Multi-view bearing fault diagnosis method based on noise, edge and difficulty perception

Abstract

The invention discloses a multi-view bearing fault diagnosis method based on noise, edge and difficulty perception, and the method constructs a multi-view perception framework of an input view, an operator view and a training view, and specifically comprises the steps: constructing a noise perception multi-scale expansion gating modulator at the input view; designing an adaptive noise perception graph to generate directional Gaussian perturbation so as to actively enhance the noise distinguishing capability; constructing a boundary edge sensing module at an operator perspective, and designing boundary adaptive filling and edge adaptive gradient sensitive masks to effectively suppress convolution boundary artifacts and enhance fault edge features; an inverse distillation strategy guided by difficulty perception clustering is put forward from a training perspective, a sample learning state is dynamically recorded by constructing a difficulty memory mechanism, and an intra-class multi-state soft label is generated by designing online clustering and inverse neighborhood sampling, so that a model is guided to carry out targeted optimization on a difficult sample, and the fault discrimination capability is remarkably improved. Experiments prove that the method has higher fault diagnosis performance.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically to a multi-view bearing fault diagnosis method (MVPNet) based on noise, edge, and difficulty perception, which is particularly suitable for fault diagnosis in complex industrial environments such as high noise levels. This method addresses the problems of existing methods, such as weak fault pulse discrimination under noise interference, destruction of boundary topology due to convolution operations, insufficient fault edge feature response, and lack of targeted optimization for easy and difficult samples during training, by constructing noise perception from the input viewpoint, boundary edge perception from the operator viewpoint, and difficulty perception from the training viewpoint. Background Technology

[0002] With the increasing intelligence and complexity of industrial equipment, fault diagnosis technology for rotating machinery such as bearings has become a core means to ensure the safe operation of equipment. While current diagnostic methods based on two-dimensional time-frequency representation have made significant progress, their development is limited by the widely adopted single-view perception methods, resulting in multi-view perception defects: the inherent strong noise of industrial vibration signals easily masks weak fault pulses, and existing methods lack explicit noise modeling and perception mechanisms, leading to a sharp decline in the model's feature discrimination ability under low signal-to-noise ratio conditions; traditional convolution operations, due to their fixed zero-padding strategy, disrupt the topological integrity of the time-frequency image boundaries; simultaneously, their non-adaptive convolution kernel weights struggle to focus on key transient pulse edges, leading to degraded edge perception; furthermore, traditional training strategies treat all samples equally, ignoring the differences in learning difficulty between samples caused by damage degree and operating condition fluctuations, thus lacking targeted optimization for easy and difficult samples, resulting in suboptimal decision boundaries and limited model generalization ability. These shortcomings in the synergistic effects of noise, edge, and difficulty perception fundamentally restrict the robustness and accuracy of bearing fault diagnosis models in complex industrial environments. To address the aforementioned issues, this invention proposes a multi-view bearing fault diagnosis method (MVPNet) based on noise, edge, and difficulty perception. This method enhances the ability to distinguish input noise by designing a noise-perceived multi-scale dilatation-gated modulator, constructs a boundary edge perception module to suppress convolution boundary artifacts and strengthen the bearing fault edge response, and proposes a difficulty-perceived clustering-guided retrodistillation strategy to achieve targeted optimization of easy and difficult samples, significantly improving the bearing fault diagnosis performance. Summary of the Invention

[0003] This invention addresses the shortcomings of existing bearing fault diagnosis methods based on two-dimensional time-frequency representation, such as insufficient noise perception, degraded edge perception, and lack of difficulty perception. It proposes a multi-view bearing fault diagnosis method based on noise, edge, and difficulty perception. The specific implementation steps of this invention are as follows: 1. Construct a noise-aware multi-scale dilation-gated modulator from the input perspective, design an adaptive noise-aware map to generate directional Gaussian perturbations, combine multi-scale dilation convolution to construct complementary receptive fields, and utilize global semantic gating to achieve feature modulation to actively enhance noise discrimination capability. The steps are as follows: (1a) Noise sensing and injection based on two-dimensional time-frequency representation Given input tensor ,in Indicates batch size. Indicates the number of channels. Indicates altitude, To represent the width, first perform batch regularization to obtain... ,Will Input to a noise predictor An adaptive noise-aware map is generated by applying a sigmoid function and regularization constraints. : in and Describe the regularization constraint interval, and Mapping to an appropriate range to ensure training stability, Gaussian noise is generated based on the input during the training phase. And it uses an adaptive noise perception map to scale it channel-by-channel and position-by-position to generate directional noise. : This is then injected into the network to obtain noise-aware input. : (1b) Multi-scale feature extraction Will Input three depthwise separable convolutions with different dilation rates respectively. In order to construct complementary multi-scale receptive fields: in, for The corresponding batch regularization layer, It is the ReLU activation function; (1c) Global semantic gating modulation right and characteristics of each branch Perform global average pooling separately to obtain semantic vectors. and And through the projection function , , Map the semantic vectors to global query vectors respectively. Key vectors of each branch AND value vector Then, the scaling dot product similarity score of each branch is calculated. The cross-branch attention score vector is obtained through the Softmax function. : in, The internal feature dimension is then used, and the attention score vector is used to weight and fuse the value vectors of each branch to obtain the global fusion vector. Spatial broadcasting of the vector and point convolution. The data is then refined, followed by batch regularization and ReLU activation to obtain the modulated feature map. Finally, it is connected to the original input via residual connection. Add: in This is the final output of the noise-aware multiscale expansion gated modulator.

[0004] 2. Construct a boundary edge sensing module from the operator's perspective, and design boundary adaptive padding and edge adaptive gradient sensitive masks to effectively suppress convolution boundary artifacts and enhance fault edge features. The steps are as follows: (2a) Design of boundary adaptive filling module Let the fill size be , Extracting the kernel size left and right boundaries List: in and This represents the extracted leftmost and rightmost column boundary slices, which are then input into a lightweight style-aware generator. middle: in and To fill in content in a style consistent with the original content, then stitch the pieces horizontally: in For the horizontally expanded feature map, the same operation is then applied to the top and bottom boundaries to produce the complete boundary adaptive padding feature map. : in and These are the top and bottom fill contents, respectively. (2b) Design an edge-adaptive gradient-sensitive mask Will Averaging along the channel dimension yields the grayscale feature map. and using predefined Sobel operators and Calculate the gradients in the horizontal and vertical directions of the grayscale feature map. and : in This represents the Sobel convolution operation, followed by further calculation of the overall gradient magnitude. : Intra-batch regularization of the overall gradient magnitude: in Take the spatial maximum value of each sample. The gradient magnitude represents the regularization, followed by a learnable gain. Soft threshold With offset Adaptive adjustment and cropping are performed, and then a smooth edge-sensitive mask is generated using the Sigmoid function. : Finally, the mask is upsampled along the channel dimension to obtain the final edge-sensitive mask. ; (2c) Masking Then, the features after adaptive boundary padding Execute 3 3 convolutions yield richly detailed features. and to Execute 1 1. Convolution and necessary downsampling To obtain background context features Finally, edge-sensitive masks are used for spatial adaptive fusion: in For a full 1 tensor, This is the output of the boundary edge sensing module.

[0005] 3. From a training perspective, a difficulty-aware clustering-guided reverse distillation strategy is proposed. This strategy dynamically records the learning state of samples by constructing a difficulty memory mechanism, and designs online clustering and reverse neighborhood sampling to generate multi-state soft labels within each class. This guides the model to perform targeted optimization on both easy and difficult samples, significantly improving the bearing fault detection capability. The steps are as follows: (3a) Construct a difficulty memory mechanism to dynamically record the learning status of samples. Design a memory bank consisting of a feature memory pool, a correct prediction memory pool, and an incorrect prediction memory pool. For each training batch of samples... Its characteristics Predicting probability Authentic Labels and cross-entropy loss For dynamically managed samples, the difficulty weight is first calculated. : in This represents the per-sample loss vector of the batch to which the sample belongs, and then the features of the sample are stored in the feature pool of the corresponding category. If a sample is misclassified, then the tuple will be... Store in the error prediction memory pool If the sample is correctly classified, then calculate the predicted probability entropy of that sample. : in Number of fault categories Represent a numerically stable term, then calculate the ordination metric. and the triplet according to Values ​​are stored in the correct prediction memory pool in ascending order. When the capacity of the correct prediction pool reaches its limit, remove it. The largest item is selected, and the remaining items in the pool are selected using a first-in, first-out (FIFO) strategy. (3b) Online clustering and reverse neighborhood sampling For the current sample, from its true category Feature pool Extract the feature set, and if the capacity is sufficient, perform online K-Means clustering to obtain... Cluster centers Then calculate the features of the current sample. Similarity weights with each cluster center : Then, an inverse sampling rule was used to select the candidate teacher set. If the current sample is predicted correctly, then it is removed from the error pool. Difficult samples are selected as teachers; if the prediction is incorrect, the correct samples are selected from the pool. Easy samples were selected as teachers, and then for each cluster center... Find the nearest neighbor sample in the middle: in for Features in express Mid-distance cluster centers The characteristics of the most recent samples are used to derive the predicted distribution. ; (3c) Optimization of Backdistillation Based on Multi-State Soft Labels Through similarity weight The predicted distributions of the nearest neighbor samples corresponding to each cluster center are weighted and fused to construct soft labels with intra-class multi-state distribution characteristics. : This soft label serves as a refined teacher signal, measured through KL divergence. To guide the model in targeted optimization of both easy and difficult samples: in The reverse distillation loss is used to guide model training. Finally, when training ends, the model's output is fed into the linear classification head to be projected onto the category dimension, thereby achieving bearing fault diagnosis.

[0006] The method of the present invention has the following advantages: (1) Enhanced noise perception capability: By constructing a noise perception multi-scale dilation gated modulator, an adaptive noise perception map is designed to generate directional Gaussian perturbations. Multi-scale dilation convolution is used, combined with a gated modulation mechanism to integrate adaptive noise into the backbone network, forcing the model to effectively distinguish between bearing fault pulses and noise interference, thereby improving the bearing fault feature discrimination capability in a strong noise environment.

[0007] (2) Enhanced edge perception: A boundary edge perception module with a collaborative dual-branch design is proposed. The boundary adaptive filling module maintains the integrity of the time-frequency image topology through adaptive filling, while the edge adaptive gradient sensitive mask uses gradient information to guide and amplify the response of significant fault edges. This combined method effectively suppresses convolution boundary artifacts and enhances the edge features of bearing fault pulses.

[0008] (3) Targeted optimization of easy and difficult samples: A reverse distillation strategy guided by difficulty-aware clustering is proposed, a difficulty memory mechanism is constructed to record the sample-level learning state, and online clustering and reverse neighborhood sampling are designed to generate soft labels, so that the model can dynamically adjust the feature learning of easy and difficult samples and promote the formation of a more discriminative and robust feature space.

[0009] (4) Complete construction of multi-view collaborative perception framework: By integrating the above three perception enhancement components, a multi-view bearing fault diagnosis method based on noise, edge and difficulty perception is constructed. The overall design achieves complementary enhancement of feature robustness, structural integrity and discrimination specificity, and provides an end-to-end bearing fault diagnosis solution that is superior to existing methods. Attached Figure Description

[0010] Figure 1 This is the overall framework diagram of the present invention. Figure 2 This is a classification result diagram based on the confusion matrix of a random experiment on the CWRU dataset. Figure 3 This is a classification result using a confusion matrix from a random experiment on the AUST dataset. Detailed Implementation

[0011] 1. Data Preprocessing Data acquisition and standardization: Collect raw vibration signals of industrial bearings, calculate the mean and standard deviation of training set signals, perform standardization processing on raw signals to eliminate dimensional differences and improve the convergence efficiency and stability of model training; Two-dimensional time-frequency representation acquisition: To effectively extract the time-frequency representation from the vibration signal, a time-frequency analysis method based on continuous wavelet transform is adopted. A sliding window strategy is used for sample segmentation, with a window length of 1024 sampling points. The standardized vibration signal is converted into a two-dimensional time-frequency representation through continuous wavelet transform. The resulting two-dimensional time-frequency representation has a dimension of 224×224, with the horizontal and vertical axes representing the time series and frequency distribution, respectively.

[0012] 2. Noise-sensing multiscale dilatation-gated modulation The generated two-dimensional time-frequency representation is input into a noise-aware multi-scale dilation-gated modulator, and the following operations are performed sequentially: After batch regularization of the input time-frequency tensor, an adaptive noise-aware map constrained by Sigmoid is generated by a noise predictor. This map is then used to scale dynamically generated Gaussian noise channel-wise and position-wise to form directional and regional noise, which is then injected into the input to enhance noise discriminability. Subsequently, the noise-aware features are fed into three sets of depthwise separable convolutions with different dilation rates to extract multi-scale representations. Global average pooling is then performed on the input and features at each scale, and the results are mapped to query, key, and value vectors. Based on the scaling dot product similarity, cross-branch attention weights are formed using Softmax. The semantics of each branch are weighted and fused and broadcast back to the feature space. Finally, after point convolution and nonlinear refinement, the result is added to the initial input residual to obtain the final noise-aware multi-scale output.

[0013] 3. Suppress convolution boundary artifacts and enhance fault edge features The feature map from the previous stage is input into the boundary edge perception module, and the following operations are performed sequentially: First, slices of the left, right, top, and bottom boundaries are extracted from the input features and fed into a lightweight style generator to produce texture-consistent fill content. This fill content is then stitched with the original image to form an artifact-free, adaptive fill feature. Next, the input is averaged along the channels to obtain a grayscale image. The horizontal and vertical gradients are calculated using the Sobel operator, and the gradient magnitudes are obtained. After batch regularization, learnable gain, soft thresholding, and offset adjustments, a smooth gradient-sensitive mask is generated by Sigmoid and extended to the channel dimension. Finally, a 3×3 convolution is performed on the fill feature to obtain detailed features, and a 1×1 convolution and necessary downsampling are performed on the original image to obtain background features. These features are then fused point-by-point in the spatial dimension based on the gradient-sensitive mask, so that high-gradient regions retain sharp edges and low-gradient regions retain robust backgrounds, thus outputting a feature map with friendly and enhanced edges.

[0014] 4. Targeted optimization of easy and difficult samples A memory pool consisting of a feature pool, a correct pool, and an error pool is maintained. In each training batch, the features, prediction probabilities, loss, and difficulty weights of samples are recorded, and samples are added to their respective pools based on their prediction accuracy. Online K-Means is used to construct multi-state cluster centers in the feature pool of the sample's category, and the similarity between the current sample and the center is calculated. Teacher candidates are selected from either the error pool or the correct pool based on backsampling principles, and the corresponding prediction distribution is obtained by searching for nearest-neighbor features from the teacher set for each center. Finally, the teacher distributions of each center are weighted and fused using similarity as the weight to obtain intra-class multi-state soft labels. KL divergence is used as the backdistillation loss to guide model learning, making it more sensitive to difficult samples and more robust to easy samples during training, thereby improving fault detection accuracy. After optimization, bearing fault diagnosis is performed using a classification head.

[0015] The effectiveness of this invention was further verified through the following experiments: On the CWRU and self-built AUST bearing datasets, the average diagnostic accuracy of the proposed method reached 100% and 99.50%, respectively. In noise robustness testing, under a -10dB strong noise environment, the proposed method maintained an accuracy of 84.23% on the CWRU dataset, and recovered to 100% perfect classification performance under a 0dB signal-to-noise ratio condition. The ablation study of the system confirmed the necessity of the collaboration of each module. The accuracy of the baseline model on the CWRU and AUST datasets was 98.50% and 95.17%, respectively. After introducing the boundary edge perception module, the accuracy improved to 99.33% and 98.77%. After further integrating the noise-aware multi-scale expansion gated modulator, the accuracy reached 99.90% and 99.33%. Finally, after combining the difficulty-aware clustering-guided reverse distillation strategy to form a complete method, the accuracy reached 100% and 99.50%, respectively, verifying the advantages of the multi-view perception mechanism in complex fault feature extraction and difficult / easy sample optimization. Through the above experiments, the bearing fault diagnosis effect of the proposed method has been further verified.

Claims

1. A multi-view bearing fault diagnosis method based on noise, edge, and difficulty perception, characterized in that, The method includes the following steps: (1) Construct a noise-aware multi-scale dilation gated modulator from the input perspective, design an adaptive noise-aware map to generate directional Gaussian perturbations, construct complementary receptive fields by combining multi-scale dilation convolution, and use global semantic gating to achieve feature modulation to actively enhance noise discrimination ability. (2) Construct a boundary edge sensing module from the perspective of operators, and design boundary adaptive filling and edge adaptive gradient sensitive mask to effectively suppress convolution boundary artifacts and enhance fault edge features; (3) From the perspective of training, a reverse distillation strategy guided by difficulty-aware clustering is proposed. By constructing a difficulty memory mechanism, the learning state of samples is dynamically recorded. Online clustering and reverse neighborhood sampling are designed to generate soft labels with multiple states within the class, thereby guiding the model to perform targeted optimization on easy and difficult samples, significantly improving the fault discrimination ability.

2. The multi-view bearing fault diagnosis method based on noise, edge, and difficulty perception according to claim 1, characterized in that... Step (1) involves constructing a noise-aware multi-scale dilation-gated modulator from the input perspective, designing an adaptive noise-aware map to generate directional Gaussian perturbations, constructing complementary receptive fields by combining multi-scale dilation convolutions, and utilizing global semantic gating to achieve feature modulation to actively enhance noise discrimination capabilities. The steps are as follows: (2a) Noise sensing and injection based on two-dimensional time-frequency representation Given input tensor ,in Indicates batch size. Indicates the number of channels. Indicates altitude, To represent the width, first perform batch regularization to obtain... ,Will Input to a noise predictor An adaptive noise-aware map is generated by applying a sigmoid function and regularization constraints. : in and Describe the regularization constraint interval, and Mapping to an appropriate range to ensure training stability, Gaussian noise is generated based on the input during the training phase. And it uses an adaptive noise perception map to scale it channel-by-channel and position-by-position to generate directional noise. : This is then injected into the network to obtain noise-aware input. : (2b) Multi-scale feature extraction Will Input three depthwise separable convolutions with different dilation rates respectively. In order to construct complementary multi-scale receptive fields: in, for The corresponding batch regularization layer, It is the ReLU activation function; (2c) Global semantic gating modulation right and characteristics of each branch Perform global average pooling separately to obtain semantic vectors. and And through the projection function , , Map the semantic vectors to global query vectors respectively. Key vectors of each branch AND value vector Then, the scaling dot product similarity score of each branch is calculated. The cross-branch attention score vector is obtained through the Softmax function. : in, The internal feature dimension is then used to weight and fuse the branch value vectors using this attention score vector, resulting in a global fusion vector. Spatial broadcasting of the vector and point convolution. The data is then refined, followed by batch regularization and ReLU activation to obtain the modulated feature map. Finally, it is connected to the original input via residual connection. Add: in This is the final output of the noise-aware multiscale expansion gated modulator.

3. The multi-view bearing fault diagnosis method based on noise, edge, and difficulty perception according to claim 1, characterized in that... Step (2) involves constructing a boundary edge sensing module from the operator's perspective, designing boundary adaptive padding and edge adaptive gradient-sensitive masks to effectively suppress convolution boundary artifacts and enhance fault edge features. The steps are as follows: (3a) Design of boundary adaptive filling module Let the fill size be , Extracting the kernel size left and right boundaries List: in and This represents the extracted leftmost and rightmost column boundary slices, which are then input into a lightweight style-aware generator. middle: in and To fill in content in a style consistent with the original content, then stitch the pieces horizontally: in For the horizontally expanded feature map, the same operation is then applied to the top and bottom boundaries to produce the complete boundary adaptive padding feature map. : in and These are the top and bottom fill contents, respectively. (3b) Design an edge-adaptive gradient-sensitive mask Will Averaging along the channel dimension yields the grayscale feature map. and using predefined Sobel operators and Calculate the gradients in the horizontal and vertical directions of the grayscale feature map. and : in This represents the Sobel convolution operation, followed by further calculation of the overall gradient magnitude. : Intra-batch regularization of the overall gradient magnitude: in Take the spatial maximum value of each sample. The gradient magnitude represents the regularization, followed by a learnable gain. Soft threshold With offset Adaptive adjustment and cropping are performed, and then a smooth edge-sensitive mask is generated using the Sigmoid function. : Finally, the mask is upsampled along the channel dimension to obtain the final edge-sensitive mask. ; (3c) Masking Then, the features after adaptive boundary padding Execute 3 3 convolutions yield richly detailed features. and to Execute 1 1. Convolution and necessary downsampling To obtain background context features Finally, edge-sensitive masks are used for spatial adaptive fusion: in For a full 1 tensor, This is the output of the boundary edge sensing module.

4. The multi-view bearing fault diagnosis method based on noise, edge, and difficulty perception according to claim 1, characterized in that... Step (3) proposes a back-distillation strategy guided by difficulty-aware clustering from a training perspective. This strategy dynamically records the learning state of samples by constructing a difficulty memory mechanism and designs online clustering and inverse neighborhood sampling to generate intra-class multi-state soft labels. This guides the model to perform targeted optimization on samples of varying difficulty, significantly improving the bearing fault discrimination capability. The steps are as follows: (4a) Construct a difficulty memory mechanism to dynamically record the learning status of samples. Design a memory bank consisting of a feature memory pool, a correct prediction memory pool, and an incorrect prediction memory pool. For each training batch of samples... Its characteristics Predicting probability Authentic Labels and cross-entropy loss For dynamically managed samples, the difficulty weight is first calculated. : in This represents the per-sample loss vector of the batch to which the sample belongs, and then the features of the sample are stored in the feature pool of the corresponding category. If a sample is misclassified, then the tuple will be... Store in the error prediction memory pool If the sample is correctly classified, then calculate the predicted probability entropy of that sample. : in Number of fault categories Represent a numerically stable term, then calculate the ordination metric. and the triplet according to Values ​​are stored in the correct prediction memory pool in ascending order. When the capacity of the correct prediction pool reaches its limit, remove it. The largest item is selected, and the remaining items in the pool are selected using a first-in, first-out (FIFO) strategy. (4b) Online clustering and reverse neighborhood sampling For the current sample, from its true category Feature pool Extract the feature set, and if the capacity is sufficient, perform online K-Means clustering to obtain... Cluster centers Then calculate the features of the current sample. Similarity weights with each cluster center : Then, an inverse sampling rule was used to select the candidate teacher set. If the current sample is predicted correctly, then it is removed from the error pool. Difficult samples are selected as teachers; if the prediction is incorrect, the correct samples are selected from the pool. Easy samples were selected as teachers, and then for each cluster center... Find the nearest neighbor sample in the middle: in for Features in express Mid-distance cluster centers The characteristics of the most recent samples are used to derive the predicted distribution. ; (4c) Optimization of Backdistillation Based on Multi-State Soft Labels Through similarity weight The predicted distributions of the nearest neighbor samples corresponding to each cluster center are weighted and fused to construct soft labels with intra-class multi-state distribution characteristics. : This soft label serves as a refined teacher signal, measured through KL divergence. To guide the model in targeted optimization of both easy and difficult samples: in The reverse distillation loss is used to guide model training. Finally, when training ends, the model's output is fed into the linear classification head to be projected onto the category dimension, thereby achieving bearing fault diagnosis.

Images