Lightweight network bearing fault diagnosis method and model based on fusion attention mechanism

Abstract

The application provides a lightweight network bearing fault diagnosis method and model based on a fusion attention mechanism; including an Inception structure of a network shallow layer, a down-sampling DownSample structure, a repeated stacking information extraction module IRGShuffle Block structure and the like. The fault type is determined through a deep learning network, and a Flooding regularization method is used in the model learning process to avoid zero training error loss, and a Stochastic depth method is used to strengthen the adaptability of the network. An improved MSKE-CSS attention structure is introduced, which respectively enhances the spatial dimension information extraction capability based on a multi-subspace feature map shuffle operation and enhances the channel dimension information extraction capability based on a multi-scale adaptive convolution kernel operation, and finally inserts the fusion attention structure into the backbone network to improve the network performance. The application can realize high-accuracy bearing fault diagnosis classification, and breaks the problems of traditional bearing fault diagnosis, such as strong noise condition, insufficient feature information extraction, low model diagnosis accuracy, slow training speed and large network parameter quantity, which is beneficial to ensure the normal operation of equipment and quickly and accurately identify faults.

Description

Technical Field

[0001] This application relates to the field of industrial intelligent manufacturing technology, and in particular to a lightweight network bearing fault diagnosis method and model based on a fusion attention mechanism. Background Technology

[0002] With the rapid development of industrial manufacturing, bearings, as core components in various large-scale machinery, play a crucial role in the operation of the entire mechanical equipment. In large-scale automated equipment, all internal components are closely interconnected; therefore, a failure in any component can lead to abnormal operation or even paralysis of the entire system, often resulting in serious personal injury accidents and irreparable economic losses. Therefore, in the development of the Industrial Internet, fault diagnosis technology for mechanical equipment is playing an increasingly important role.

[0003] Bearing faults are categorized into rolling element faults, inner ring faults, and outer ring faults, each corresponding to a fault size of 0.007, 0.014, and 0.021 inches. Since the 1960s, bearing fault diagnosis has primarily relied on expert systems and accumulated experience, such as signal analysis based on collected vibration signals. These methods inevitably possess a degree of subjectivity and cannot provide real-time information about the fault, potentially leading to serious consequences. With the advent of artificial intelligence, more and more researchers have begun data-driven fault diagnosis research, integrating it with deep learning. Existing technologies combine CBAM structures with convolutional neural networks, focusing on important feature information and adjusting weight parameters, but fail to capture spatial information at different scales to enrich the feature space. Other technologies use Inception and Google's EfficientNet V2 structures as fault diagnosis models, demonstrating good recognition performance under varying loads and high-noise environments, but these models have a large number of parameters. Still others combine fault diagnosis with the ShuffleNetv2 network structure, introducing hybrid convolutions to improve the network's feature extraction capabilities, but this does not consider the feature map redundancy inherent in deep networks. Therefore, this invention addresses the technical problem of how to achieve a diagnostic model with fewer network parameters, higher diagnostic accuracy, and faster training speed in bearing fault diagnosis methods. Summary of the Invention

[0004] This invention addresses existing deep learning network models by designing a high-performance, lightweight convolutional network. It primarily employs pointwise grouped convolutions to reduce the number of model parameters and uses Ghost convolutions, a computationally less computationally intensive operation, to generate redundant feature maps, further reducing model complexity. Recognizing the limitations of using single-scale convolutional kernels in shallow network layers, this invention replaces the original network's convolutional layers and max-pooling downsampling layers with an Inception structure. It adaptively selects the optimal kernel size for different training epochs and performs convolution and aggregation simultaneously at multiple scales during feature extraction, effectively expanding the network width.

[0005] Simultaneously, the HardSwish activation function is employed, which maintains model performance while simplifying calculation formulas and derivatives. A fusion attention mechanism algorithm is designed based on efficient channel attention units with multi-scale selectable convolutional kernels and subspace channel-shuffling spatial attention units. Channel attention aims to learn the weight relationships between different channels of the feature map, reducing the interference of redundant channels on image recognition; spatial attention aims to guide the network to focus on key regions of the image, using adaptively learned weights for different regions. Fusing these two dimensions of attention aggregates information from multiple scales, ensuring more effective extraction of latent feature information within the same receptive field.

[0006] Finally, the Stochastic depth method is used to randomly discard some neurons and some layer structures to enhance the network's adaptability. The Flooding regularization method is used to adjust the hyperparameter values ​​to maintain a fixed / preset training loss level and suppress overfitting of the model.

[0007] The present invention provides a lightweight network bearing fault diagnosis method based on a fusion attention mechanism, the method comprising: S1: Collect bearing vibration signals and perform data preprocessing on the bearing vibration signals; S2: Initialize model parameters, use the constructed IRGShuffleNet network based on MSKE-CSS fusion attention mechanism for learning and training, calculate the loss function between the obtained output vector and the expected value, and use the Adam optimizer for tuning; S3: Determine whether the accuracy of each generation of training has reached the optimal value. If it has, save the current best training model and go to S4; otherwise, continue iterative training. S4: Use the optimal training model to diagnose bearing faults.

[0008] Furthermore, S1 also includes: S101: Normalize the vibration signal of each bearing segment; S102: Perform augmentation processing on the sliding window sampling dataset; S103: Transform the processed one-dimensional data signal into a two-dimensional form using wavelet time-frequency conversion; S104: Divide the data into training and test sets.

[0009] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The Inception structure is used to replace the initial convolutional and max-pooling downsampling layers in the original network structure to adaptively select the optimal convolutional kernel size at different training epochs. Furthermore, when extracting features, convolution and aggregation are performed simultaneously at multiple sizes to extract features at different scales.

[0010] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The DownSample structure is designed so that for repeatedly stacked Shuffle Blocks, a depthwise splitting convolution with a stride of 2 is first used to perform downsampling operation to reduce the feature map size during each stacking, and the number of channels is doubled through the stitching operation. Introducing the Ghost module to replace the downsampling operation convolution; Design the IRGSHuffle Block structure, and introduce the Ghost module and channel shuffling operation for the part with a step size of 1; Assuming the input layer has c channels, the output layer has n channels, the ordinary convolution output has m channels, and the feature map has h height and w width, then the Ghost module includes an identity mapping and Linear operations; the theoretical speedup of the Ghost module is: ; Where k is The kernel size of the convolution is d; the kernel size of the depthwise separable convolution is s; the number of Ghost feature maps is s.

[0011] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of MSKE attention units: An adaptive method is used to determine the 1D convolution kernel size for MSKE attention units; Define a nonlinear function: ; Where k is the size of the convolution kernel.

[0012] Given C, we have: ; in , Based on the original k value, select a nearby scale and increase the k value to [value]. , , ;in, The calculation method and The values ​​are calculated in the same way; , The values ​​are respectively , , , ; The final result is obtained by fusing the three pieces of information generated with different k values. Feature map of scale.

[0013] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of CSS attention units: definition For one set of intermediate feature mappings: ; in The kernel size is Depth convolution, The kernel size is , It is a point-to-point convolution with only one filter; a maxpool operation with padding of 1. It is an attention map inferred from intermediate feature maps; each set of feature maps is refined into a set of feature maps after feature redistribution; ; in This indicates element-wise multiplication. This represents element-wise addition. The final result is obtained by concatenating each set of feature maps across channels using a concatenation operation. :

[0014] Introducing channel mixing operation, Different intergroup channels are merged.

[0015] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism also includes: Using the HardSwish activation function, the formula is: ; Where x represents the feature output by the network.

[0016] Replace the original Dropout layer with the Stochastic Depth method, and assume... Let Bernoulli random variable be used to represent The backbone transformation function of the layer exists. Or it doesn't exist. , will function and Multiply to bypass Layer backbone transformation function: ; in, express The probability of the existence of the layer backbone transformation function; express The output of the layer, Indicates a unit transformation. When hour Remain unchanged; when Then it simplifies to the identity function. ; Among them, from input To the last block The update rules are as follows: .

[0017] Where L represents the total number of layers.

[0018] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism also includes: The Flooding regularization method is introduced to intentionally prevent the training loss from decreasing further when it drops to a reasonably small value. Let the loss function be for: ; in This is the expected training loss value. These are model parameters; when hour At this point, normal mini-batch gradient descent is performed; when hour At this point, gradient ascent is performed; during model training, once the loss function reaches the threshold, the training process alternates between gradient descent and gradient ascent to keep the loss near the threshold.

[0019] S3 further includes: During iterative training, the model is trained using the training set data, and the best model parameters saved during training are transferred. The final fault diagnosis model is obtained by combining the best model with the test set data.

[0020] Based on the same concept, this application also provides a lightweight network bearing fault diagnosis model based on a fusion attention mechanism, wherein the fault diagnosis model adopts the IRGShuffleNet network model based on the MSKE-CSS fusion attention mechanism in the lightweight network bearing fault diagnosis method based on a fusion attention mechanism described above.

[0021] This invention employs an improved IRGSHuffleNet network. The original ShuffleNet network structure primarily uses pointwise grouped convolutions to reduce model complexity. With a stride of 1 in the backbone network, the input feature matrix channels of each unit are first split into two branches. The left branch is left unprocessed, while the right branch is processed... convolution, Depthwise separable convolution Convolution is used for feature extraction. Then, the two branches are concatenated to maintain the overall number of channels, and finally, a channel shuffling operation is performed. In cases where downsampling is required in the backbone network, without channel splitting, both branches use DW convolutions with a stride of 2 to reduce the feature map size. The number of channels is adjusted by convolution, and the number of channels in the output feature matrix is ​​doubled by concatenation operation, followed by channel shuffling operation.

[0022] Compared with the prior art, the beneficial effects of this application are as follows: This invention enables high-accuracy bearing fault diagnosis and classification, overcoming the limitations of traditional bearing fault diagnosis methods, such as insufficient feature information extraction, low model diagnostic accuracy, slow training speed, and large number of network parameters, under conditions of strong noise. This facilitates the normal operation of equipment and the rapid and accurate identification of faults. Attached Figure Description

[0023] Figure 1 This is a flowchart of the lightweight network bearing fault diagnosis method based on the fusion attention mechanism described in this invention.

[0024] Figure 2 This is a structural diagram of the bearing fault diagnosis system described in this invention.

[0025] Figure 3 This is a schematic diagram of the improved Inception shallow feature extraction structure described in this invention.

[0026] Figure 4 This is a schematic diagram of the Ghost structure described in this invention.

[0027] Figure 5 This is a schematic diagram of the channel mixing operation described in this invention.

[0028] Figure 6This is a schematic diagram of the improved DownSample structure described in this invention.

[0029] Figure 7 This is a schematic diagram of the improved IRGShuffle Block structure described in this invention.

[0030] Figure 8 This is a schematic diagram of the improved MSKE attention unit described in this invention.

[0031] Figure 9 This is a schematic diagram of the improved CSS attention unit described in this invention.

[0032] Figure 10 This is a schematic diagram of the Stochastic Depth described in this invention.

[0033] Figure 11 This is a graph showing the changes in training and testing loss before and after Flooding regularization as described in this invention. Detailed Implementation

[0034] The following describes in further detail a lightweight network bearing fault diagnosis method based on a fusion attention mechanism, in conjunction with specific embodiments and accompanying drawings.

[0035] The working process of this invention is as follows: After acquiring the original vibration signal, each signal segment is normalized and transformed into a form easily processed by a convolutional network. A fixed-window sliding window sampling dataset enhancement method is applied to the original data signal to increase the number of samples. The one-dimensional original vibration signal after data enhancement is transformed into a two-dimensional time-frequency graph using wavelet time-frequency transformation, fully preserving the time-domain and frequency-domain information of the signal. Finally, the dataset is divided into training and test sets. Model parameters are initialized, and the constructed IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism is used for training. The accuracy of each training iteration is compared to determine whether to save the current model as the best model. Training stops after reaching the maximum number of training iterations, at which point the best training model saved during the training process is obtained. The parameters of the best model saved during training are transferred, and the test set dataset is combined with the saved best model to obtain the final fault diagnosis output result.

[0036] The lightweight network bearing fault diagnosis method based on fusion attention mechanism of the present invention, such as Figure 1 As shown, the method includes: S1: Collect bearing vibration signals and perform data preprocessing on the bearing vibration signals; S2: Initialize model parameters, use the constructed IRGShuffleNet network based on MSKE-CSS fusion attention mechanism for learning and training, calculate the loss function between the obtained output vector and the expected value, and use the Adam optimizer for tuning; S3: Determine whether the accuracy of each generation of training has reached the optimal value. If it has, save the current best training model and go to S4; otherwise, continue iterative training. S4: Use the optimal training model to diagnose bearing faults.

[0037] The lightweight network-based bearing fault diagnosis method based on a fusion attention mechanism described in this invention has the following technical advantages: Utilizing the characteristics of the fusion attention mechanism and lightweight networks, it can accurately locate and diagnose bearing faults. By calculating the loss function and tuning the optimizer, the difference between the output vector and the expected value can be minimized, thereby improving the accuracy of the diagnostic results. The construction and training of lightweight networks can reduce computational load and model complexity, improving computational efficiency. This is crucial for real-time diagnosis of a large number of bearings in practical applications. The proposed method, based on an attention mechanism, can diagnose faults in different types of bearings. This method can be adaptively adjusted according to specific bearing types and vibration signal characteristics, exhibiting a certain degree of scalability. The method described in this invention has high accuracy and efficiency in bearing fault diagnosis, providing reliable diagnostic results and offering technical support for bearing maintenance and fault prevention.

[0038] MSKE-CSS is a feature selection method that incorporates an attention mechanism, enabling it to automatically select the most relevant features for bearing fault diagnosis. This method utilizes an attention-based mechanism to model and evaluate the importance of each feature during the feature selection process.

[0039] IRGShuffleNet is a lightweight network model that combines channel shuffling and group convolution techniques to reduce computational cost and complexity while maintaining model accuracy. This lightweight network model is suitable for bearing fault diagnosis under conditions of limited computational resources.

[0040] The lightweight network bearing fault diagnosis model based on the fusion attention mechanism of this invention adopts the IRGShuffleNet network model based on the MSKE-CSS fusion attention mechanism, which can effectively diagnose bearing faults and has high efficiency and accuracy.

[0041] Furthermore, S1 also includes: S101: Normalize the vibration signal of each bearing segment; S102: Perform augmentation processing on the sliding window sampling dataset; S103: Transform the processed one-dimensional data signal into a two-dimensional form using wavelet time-frequency conversion; S104: Divide the data into training and test sets.

[0042] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The Inception structure is used to replace the initial convolutional and max-pooling downsampling layers in the original network structure to adaptively select the optimal convolutional kernel size at different training epochs. Furthermore, when extracting features, convolution and aggregation are performed simultaneously at multiple sizes to extract features at different scales.

[0043] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The DownSample structure is designed so that for repeatedly stacked Shuffle Blocks, a depthwise splitting convolution with a stride of 2 is first used to perform downsampling operation to reduce the feature map size during each stacking, and the number of channels is doubled through the stitching operation. Introducing the Ghost module to replace the downsampling operation convolution; Design the IRGSHuffle Block structure, and introduce the Ghost module and channel shuffling operation for the part with a step size of 1; Assuming the input layer has c channels, the output layer has n channels, the ordinary convolution output has m channels, and the feature map has h height and w width, then the Ghost module includes an identity mapping and Linear operations; the theoretical speedup of the Ghost module is: ; Where k is The kernel size of the convolution is d; the kernel size of the depthwise separable convolution is s; the number of Ghost feature maps is s.

[0044] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of MSKE attention units: An adaptive method is used to determine the 1D convolution kernel size for MSKE attention units; Define a nonlinear function: ; Where k is the size of the convolution kernel.

[0045] Given C, we have: ; in , Based on the original k value, select a nearby scale and increase the k value to [value]. , , ;in, The calculation method and The values ​​are calculated in the same way; , The values ​​are respectively , , , ; The final result is obtained by fusing the three pieces of information generated with different k values. Feature map of scale.

[0046] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of CSS attention units: definition For one set of intermediate feature mappings: ; in The kernel size is Depth convolution, The kernel size is , It is a point-to-point convolution with only one filter; a maxpool operation with padding of 1. It is an attention map inferred from intermediate feature maps; each set of feature maps is refined into a set of feature maps after feature redistribution; ; in This indicates element-wise multiplication. This represents element-wise addition. The final result is obtained by concatenating each set of feature maps across channels using a concatenation operation. :

[0047] Introducing channel mixing operation, Different intergroup channels are merged.

[0048] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism also includes: Using the HardSwish activation function, the formula is: ; Where x represents the feature output by the network.

[0049] Replace the original Dropout layer with the Stochastic Depth method, and assume... Let Bernoulli random variable be used to represent The backbone transformation function of the layer exists. Or it doesn't exist. , will function and Multiply to bypass Layer backbone transformation function: ; in, express The probability of the existence of the layer backbone transformation function; express The output of the layer, Indicates a unit transformation. When hour Remain unchanged; when Then it simplifies to the identity function. ; Among them, from input To the last block The update rules are as follows: .

[0050] Where L represents the total number of layers.

[0051] The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism also includes: The Flooding regularization method is introduced to intentionally prevent the training loss from decreasing further when it drops to a reasonably small value. Let the loss function be for: ; in This is the expected training loss value. These are model parameters; when hour At this point, normal mini-batch gradient descent is performed; when hour At this point, gradient ascent is performed; during model training, once the loss function reaches the threshold, the training process alternates between gradient descent and gradient ascent to keep the loss near the threshold.

[0052] S3 further includes: During iterative training, the model is trained using the training set data, and the best model parameters saved during training are transferred. The final fault diagnosis model is obtained by combining the best model with the test set data.

[0053] Please see Figure 2 The bearing fault diagnosis system described in this invention consists of three stages: data preprocessing, model training, and model testing. In the model structure, an improved Inception structure, DownSample structure, IRGShuffle Block structure, and MSKE-CSS fused attention unit are used to enhance network performance. The Stochastic Depth method is employed to enhance the network's anti-interference performance and generalization ability. Flooding regularization is used to avoid zero-training loss error.

[0054] Please see Figure 3 The Inception shallow feature extraction structure described in this invention first employs... Convolution performs dimensionality reduction by using two consecutive... Convolutional layers instead Convolutional layers reduce the number of parameters while maintaining the same receptive field. In the first stage of information aggregation, a concatenation operation is used to stitch together channels; in the second stage, the idea of ​​residual networks is adopted to pass the original information through... After adjusting the number of channels in the convolution, it is added to the multi-scale information from multiple branches, and the information from the initial layer is passed to deeper layers through matrix addition, thus avoiding the problems of gradient vanishing and gradient exploding.

[0055] Please see Figure 4 The Ghost module described in this invention first performs a normal... Convolution yields intrinsic feature maps, which are then integrated to generate feature condensations for the input feature layer; finally, ordinary features are applied... The convolution result is used to perform layer-by-layer convolution of depthwise separable convolution to obtain the Ghost feature map; finally, the intrinsic feature map and the Ghost feature map are concatenated to obtain the output feature map.

[0056] Please see Figure 5 For ordinary group convolutions, the convolution operation is performed on the channel information within a group, and there is no information exchange between groups. Therefore, the feature matrix is ​​further divided into three subgroups, and the corresponding subgroups in each group are put together to form a new feature matrix. Information exchange between groups is achieved through channel shuffling. Assume the input layer is divided into g groups, ensuring the total number of channels is a multiple of g. First, the channel dimension is split into... Two dimensions, each row is a group, and each group has n elements. Then transpose these two dimensions to become The dimensions are then converted back into the dimensions of the input channels. For example, when... , The operation process is as follows: (18) The channel shuffling operation used here is a uniform shuffling operation, and the number of groups is set to 4. This is because when the number of groups in the grouped convolution increases, the memory access time cost (MAC) will increase. Although the number of FLOPs remains unchanged, it will still directly affect the model inference speed.

[0057] Please see Figure 6 The DownSample structure described in this invention generates redundant feature maps using a Ghost module with lower computational cost, and then reduces the feature map size using a depthwise separable convolution with a stride of 2. For the residual connection part, depthwise separable convolution is used to adjust the feature map size and number of channels, and then a concatenation operation is used to double the number of channels. Finally, the output feature map is shuffled through a channel mixing operation to enhance the information exchange capability between different groups of channels.

[0058] Please see Figure 7 The inverse residual structure in the IRGSHuffle Block structure of this invention first processes the input feature map through... The convolution is then expanded to a dimension of 4; then... DW convolution is used for feature integration; finally, through... Convolution dimensionality reduction.

[0059] Please see Figure 8 The fusion attention structure described in this invention consists of MSKE channel-dimensional attention units and CSS spatial-dimensional attention units. The MSKE attention unit comprises three stages: splitting, fusion, and selection. The splitting operator generates multiple paths of different kernel sizes, which are then processed separately... Depthwise separable convolution and Dilated convolution yields the corresponding feature map. and And reduce the number of model parameters, then add and fuse the feature maps to obtain... These paths correspond to different neuron receptive field sizes. The fusion operator combines and aggregates information from multiple paths through global average pooling and global max pooling to obtain a global comprehensive representation of the selection weights. After global average pooling, it is learned through an adaptive 1D convolution, which can significantly reduce the model complexity while maintaining performance with appropriate cross-channel interactions. Finally, it is passed through the corresponding SoftMax activation function layers and then combined with the original... and Multiplication assigns weights to the feature maps in the branch. The selection operator aggregates feature maps with kernels of different sizes based on the selected weights.

[0060] Please see Figure 9 The CSS attention unit described in this invention divides the input feature map into G groups of depthwise separable convolutions, performs DW convolution followed by max pooling to highlight local information regions and collect spatial information. Then, PW convolution is used to adjust the number of channels, and after passing through the SoftMax activation function and expanding the dimensions, it is multiplied with the input feature map, and the linear relationship between different feature map subspaces is used to integrate information.

[0061] Please see Figure 10 This is a schematic diagram of the Stochastic Depth described in this invention. A gradually increasing scaling factor is used to randomly discard a portion of the network structure during training, accelerating the training process. During testing, the complete network structure is used, resulting in a larger expected depth and improved model generalization ability. For each batch, a set of layers is randomly selected, and their corresponding transformation functions are deleted, retaining only the identity skip connections.

[0062] Please see Figure 11 This is the Flooding regularization method described in this invention, for the threshold The settings are typically configured so that the test loss begins to rise. The optimal result is even smaller, because when the test loss starts to rise, it not only indicates that the model is overfitting, but may also mean that the test set and the training set are not independently and identically distributed. In this case, it is often necessary to continue learning along the gradient descent direction. In this invention, it is set as follows: .

[0063] Example 2: Based on the same concept, this application also provides a lightweight network bearing fault diagnosis model based on a fusion attention mechanism, wherein the fault diagnosis model adopts the IRGShuffleNet network model based on the MSKE-CSS fusion attention mechanism in the lightweight network bearing fault diagnosis method based on a fusion attention mechanism described above.

[0064] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0065] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0066] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0067] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

Claims

1. A lightweight network bearing fault diagnosis method based on a fusion attention mechanism, characterized in that, The method includes: S1: Collect bearing vibration signals and perform data preprocessing on the bearing vibration signals; S2: Initialize model parameters, use the constructed IRGShuffleNet network based on MSKE-CSS fusion attention mechanism for learning and training, calculate the loss function between the obtained output vector and the expected value, and use the Adam optimizer for tuning; S3: Determine whether the accuracy of each generation of training has reached the optimal value. If it has, save the current best training model and go to S4; otherwise, continue iterative training. S4: Use the optimal training model described above for bearing fault diagnosis; The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The Inception structure is used to replace the initial convolution and max pooling downsampling layers in the original network structure to adaptively select the optimal convolution kernel size at different training epochs. Furthermore, when extracting features, convolution and aggregation are performed simultaneously at multiple sizes to extract features at different scales. The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: The DownSample structure is designed so that for repeatedly stacked Shuffle Blocks, a depthwise splitting convolution with a stride of 2 is first used to perform downsampling operation to reduce the feature map size during each stacking, and the number of channels is doubled through the stitching operation. Introducing the Ghost module to replace the downsampling operation convolution; Design the IRGSHuffle Block structure, and introduce the Ghost module and channel shuffling operation for the part with a step size of 1; Assuming the input layer has c channels, the output layer has n channels, the ordinary convolution output has m channels, and the feature map has h height and w width, then the Ghost module includes an identity mapping and Linear operations; the theoretical speedup of the Ghost module is: ； Where k is The kernel size of the convolution; d is the kernel size of the depthwise separable convolution; s is the number of Ghost feature maps; The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of MSKE attention units: An adaptive method is used to determine the 1D convolution kernel size for MSKE attention units; Define a nonlinear function: ； Where k is The kernel size of the convolution is C, which is a given number of channels; Given C, we have: ； in Based on the original k value, select a nearby scale and increase the k value to [value]. , , ; The final result is obtained by fusing the three pieces of information generated with different k values. Scale feature map; The IRGSHuffleNet network based on the MSKE-CSS fusion attention mechanism in S2 also includes: the design of CSS attention units: definition For one set of intermediate feature mappings: ； in The kernel size is Depth convolution, The kernel size is , maxpool operation with filler set to 1; It is a point-to-point convolution with only one filter; It is an attention map inferred from intermediate feature maps; each set of feature maps is refined into a set of feature maps after feature redistribution; ； in This indicates element-wise multiplication. This represents element-wise addition; by concatenating each set of feature maps across channels, the final result is obtained. : ； Introducing channel mixing operation, Different intergroup channels are merged.

2. The lightweight network bearing fault diagnosis method based on fusion attention mechanism according to claim 1, characterized in that, S1 further includes: S101: Normalize the vibration signal of each bearing segment; S102: Perform augmentation processing on the sliding window sampling dataset; S103: Transform the processed one-dimensional data signal into a two-dimensional form using wavelet time-frequency conversion; S104: Divide the data into training and test sets.

3. The lightweight network bearing fault diagnosis method based on fusion attention mechanism according to claim 2, characterized in that, Also includes: Using the HardSwish activation function, the formula is: ； Where x represents the feature of the network output; Replace the original Dropout layer with the Stochastic Depth method, and assume... Let Bernoulli random variable be used to represent The backbone transformation function of the layer exists. Or it doesn't exist. , will function and Multiply to bypass Layer backbone transformation function: ； in, express The probability of the existence of the layer backbone transformation function; express The output of the layer, Indicates unit transformation; when hour Remain unchanged; when Then it simplifies to the identity function. ; Among them, from input To the last block The update rules are as follows: ； Where L represents the total number of layers.

4. The lightweight network bearing fault diagnosis method based on fusion attention mechanism according to claim 3, characterized in that, It also includes: introducing a Flooding regularization method, which intentionally prevents the training loss from decreasing further when it drops to a reasonably small value; Let the loss function be for: ； in This is the expected training loss value. These are model parameters; when hour At this point, normal mini-batch gradient descent is performed; when hour At this point, gradient ascent is performed; during model training, once the loss function reaches the threshold, the training process alternates between gradient descent and gradient ascent to keep the loss near the threshold.

Images