide convolution kernel in the first layer can also play a role in suppressing high-frequency noise. Then, Gram noise reduction is used to perform preliminary noise reduction on the signal. The basic principle of GNR is as follows: Since the normal vibration signal contains periodic pulses, it has periodic self-similarity, and the rows of the corresponding matrix are relatively similar; the matrix converted from the noisy signal will break the low-rank mode, resulting in a higher rank of the matrix. Therefore, matrix operations are performed to reduce the influence of outliers and achieve the purpose of preliminary noise reduction.

[0020] After obtaining the global features and preliminary noise reduction, the model needs to capture features of different granularities under noise interference, adopting a dual-channel feature extraction strategy: the size of the convolution kernel for each channel decreases according to an arithmetic progression rule to ensure that feature loss does not occur due to excessive changes in the feature map; at the same time, the sizes of the convolution kernels for the two channels are different to capture features of different scales, and the learned feature information is more diverse. Finally, the features of the two channels are fused. In feature extraction, quadratic convolution is used instead of traditional convolution: quadratic convolution uses three groups of different weights and biases to introduce non-linearity in addition to the activation function, and can extract more complex features under external interference conditions.

[0021] 2.2) Through the grouped parallel attention mechanism, reduce the interference of external irrelevant information, and adopt a grouping strategy to parallelly use channel attention and spatial attention to extract the feature information on important channels and key positions.

[0022] In 2.2), the grouped parallel attention mechanism is as follows:

[0023] The grouped parallel attention adopts a grouped learning strategy to capture the features within the group, and the parallel execution of channel attention and spatial attention also speeds up the learning speed of the model; after grouped learning, in order to enable information interaction between different groups, the channel shuffle technology is introduced. The channel shuffle ensures feature sharing between different groups and can extract detailed global features;

[0024] First, assume that the size of the upper-layer input feature map is 1×h×c, that is, an input with a length of h and c channels. The c channels are evenly divided into g groups, and the number of channels in each group is c / g. Each group is further evenly divided into two sub-groups: Sub-group 1 uses channel attention to extract key channel features within the group and suppress external information interference; Sub-group 2 uses spatial attention to make the model pay more attention to key position information;

[0025] Sub-group 1 adopts a variant of the efficient channel attention. First, one branch performs global average pooling on the channels, and the channel mean value is obtained for each channel; then, a 3×1 convolutional kernel is used to perform convolution on the mean value, and information exchange is completed between each channel and its adjacent channels; finally, the Sigmoid function is used to generate channel attention weights; on the other branch, a 3×1 depth convolution is adopted: each convolutional kernel of the depth convolution only performs convolution on one input channel, so independent feature extraction for each channel can explore deep information within the channel, and finally the internal features of each channel are multiplied by the channel weights;

[0026] The variant of spatial attention is applied to Sub-group 2. On one branch, group normalization is performed: compared with batch normalization and layer normalization, it utilizes the local correlation between channels within the group and is not affected by the batch size; then, it passes through a linear transformation layer, and finally, the position weights are generated through an activation function; the other branch does not use depth convolution like Sub-group 1 because the importance of each spatial position is jointly determined by the information at the same position on multiple channels. Therefore, ordinary convolution operations are performed, and finally, the convolved feature map is multiplied by its respective position weights;

[0027] Finally, the output feature maps of Sub-group 1 and Sub-group 2 are concatenated, and then channel shuffling is performed among all groups to strengthen information sharing between different groups and jointly suppress noise interference.

[0028] 2.3) Each channel is regarded as a keyword through a bidirectional gated recurrent unit, so that each channel can fully exchange the feature information it extracts with its front and back channels. Finally, the final diagnostic result is obtained through a fully connected layer and a classifier.

[0029] 2.4) Model training is carried out. The initial learning rate is set to lr = 0.0012. The learning rate adjustment strategy selects the exponential decay adjustment strategy. The optimizer selects the Adam optimizer, and the loss function selects the cross-entropy loss function. Early stopping is used during model training, that is, when the loss no longer decreases, it is defaulted that the model converges.

[0030] 3) Use the test set data to verify the diagnostic performance of the model. Accuracy comparison experiments are carried out under three different noise and different signal-to-noise ratio conditions:

[0031] If the accuracy rate of the proposed model is generally higher than that of the comparative model under different types and intensities of noise conditions, it indicates that the proposed model has excellent noise suppression ability and can excellently complete the bearing fault diagnosis task in a noisy environment;

[0032] If the diagnosis accuracy rate of the proposed model is inferior to that of other comparative models, it indicates that the parameter settings in the training process of the proposed model are unreasonable, and it is necessary to adjust the hyperparameters or the internal parameters of the model: the hyperparameter adjustment measures include changing the initial learning rate, selecting other learning rate adjustment strategies, and introducing hyperparameter optimization algorithms, while the internal parameter adjustment measures of the model include selecting the number of groups and choosing activation functions.

[0033] 4) Import the data to be diagnosed into the anti-noise fault diagnosis model based on grouped parallel attention for fault diagnosis. Determine the specific corresponding bearing fault type according to the label type output by the model.

[0034] The beneficial effects of the present invention are as follows: through the dual-channel feature extraction unit, the proposed model not only effectively suppresses external interference in a complex noise environment, but also helps the model effectively extract signal features from both local and global levels; the proposed grouped parallel attention mechanism adopts a grouping strategy to enable the model to focus on enhancing the weights of important channels and key positions, reduce the attention to redundant information, and effectively mine the feature information "hidden" by noise. The overall model can handle various bearing fault diagnosis tasks under complex noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the one-dimensional time series signal diagram of the fault;

[0036] Figure 2 is the change diagram of the loss accuracy rate of model training;

[0037] Figure 3 is the framework diagram of the model in practical application;

[0038] Figure 4 is the structural diagram of the anti-noise diagnosis model with grouped parallel attention;

[0039] Figure 5 is the diagram of the comparative test results;

[0040] Figure 6 is the confusion matrix diagram of the classification results;

[0041] Figure 7 is the algorithm flow chart of the grouped parallel attention mechanism;

[0042] Figure 8 is the model hierarchical parameter diagram. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0044] A noise-resistant bearing fault diagnosis method based on grouped parallel attention, the steps of which are as follows:

[0045] A noise-resistant bearing fault diagnosis method based on grouped parallel attention, the steps of which are as follows:

[0046] 1) Collect the acceleration signals in different bearing states, save the signals to the dataset, preprocess the data, and perform dataset division.

[0047] 1.1) After collecting the bearing acceleration signals in different states, perform preprocessing operations on the data: normalize the data, use the maximum-minimum normalization, and the mathematical expression is as follows:

[0048]

[0049] In the formula: z is the output after normalization, x is the initial value of the data, μ is the data mean, and σ is the standard deviation.

[0050] 1.2) After data normalization, since the number of actual fault samples that can be collected is small, expand the number of samples by the method of data augmentation, and use the sliding window method for data augmentation. The sliding window formula is as follows:

[0051]

[0052] In the formula: L is the total length of the signal, s is the step size, and n is the number of samples.

[0053] 1.3) Continue to divide the data according to the ratio of the training set to the test set of 2:1. The training set is used for model training, and the test set is used for model performance verification.

[0054] 2) Build a noise-resistant bearing fault diagnosis model based on grouped parallel attention, and use the training set data to train the model.

[0055] 2.1) The dual-path feature extraction unit first inputs the signal into a wide convolutional kernel for convolution, suppressing high-frequency noise interference through wide-kernel convolution and capturing low-frequency global features; then, the Gram noise reduction strategy is used to perform preliminary noise reduction on the features after wide-kernel convolution, reducing outliers in the feature map. The transformed two-dimensional matrix is reduced from a high-rank mode to a low-rank mode, and the low-rank two-dimensional matrix is transformed into a one-dimensional vector; finally, the secondary convolutional kernels of different sizes in the dual-path feature extraction unit enable the model to extract features of different granularities; at the same time, the sizes of the secondary convolutions in each branch decrease according to an arithmetic progression rule, avoiding the loss of key feature information of intermediate granularities. The mathematical formula for the secondary convolution is as follows:

[0056] σ(f(x)) = σ((x T w r +b r )(x T w g +b g )+(x⊙x) T w b +b b )

[0057] where σ(.) is the activation function, w r , w g and w b are weight matrices, and b r , b g and b b are biases.

[0058] In 2.1), the dual-path feature extraction unit is as follows:

[0059] First, the wide-kernel convolution is used to extract the global features of the input signal. At the same time, using a wide convolutional kernel in the first layer can also play a role in suppressing high-frequency noise; then, the Gram noise reduction is used to perform preliminary noise reduction on the signal. The basic principle of GNR is as follows: Since the normal vibration signal contains periodic pulses, it has periodic self-similarity, and the rows of the corresponding matrix are relatively similar; the matrix converted from the noisy signal will break the low-rank mode, resulting in a higher rank of the matrix. Therefore, matrix operations are performed to reduce the influence of outliers and achieve the purpose of preliminary noise reduction;

[0060] After obtaining the global features and preliminary noise reduction, the model needs to capture features of different granularities under noise interference, adopting a dual-path feature extraction strategy: the sizes of the convolutional kernels in each path decrease according to an arithmetic progression rule to ensure that features are not lost due to excessive changes in the feature map; at the same time, the sizes of the convolutional kernels in the two paths are different, capturing features of different scales, and the learned feature information is more diverse. Finally, the features of the two paths are fused. In feature extraction, secondary convolution is used instead of traditional convolution: the secondary convolution introduces non-linearity other than the activation function by using three groups of different weights and biases, and can extract more complex features under external interference conditions.

[0061] 2.2) Reduce the interference of external irrelevant information through the grouped parallel attention mechanism. Adopt a grouping strategy to parallelly use channel attention and spatial attention to extract feature information on important channels and key positions.

[0062] In 2.2), the grouped parallel attention mechanism is as follows:

[0063] The grouped parallel attention adopts a grouped learning strategy to capture intra-group features, and the parallel execution of channel attention and spatial attention also speeds up the learning speed of the model; after grouped learning, in order to enable information interaction between different groups, a channel shuffle technique is introduced. The channel shuffle ensures feature sharing between different groups and can extract detailed global features;

[0064] First, assume that the size of the upper-layer input feature map is 1×h×c, that is, an input with a length of h and c channels. The c channels are evenly divided into g groups, and the number of channels in each group is c / g. Each group is further evenly divided into two subgroups inside: Subgroup 1 uses channel attention to extract key channel features within the group and suppress external information interference; Subgroup 2 uses spatial attention to make the model pay more attention to key position information;

[0065] Subgroup 1 adopts a variant of the efficient channel attention. First, one branch performs global average pooling on the channels, and the channel mean value is obtained for each channel; then, a convolutional kernel with a size of 3×1 is used to perform convolution on the mean value, and information exchange is completed between each channel and adjacent channels; finally, the channel attention weights are generated through the Sigmoid function; on the other branch, a depth convolution of 3×1 is adopted: each convolutional kernel of the depth convolution only performs convolution on one input channel, so independent feature extraction for each channel can discover deep information inside the channel. Finally, the internal features of each channel are multiplied by the channel weights;

[0066] The variant of spatial attention is applied to Subgroup 2. Group normalization is performed on one branch: compared with batch normalization and layer normalization, it utilizes the local correlation between channels within the group and is not affected by the batch size; then, through a linear transformation layer, and finally, position weights are generated through an activation function; the other branch does not use depth convolution like Subgroup 1 because the importance of each spatial position is jointly determined by the information at the same position on multiple channels. Therefore, ordinary convolution operations are performed, and finally, the convolved feature map is multiplied by its respective position weights;

[0067] Finally, the output feature maps of Subgroup 1 and Subgroup 2 are concatenated, and then channel shuffling is performed between all groups to strengthen information sharing between different groups and jointly suppress noise interference.

[0068] 2.3) Use a bidirectional gated recurrent unit to take each channel as a keyword, enabling each channel to fully communicate the feature information it extracts with its front and back channels. Finally, obtain the final diagnostic result via a fully connected layer and a classifier.

[0069] 2.4) Conduct model training. Set the initial learning rate to lr = 0.0012, select the exponential decay adjustment strategy for the learning rate adjustment strategy, choose the Adam optimizer as the optimizer, and select the cross-entropy loss function as the loss function. Use early stopping during model training, that is, by default, the model converges when the loss no longer decreases.

[0070] 3) Use the test set data to verify the diagnostic performance of the model. Conduct accuracy comparison experiments under three different noise and different signal-to-noise ratio conditions:

[0071] If the accuracy of the proposed model is generally higher than that of the comparison model under different types and intensities of noise conditions, it indicates that this model has excellent noise suppression ability and can excellently complete the bearing fault diagnosis task in a noisy environment;

[0072] If the diagnostic accuracy of the proposed model is inferior to that of other comparison models, it indicates that the parameter settings in the training process of this model are unreasonable, and it is necessary to adjust the hyperparameters or the internal parameters of the model: The hyperparameter adjustment measures include changing the initial learning rate, selecting other learning rate adjustment strategies, and introducing hyperparameter optimization algorithms, while the internal parameter adjustment measures of the model include selecting the number of groups and choosing activation functions.

[0073] 4) Import the data to be diagnosed into the anti-noise fault diagnosis model based on grouped parallel attention for fault diagnosis. Determine the specific corresponding bearing fault type according to the label type output by the model.

[0074] Specific application example:

[0075] Step 1: Use an acceleration sensor to collect vibration data signals, and select the sampling frequency of the digital signal to be 12000 Hz; then perform data augmentation on the data through a sliding window, set the sample length to 1024, and the window size to 400; use the mean standard deviation normalization method for normalization.

[0076] Step 2: Build an anti-noise bearing fault diagnosis model with grouped attention. The main hierarchical structure of this model mainly includes a preliminary noise reduction module: a dual-channel feature extraction unit (including wide-kernel convolution, Gram noise reduction kernel feature extraction), grouped parallel attention, a bidirectional gated recurrent unit, a fully connected layer, and a classifier. The changes in its model parameters are as Figure 8 shown:

[0077] Step 3: Set the model running environment and parameters

[0078] The model evaluation experiment was conducted in the Windows 11 operating system with a 12th Gen Intel(R) Core(TM) i7-12700H 2.30GHz processor and an NVIDIA GeForce RTX 3050 discrete graphics card; the initial learning rate was set to 0.0012, and an exponential decay adjustment strategy was adopted, with the cross-entropy loss function selected as the loss function.

[0079] Step 4: The training period was set to 100 epochs, and the model was trained using the early stopping strategy. After the model converged, the weight parameters of the model were saved for subsequent diagnosis.

[0080] Step 5: The obtained model was compared with the comparative model under different noise and different signal-to-noise ratio conditions, and the proposed model achieved excellent results. The results of the comparative experiment are as Figure 5 shown.

Claims

1. A method for diagnosing bearing faults based on grouped parallel attention and anti-noise, characterized in that The steps are as follows: 1) Collect the acceleration signals in different bearing states, save the signals to the dataset, perform data preprocessing and divide the dataset; 2) Build a noise-resistant bearing fault diagnosis model based on grouped parallel attention, and use the training set data to train the model; 2.1) The dual-channel feature extraction unit first inputs the signal into a wide convolutional kernel for convolution, suppresses high-frequency noise interference through wide-kernel convolution, and captures low-frequency global features; Then, the Gram noise reduction strategy is used to perform preliminary noise reduction on the features after wide-kernel convolution, reduce the outliers in the feature map, transform the two-dimensional matrix after transformation from a high-rank mode to a low-rank mode, and transform the low-rank two-dimensional matrix into a one-dimensional vector; finally, the quadratic convolutional kernels of different sizes in the dual-channel feature extraction unit enable the model to extract features of different granularities; at the same time, the quadratic convolution sizes in each branch decrease in an arithmetic progression rule to avoid the loss of key feature information in the intermediate granularity. The mathematical formula of the quadratic convolution is as follows: σ(f(x)) = σ((x T w r + b r )(x T w g + b g ) + (x ⊙ x) T w b + b b ) where σ(.) is the activation function, w r 、w g and w b are weight matrices, b r 、b g and b b are biases; 2.2) Through the grouped parallel attention mechanism, reduce the interference of external irrelevant information, and adopt a grouping strategy to parallelly use channel attention and spatial attention to extract feature information on important channels and key positions; 2.3) Use each channel as a keyword through a bidirectional gated recurrent unit, so that each channel can fully communicate the feature information extracted by itself with its front and back channels, and finally obtain the final diagnosis result through a fully connected layer and a classifier; 2.4) Conduct model training, set the initial learning rate to lr = 0.0012, select the exponential decay adjustment strategy for the learning rate adjustment strategy, select the Adam optimizer for the optimizer, select the cross-entropy loss function for the loss function, and use the early stopping method during model training, that is, when the loss no longer decreases, it is default that the model converges; 3) Use the test set data to verify the diagnostic performance of the model, and conduct an accuracy comparison experiment under three different noise and different signal-to-noise ratio conditions: If the accuracy of the proposed model is generally higher than that of the comparison model under different types and intensities of noise conditions, it indicates that the proposed model has excellent noise suppression ability and can excellently complete the bearing fault diagnosis task in a noisy environment; If the diagnostic accuracy of the proposed model is inferior to that of other comparison models, it indicates that the parameter settings in the training process of the proposed model are unreasonable, and it is necessary to adjust the hyperparameters or the internal parameters of the model: the hyperparameter adjustment measures include changing the initial learning rate, selecting other learning rate adjustment strategies, and introducing a hyperparameter optimization algorithm, while the internal parameter adjustment measures of the model include selecting the number of groups and choosing an activation function; 4) Import the data to be diagnosed into the noise-resistant fault diagnosis model based on grouped parallel attention for fault diagnosis. Judge the specific corresponding bearing fault type according to the label type output by the model.

2. The anti-noise bearing fault diagnosis method based on grouped parallel attention according to claim 1, wherein, In the above-mentioned 1), the specific method is: 1.1) After collecting the bearing acceleration signals in different states, perform preprocessing operations on the data: normalize the data, using the maximum-minimum normalization, and the mathematical expression is as follows: In the formula: z is the output after normalization, x is the initial value of the data, μ is the data mean, and σ is the standard deviation; 1.2) After data normalization, since the number of actual fault samples that can be collected is small, the number of samples is expanded by data augmentation. The sliding window method is used for data augmentation, and the sliding window formula is as follows: Where: L is the total length of the signal, s is the step size, and n is the number of samples; 1.3) The data is further divided according to the ratio of 2:1 for the training set and the test set. The training set is used for model training, and the test set is used for model performance verification.

3. The method for diagnosing bearing faults based on grouped parallel attention and anti-noise according to claim 1, wherein, In the above-mentioned 2.1), the dual-channel feature extraction unit is specifically as follows: First, wide kernel convolution is used to extract the global features of the input signal. At the same time, using a wide convolution kernel in the first layer can also play a role in suppressing high-frequency noise; then Gram noise reduction is used to perform preliminary noise reduction on the signal. The basic principle of GNR is as follows: Since the normal vibration signal contains periodic pulses, it has periodic self-similarity, and the rows of its corresponding matrix are relatively similar; the matrix converted from the noisy signal will break the low-rank pattern, resulting in a higher rank of the matrix. Therefore, matrix operations are performed to reduce the influence of outliers and achieve the purpose of preliminary noise reduction; After obtaining the global features and preliminary noise reduction, the model needs to capture features of different granularities under noise interference, and adopt a dual-channel feature extraction strategy: the size of the convolution kernel for each channel decreases according to the arithmetic progression rule to ensure that feature loss will not occur due to excessive changes in the feature map; at the same time, the sizes of the two-channel convolution kernels are different to capture features of different scales, and the learned feature information is more diverse. Finally, the features of the two channels are fused. In feature extraction, quadratic convolution is used instead of traditional convolution: quadratic convolution introduces non-linearity except for the activation function by using three groups of different weights and biases, and can extract more complex features under external interference conditions.

4. A method for diagnosing bearing faults based on grouped parallel attention and anti-noise, according to claim 1, characterized in that In the above-mentioned 2.2), the grouped parallel attention mechanism is specifically as follows: Grouped parallel attention adopts a grouped learning strategy to capture intra-group features, and the parallel execution of channel attention and spatial attention also speeds up the learning speed of the model; after grouped learning, in order to enable information interaction between different groups, the channel shuffle technique is introduced. Channel shuffle ensures feature sharing between different groups and can extract detailed global features; First, assume that the size of the input feature map of the upper layer is 1×h×c, that is, an input with a length of h and a channel number of c. The c channels are evenly divided into g groups, and the number of channels in each group is c / g. Each group is further evenly divided into two subgroups inside: Subgroup 1 uses channel attention to extract the key channel features within the group and suppress external information interference; Subgroup 2 uses spatial attention to make the model pay more attention to key position information; Subgroup 1 adopts a variant of efficient channel attention. First, one branch performs global average pooling on the channels, and the channel mean value is obtained for each channel; then a convolutional kernel with a size of 3×1 is used to perform convolution on the mean value, and information exchange is completed between each channel and its adjacent channels; finally, the channel attention weight is generated through the Sigmoid function; on the other branch, a 3×1 depth convolution is adopted: each convolutional kernel of the depth convolution only performs convolution on one input channel. Therefore, independent feature extraction for each channel can explore the deep information inside the channel, and finally the internal features of each channel are multiplied by the channel weight; A variant of spatial attention is applied to Group 2. Group normalization is performed on one branch: compared with batch normalization and layer normalization, it utilizes the local correlation among channels within a group and is not affected by the batch size; then it passes through a linear transformation layer, and finally generates position weights through an activation function; the other branch does not use depth convolution like Group 1 because the importance of each spatial position is jointly determined by the information at the same position on multiple channels, so ordinary convolution operations are performed, and finally the convolution feature maps are multiplied by their respective position weights; Finally, the output feature maps of Group 1 and Group 2 are concatenated, and then channel shuffling is performed among all groups to strengthen information sharing between different groups and jointly suppress noise interference.

Images