Bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction

Abstract

The invention discloses a bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction. The specific implementation process is as follows: constructing a dynamic kernel evolution network in a time domain flow, guiding convolution kernel weight adaptive evolution by using a time context, capturing transient impact and constructing a multi-resolution graph topology; in the frequency domain flow, a spectrum sparse purification network is constructed, a significance evaluation mechanism based on a learnable quantization matrix is established, and high-fidelity purification and denoising of fault key information are realized by using a soft threshold switch; in order to make full use of the flow features, a heterogeneity perception cross-flow interaction mechanism is provided, time domain energy heterogeneity is recovered through a latent feature re-calibration module, frequency domain features are fused to form cross-flow enhanced representation, the enhanced representation is introduced into an adaptive graph Transform, graph topology is used as a skeleton, and the cross-flow interaction mechanism is constructed. And the node features are driven to perform cross-flow spatial dynamic interaction in the multi-scale topology. Experiments show that the problems of strong noise interference and feature homogenization can be effectively solved, and the fault diagnosis precision is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically a bearing fault diagnosis method based on cross-current feature enhancement and graph topology interaction, particularly suitable for rolling bearing fault diagnosis scenarios. This invention achieves dynamic calibration of time-varying impact features and multi-resolution graph topology construction by constructing a dynamic kernel evolution network, adaptively selecting key frequency bands using a spectral sparse purification network, and achieving deep time-frequency-space semantic alignment through a heterogeneity-aware cross-current interaction mechanism. This addresses the problems of strong background noise interference, non-stationary geometric distortion, and easy loss of feature heterogeneity in rolling bearing vibration signals under complex operating conditions, which lead to difficulties in extracting weak fault features and low fault diagnosis accuracy. Background Technology

[0002] With the increasing size and speed of rotating machinery, fault diagnosis technology for rolling bearings has become a core means of ensuring the safe operation of equipment. However, traditional methods have significant limitations: traditional rolling bearing fault diagnosis methods usually rely on manual feature extraction, which is time-consuming and highly dependent on expert experience; at the same time, traditional convolutional neural networks, which use static convolution weights, lack the ability to adaptively capture temporal geometric distortions in non-stationary signals, and the widely used statistical normalization strategy tends to smooth the energy heterogeneity between feature channels, making early weak fault features easily masked by low-frequency trends; in addition, existing frequency domain analysis methods lack an active discrimination mechanism for spectral saliency, making it difficult to accurately screen key fault frequency bands under strong background noise coupling, which restricts the discrimination capability of bearing fault diagnosis under complex operating conditions. To address the aforementioned issues, this invention proposes a bearing fault diagnosis method based on cross-current feature enhancement and graph topology interaction. This invention achieves dynamic calibration of time-varying impact features and multi-resolution graph topology construction by constructing a dynamic kernel evolution network. Simultaneously, a spectral sparsity purification network is designed to adaptively filter and enhance noise resistance in key frequency bands. Finally, a heterogeneity-sensing cross-current interaction mechanism is proposed, achieving deep alignment of temporal, frequency, and spatial semantics through latent feature recalibration and graph topology inference, significantly improving fault diagnosis capabilities. Summary of the Invention

[0003] This invention addresses the problems of existing rolling bearing fault diagnosis methods under strong noise and non-stationary operating conditions, such as rigid static feature extraction, the submergence of weak features due to normalization, and insufficient frequency domain noise immunity. It proposes a bearing fault diagnosis method based on cross-current feature enhancement and graph topology interaction. The specific implementation steps of this invention are as follows: 1. Construct a rolling bearing fault test platform consisting of a servo motor, a driver, and a manual hydraulic cylinder. Apply a 100N radial force to the bearing using the manual hydraulic cylinder to simulate the load in the actual production environment. Use vibration sensors to collect the original vibration signals of the bearing under inner ring fault, rolling element fault, combined fault, and healthy conditions. For the collected vibration signals, calculate the mean and standard deviation for standardization. Use sliding window technology to divide the continuous time domain signal into a fixed-length sample sequence.

[0004] 2. Construct a dynamic kernel evolution network to achieve dynamic calibration of temporal features and multi-resolution graph topology construction. The steps are as follows: (2a) Generate dynamic calibration coefficients For the input feature sequence Where B is the batch size, C is the number of feature channels, and T is the sequence length, global and local context descriptors are extracted separately. First, global average pooling is performed on the input features in the time dimension to obtain the global context descriptor g: Where T represents the time length of the feature sequence. Let represent the feature vector at time step t; then, a linear mapping function is used to project the global context and fuse it with the local features. in Let X be the local feature vector of the input sequence X at time t. This means projecting the global descriptor g onto a linear fully connected layer with the same dimensions as the local features. This represents the fused contextual features at time t. Dynamic calibration coefficients are generated based on fusion features. , in yes Activation function This refers to the convolutional layer used for dimensionality restoration to recover the number of channels. Represents the ReLU activation function. This represents a convolutional layer used for dimensionality reduction. The dynamic calibration coefficients generated at time t; (2b) Perform dynamic convolution operation The convolution kernel at time t It is decomposed into the tensor product of the baseline weights and the dynamic calibration coefficients. in Represents the static baseline convolution weights. This indicates a broadcast multiplication operation. The adaptive dynamic convolution kernel is generated as the input at time t; subsequently, the input signal is convolved using the dynamic convolution kernel. in This represents the convolution operation. The output characteristics after dynamic calibration; (2c) Constructing a multi-resolution graph topology After obtaining multi-resolution time series embeddings through a dynamic kernel evolution network, a set of resolution-related graph structures is constructed for each resolution level. ,set up This is the feature representation output at this resolution. ,in Let represent the time length at the m-th resolution. To construct a channel-based node feature representation, the embedding tensor is rearranged in dimensions to obtain the node feature matrix. Construct the corresponding graph based on the node feature matrix The connections between nodes are represented by a learnable adjacency matrix. Perform implicit modeling. in This indicates a dimension permutation operation, which involves adjusting the order of dimensions in a tensor. Each row corresponds to the temporal characteristics of a channel at that resolution.

[0005] 3. Construct a spectral sparse purification network to achieve adaptive screening and noise reduction enhancement of key frequency bands. The steps are as follows: (3a) Pre-filtering and frequency domain transformation First, a gated convolutional structure is introduced as a pre-filter. The gated linear units are then used to initially refine the features Z extracted by the dynamic kernel evolution network, resulting in the pre-filter features. , in Indicates linear projection. This represents a gated convolution operation with a 1×1 kernel; subsequently, a frequency domain transformation is performed on the features, mapping the time-domain features to the frequency domain space. in This indicates a feature unpacking operation used to construct local image patches. This represents the Fast Fourier Transform, used to transform the expanded features from the time domain to the frequency domain. The generated frequency domain feature spectrum; (3b) Quantization matrix selection and frequency domain reconstruction A learnable quantization matrix Q is constructed that is adaptively aligned with the spectral dimension. Frequency domain features are then filtered element-wise with weights. Each element in the quantization matrix acts as a soft threshold switch, adaptively enhancing fault-related key frequency bands and suppressing background noise. Where Q is a learnable quantized weight matrix, This indicates element-wise multiplication. This represents the inverse fast Fourier transform, used to reconstruct the filtered spectrum back to the time domain. Finally, the feature dimension is restored through an inverse folding operation, and residual connections and the GEGLU activation function are introduced to output high signal-to-noise ratio frequency domain enhanced features. , in This indicates a reverse folding operation.

[0006] 4. Design a heterogeneity-aware cross-flow interaction mechanism to achieve deep alignment of time-frequency-spatial semantics. Combine this with a graph Transformer to realize deep interaction of cross-flow features in a multi-resolution graph topology. The steps are as follows: (4a) Temporal latent feature recalibration To address the issue of low-frequency trends masking subtle fault impacts, a first-order finite-difference approach is introduced at the front end of attention computation. Firstly, an appropriate padding operation is introduced in the temporal embedding dimension to obtain the feature representation output at the M-th resolution. Then, the first-order finite difference between adjacent time steps is calculated. in This represents the first difference of the sequence at time step t. , For the feature values ​​corresponding to the time positions, this operation is physically equivalent to an adaptive high-pass filter, eliminating low-frequency trend terms between adjacent time steps; subsequently, the attention distribution is calculated based on the difference features, and the query, key, and value matrices are defined. in , , These are the learnable weight matrices for the query, key, and value, respectively; the attention weighting result is calculated using a scaled dot product attention mechanism. , Where d is the scaling factor. The normalized exponential function is then used; a linear mapping is then applied to the attention output to obtain the attention features. , in This represents the linear layer of the final output; finally, the difference features and attention features are residually connected and input into the latent feature recalibration function. In this process, the final time-domain branch output is obtained. , in, It is a nonlinear recalibration function, specifically defined as follows: in Channel-aware scaling vector. , For affine transformation parameters, For learnable bias terms; (4b) Cross-stream fusion and graph topology interaction Temporal recalibration features With frequency domain quantization characteristics By fusing the data, a comprehensive representation containing information from multiple perspectives is obtained. , Subsequently, a local attention mechanism is introduced into graph structure sequence modeling. To accurately characterize the dependencies between nodes and their neighborhoods, an adaptive transformation of the input nodes is first performed through a nonlinear recalibration strategy at the feature dimension level. Specifically, given the representation of a node p... The result after latent feature recalibration is as follows: in The nonlinear recalibration function defined in step (4a) is used here to enhance the discriminative power of node features before spatial interaction; based on this, node p and its neighbors... Local attention weights It is expressed as follows: in , These are the node features after latent feature recalibration. Let p be the set of neighbors of node p. Represents the similarity function. Learnable biases for local attention; spatial attention features derived from this. Each node feature it contains The formula is obtained by weighted aggregation of neighboring nodes, as follows: in express The feature vector of node p is obtained; finally, a graph convolutional network is used to aggregate the features of the multi-resolution graph structure constructed in step (2c). in Represents a non-linear activation function. It is a normalized adjacency matrix. These are the learnable weights of the graph convolutional layer. This represents the output representation of the graph convolution at the Mth resolution. (4c) Multiscale aggregation and fault classification Introducing an average pooling strategy for all resolution levels The graph convolution outputs are unified and integrated to obtain a global multi-scale representation. , in The fused global multi-scale feature representation is represented by m, where m is the total number of levels in the resolution map structure. Finally, the fused representation is... Input a fully connected layer and a Softmax classifier, output the fault category classification results for the rolling bearing. in It is a learnable linear layer. This represents the predicted probability distribution of the fault category.

[0007] 5. By integrating dynamic nuclear evolution network, spectral sparse purification network and heterogeneity sensing cross-current interaction mechanism, a bearing fault diagnosis method model based on cross-current feature enhancement and graph topology interaction is formed. The vibration signal to be tested is input into the trained model and the final fault category is output, realizing intelligent diagnosis of rolling bearing faults under complex working conditions.

[0008] The method of the present invention has the following advantages: (1) Dynamic calibration and instantaneous capture function with time-varying characteristics: A dynamic kernel evolution network is constructed, which uses global and local time context as prior knowledge to guide the adaptive evolution of convolutional kernel weights. This network endows the model with the ability to respond instantaneously to changes in the content of the input signal, and can keenly capture the fleeting fault impact under non-stationary conditions and adapt to the temporal geometric distortion of the signal. It simultaneously constructs a multi-resolution graph topology structure representing the operating state of the equipment, thereby effectively overcoming the problems of lag and insufficient generalization ability of traditional static convolution when processing non-stationary signals.

[0009] (2) Discriminative frequency domain screening and strong noise resistance: Construct a spectral sparse purification network and establish a spectral saliency evaluation mechanism based on a learnable quantization matrix. This mechanism acts as a soft threshold switch, which can adaptively purify the key frequency bands carrying fault information, while strongly suppressing and filtering broadband background noise, thus solving the dilemma of spectral aliasing caused by strong background noise coupling and difficulty in extracting weak fault features.

[0010] (3) Achieving deep alignment between heterogeneity perception and time-frequency-space semantics: A heterogeneity perception cross-current interaction mechanism is proposed. Through the temporal latent feature recalibration strategy, on the basis of inheriting the numerical stability advantage of statistical normalization, it makes up for its potential limitations in characterizing fine-grained feature differences. Nonlinear recalibration is used to accurately restore the energy heterogeneity between feature dimensions. Combined with the adaptive graph Transformer to drive cross-current features to perform spatial dynamic interaction in multi-resolution graph topology, deep semantic alignment of "time-frequency-space" three-dimensional information is achieved, which significantly improves the model's discrimination accuracy for early weak faults under complex working conditions. Attached Figure Description

[0011] Figure 1 This is the overall framework diagram of the present invention. Figure 2 This is a classification result using a confusion matrix from a random experiment on the AUST dataset. Figure 3 This is a classification result diagram based on the confusion matrix of a random experiment on the CWRU dataset. Detailed Implementation

[0012] The specific implementation steps of this invention are as follows: 1. Data Preprocessing Data Acquisition and Standardization: A rolling bearing fault experimental platform consisting of a servo motor, a driver, and a manual hydraulic cylinder was built. A 100N radial force was applied by the manual hydraulic cylinder to simulate the actual load. The original vibration signals of the bearing in four states—inner ring, rolling element, combined, and healthy—were collected. The mean and standard deviation of the training set signals were calculated, and the original signals were normalized to improve the stability of the model.

[0013] Sample acquisition: The standardized vibration signal is segmented into time series data using a sliding window of fixed length. The window length is set to 1024 and the step size is 512 to ensure that the samples cover the complete vibration cycle and achieve data augmentation. Finally, the data is divided into training set and test set according to the ratio.

[0014] 2. Obtain time-varying dynamic features and multi-resolution graph topology. By constructing a dynamic kernel evolution network, a deep understanding of the non-stationary time-varying characteristics of the input signal is achieved. Specifically, global average pooling and local convolution are used in parallel to extract temporal context prior descriptors, which guide the adaptive evolution of the convolutional kernel weights. This allows the feature extraction strategy to be dynamically adjusted according to the real-time changes of the input signal, thereby keenly capturing transient impacts and geometric distortions under complex operating conditions in the time domain. Simultaneously, the extracted high-dimensional dynamic feature channels are mapped to semantic nodes, and a learnable adjacency matrix is ​​used to construct a multi-resolution graph topology representing the device's operating state, providing a structural foundation for subsequent spatial interactions.

[0015] 3. Frequency domain filtering and enhancement using spectral sparse purification networks. The Spectral Sparse Refinement Network aims to address the spectral aliasing problem in noisy environments. The network first maps temporal features to the frequency domain using gated convolution and Fast Fourier Transform. Then, it establishes a saliency evaluation mechanism based on a learnable quantization matrix, utilizing a soft threshold switch to achieve high-fidelity refinement and denoising of critical fault information. Finally, through Inverse Fast Fourier Transform and residual connections, the refined spectrum is reconstructed into high signal-to-noise ratio temporal enhancement features.

[0016] 4. Heterogeneity-aware cross-flow interaction mechanism Unlike traditional statistical normalization methods that easily lead to feature homogenization, the heterogeneity-aware cross-current interaction mechanism is a deep semantic alignment paradigm. First, a latent feature recalibration strategy is introduced in the temporal branch, using first-order finite difference and nonlinear activation functions to reconstruct the feature distribution, accurately restoring the energy heterogeneity of the feature dimensions while inheriting the numerical stability of normalization. Then, the recalibrated temporal features are fused with frequency-domain enhanced features and input into an adaptive graph Transformer. Through a local attention mechanism and a graph convolutional network, cross-current features are driven to perform spatial dynamics interaction in a multi-resolution graph topology, achieving deep fusion of "time-frequency-space" three-dimensional semantics and fault detection.

[0017] The effectiveness of this invention was further verified through the following experiments: On the self-built AUST rolling bearing dataset and the publicly available CWRU dataset, the accuracy of the proposed method reached 93.31% and 95.71% respectively under -5dB high noise conditions, significantly outperforming the comparison models (OverLoCK, CAT-KD, and ConvFormer, etc.). Specifically, on the CWRU dataset, it outperformed the suboptimal model by more than 10 percentage points. In a 0dB noise environment, the accuracy of the proposed method recovered to 95.06% on the AUST dataset and reached 98.57% on the CWRU dataset. Ablation experiments confirmed the necessity of synergy among the modules. Removing the spectral sparse purification network module significantly decreased the model accuracy by 9.21%, replacing the dynamic kernel evolution network module caused a sharp drop in accuracy of 15.07%, and removing the latent feature recalibration strategy in the heterogeneity-aware cross-current interaction mechanism resulted in a performance decrease of 8.11%. These experiments fully validated the absolute advantage of the bearing fault diagnosis method based on cross-current feature enhancement and graph topology interaction in feature extraction and noise decoupling. Through the above experiments, the fault diagnosis effect of the proposed method was further verified.

Claims

1. A bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction, comprising the following steps: (1) Collect the original vibration signals of the bearing under different health conditions on the rolling bearing failure test platform; (2) Construct a dynamic kernel evolution network to extract temporal dynamic features from the signal and construct a multi-resolution graph topology; (3) Construct a spectrum sparse purification network to achieve adaptive screening and noise reduction of key frequency bands; (4) Design a heterogeneity-sensing cross-stream interaction mechanism to achieve deep alignment between the temporal latent feature recalibration module and the temporal-frequency-spatial semantics; (5) The dynamic nuclear evolution network, the spectrum sparse purification network and the heterogeneity perception cross-current interaction mechanism are integrated to form a bearing fault diagnosis method model based on cross-current feature enhancement and graph topology interaction, so as to realize intelligent diagnosis of rolling bearing faults.

2. The bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction according to claim 1, characterized in that, Step (2) involves constructing a dynamic kernel evolution network to extract temporal dynamic features from the signal and constructing a multi-resolution graph topology, including the following steps: (2a) Generate dynamic calibration coefficients For the input feature sequence Where B is the batch size, C is the number of feature channels, and T is the sequence length, global and local context descriptors are extracted separately. First, global average pooling is performed on the input features in the time dimension to obtain the global context descriptor g: Where T represents the time length of the feature sequence. Let represent the feature vector at time step t; then, a linear mapping function is used to project the global context and fuse it with the local features. in Let X be the local feature vector of the input sequence X at time t. This means projecting the global descriptor g onto a linear fully connected layer with the same dimensions as the local features. This represents the fused contextual features at time t. Dynamic calibration coefficients are generated based on fusion features. , in yes Activation function This refers to the convolutional layer used for dimensionality restoration to recover the number of channels. Represents the ReLU activation function. This represents a convolutional layer used for dimensionality reduction. The dynamic calibration coefficients generated at time t; (2b) Perform dynamic convolution operation The convolution kernel at time t It is decomposed into the tensor product of the baseline weights and the dynamic calibration coefficients. in Represents the static baseline convolution weights. This indicates a broadcast multiplication operation. The adaptive dynamic convolution kernel is generated as the input at time t; subsequently, the input signal is convolved using the dynamic convolution kernel. in This represents the convolution operation. The output characteristics after dynamic calibration; (2c) Constructing a multi-resolution graph topology After obtaining multi-resolution time series embeddings through a dynamic kernel evolution network, a set of resolution-related graph structures is constructed for each resolution level. ,set up This is the feature representation output at this resolution. ,in Let represent the time length at the m-th resolution. To construct a channel-based node feature representation, the embedding tensor is rearranged in dimensions to obtain the node feature matrix. Construct the corresponding graph based on the node feature matrix The connections between nodes are represented by a learnable adjacency matrix. Perform implicit modeling. in This indicates a dimension permutation operation, which involves adjusting the order of dimensions in a tensor. Each row corresponds to the temporal characteristics of a channel at that resolution.

3. The bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction according to claim 1, characterized in that, The steps for constructing the spectral sparse purification network in step (3) to achieve adaptive screening and noise reduction of key frequency bands are as follows: (3a) Pre-filtering and frequency domain transformation First, a gated convolutional structure is introduced as a pre-filter. The gated linear units are then used to initially refine the features Z extracted by the dynamic kernel evolution network, resulting in the pre-filter features. , in Indicates linear projection. This represents a gated convolution operation with a 1×1 kernel; subsequently, a frequency domain transformation is performed on the features, mapping the time-domain features to the frequency domain space. in This indicates a feature unpacking operation used to construct local image patches. This represents the Fast Fourier Transform, used to transform the expanded features from the time domain to the frequency domain. The generated frequency domain feature spectrum; (3b) Quantization matrix selection and frequency domain reconstruction A learnable quantization matrix Q is constructed that is adaptively aligned with the spectral dimension. Frequency domain features are then filtered element-wise. Each element in the quantization matrix acts as a soft threshold switch, adaptively enhancing fault-related key frequency bands and suppressing background noise. Where Q is a learnable quantized weight matrix, This indicates element-wise multiplication. This represents the inverse fast Fourier transform, used to reconstruct the filtered spectrum back to the time domain. Finally, the feature dimension is restored through an inverse folding operation, and residual connections and the GEGLU activation function are introduced to output high signal-to-noise ratio frequency domain enhanced features. , in This indicates a reverse folding operation.

4. The bearing fault diagnosis method based on cross-flow feature enhancement and graph topology interaction according to claim 1, characterized in that, The design of the heterogeneity-aware cross-stream interaction mechanism described in step (4) to achieve deep alignment between the temporal latent feature recalibration module and the temporal-frequency-spatial semantics is carried out as follows: (4a) Temporal latent feature recalibration To address the issue of low-frequency trends masking subtle fault impacts, a first-order finite-difference approach is introduced at the front end of attention computation. First, an appropriate padding operation is introduced in the temporal embedding dimension to obtain the feature representation output at the M-th resolution. Then, the first-order finite difference between adjacent time steps is calculated. in This represents the first difference of the sequence at time step t. , For the feature values ​​corresponding to the time positions, this operation is physically equivalent to an adaptive high-pass filter, which eliminates the low-frequency trend terms between adjacent time steps; Subsequently, the attention distribution is calculated based on the differential features, the query, key, and value matrices are defined, and the attention weighting result is calculated using the scaled dot product attention mechanism. , Where d is the scaling factor. The normalized exponential function is then used; a linear mapping is then applied to the attention output to obtain the attention features. , in This represents the linear layer of the final output; finally, the difference features and attention features are residually connected and input into the latent feature recalibration function. In this process, the final time-domain branch output is obtained. , in, It is a nonlinear recalibration function, specifically defined as follows: in Channel-aware scaling vector. , For affine transformation parameters, For learnable bias terms; (4b) Cross-stream fusion and graph topology interaction Temporal recalibration features With frequency domain quantization characteristics By fusing the data, a comprehensive representation containing information from multiple perspectives is obtained. , Subsequently, a local attention mechanism is introduced into graph structure sequence modeling. To accurately characterize the dependencies between nodes and their neighborhoods, an adaptive transformation of the input nodes is first performed using a nonlinear recalibration strategy at the feature dimension level. Specifically, given the representation of a node p... The result after latent feature recalibration is as follows: in The nonlinear recalibration function defined in step (4a) is used here to enhance the discriminative power of node features before spatial interaction; based on this, node p and its neighbors... Local attention weights It is expressed as follows: in , These are the node features after latent feature recalibration. Let p be the set of neighbors of node p. Represents the similarity function. Learnable biases for local attention; spatial attention features derived from this. Each node feature it contains The formula is obtained by weighted aggregation of neighboring nodes, as follows: in express The feature vector of node p is obtained; finally, a graph convolutional network is used to aggregate the features of the multi-resolution graph structure constructed in step (2c). in Represents a non-linear activation function. It is a normalized adjacency matrix. These are the learnable weights of the graph convolutional layer. This represents the output representation of the graph convolution at the Mth resolution. (4c) Multiscale aggregation and fault classification Introducing an average pooling strategy for all resolution levels The graph convolution outputs are unified and integrated to obtain a global multi-scale representation. , in The fused global multi-scale feature representation is represented by m, where m is the total number of levels in the resolution map structure. Finally, the fused representation is... Input a fully connected layer and a Softmax classifier, and output the failure category classification results for the rolling bearing.

Images