Bearing fault diagnosis method based on dynamic hypergraph convolution and spatial-temporal feature fusion

Abstract

The invention provides a bearing fault diagnosis method based on dynamic hypergraph convolution and spatial-temporal feature fusion. The method comprises the following steps: acquiring a training data set; the training data set comprises a vibration signal and a fault type; constructing a bearing fault diagnosis model based on dynamic hypergraph convolution and spatial-temporal feature fusion; the bearing fault diagnosis model comprises a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatial-temporal feature fusion module and a full-connection classification layer; training the bearing fault diagnosis model based on the training data set; and inputting a to-be-diagnosed vibration signal into the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatial-temporal feature fusion to obtain a bearing fault type. According to the bearing fault diagnosis method based on dynamic hypergraph convolution and spatial-temporal feature fusion, the problem that the bearing fault diagnosis accuracy is low due to the fact that an existing bearing fault diagnosis method is insufficient in the aspects of multi-damage-degree distinguishing, dynamic feature correlation modeling and physical rule fusion is solved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis, specifically relating to a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion. Background Technology

[0002] Mechanical condition monitoring and fault diagnosis technology, as a key component of industrial equipment health management, focuses on achieving early warning and accurate identification of faults through data analysis. In the field of bearing fault diagnosis, traditional methods mainly rely on the time-domain statistical characteristics (such as kurtosis and peak factor) and frequency-domain characteristics (such as spectral analysis and wavelet decomposition) of vibration signals, combined with shallow classification models (such as SVM and random forest) to achieve fault classification. With the development of deep learning technology, feature extraction methods based on convolutional neural networks (CNN) and recurrent neural networks (RNN) have gradually become mainstream.

[0003] However, existing technologies still face significant bottlenecks: First, in industrial multi-condition scenarios, vibration signals are easily distorted in the time and frequency domain due to factors such as load fluctuations and noise interference, limiting the generalization ability of traditional feature extraction methods. Second, while existing deep learning methods can automatically extract features, they are mostly static models that fail to fully explore the dynamic correlations between fault modes and the coupling relationships between multi-scale spatiotemporal features, making it difficult to effectively distinguish fault types with similar damage levels (such as rolling element faults with different degrees of expansion). Third, in practical applications, bearing damage evolution exhibits physical regularities (such as the degree of damage increasing with fatigue accumulation), but existing models lack the integration of physical constraints, leading to diagnostic results that may not conform to the equipment degradation mechanism. To address these issues, existing research has attempted to optimize feature weight allocation by introducing attention mechanisms or graph neural networks, but these have not solved the sparsity problem in modeling high-order relationships of fault samples, and the robustness of the models under complex conditions still needs improvement.

[0004] Therefore, given the shortcomings of existing bearing fault diagnosis methods in distinguishing multiple damage levels, modeling dynamic feature correlations, and integrating physical laws, the accuracy of bearing fault diagnosis is low. How to improve the accuracy of bearing fault diagnosis is an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion. This invention, based on dynamic hypergraph convolution and spatiotemporal feature fusion, obtains a training dataset including vibration signals and fault types. A bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion is constructed. This model includes a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module, and a fully connected classification layer. The bearing fault diagnosis model is trained based on the training dataset. The vibration signal to be diagnosed is input into the trained bearing fault diagnosis model to obtain the bearing fault type. This invention, by employing a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion, solves the problem of low accuracy in bearing fault diagnosis caused by the deficiencies of existing methods in distinguishing multiple damage levels, modeling dynamic feature correlations, and fusing physical laws.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention proposes a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion, characterized by the following steps:

[0008] S1. Obtain the training dataset; the training dataset includes vibration signals and fault types, wherein the vibration signals are signals obtained from the accelerometer at the drive end;

[0009] S2. Construct a bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion;

[0010] The bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion includes: a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module, and a fully connected classification layer;

[0011] S3. The bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion is trained based on the training dataset to obtain the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion.

[0012] S4. Obtain the vibration signal to be diagnosed, and input the vibration signal to be diagnosed into the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion to obtain the bearing fault type.

[0013] Furthermore, the multi-scale feature fusion module includes a seven-dimensional signal feature extraction module, a one-dimensional convolutional neural network (1D-CNN) feature extraction module, a linear fully connected layer, and a Transformer encoder;

[0014] The vibration signal is input to the multi-scale feature fusion module to obtain temporal features; specifically including:

[0015] The vibration signal is input to the seven-dimensional signal feature extraction module and the 1D-CNN feature extraction module respectively to obtain a seven-dimensional physical feature vector and deep temporal features. The seven-dimensional physical feature vector includes temporal features, frequency domain features, and time-frequency features. The temporal features include kurtosis and impact factor; the frequency domain features include wavelet energy entropy; and the time-frequency features include STFT energy distribution and Hilbert envelope mean. The Hilbert envelope mean is positively correlated with the degree of damage. The STFT energy distribution is the time-frequency energy feature E. STFT The first three values ​​of the energy sequence;

[0016] The seven-dimensional physical feature vector and the deep temporal feature are input into a linear fully connected layer to obtain the fused features;

[0017] The fused features are input into the Transformer encoder to obtain the temporal features.

[0018] Furthermore, the dynamic hypergraph learning model includes: a similarity network module, a dynamic hypergraph construction module, and a hypergraph convolutional layer that integrates a multi-head attention mechanism;

[0019] The deep temporal features are input into the dynamic hypergraph learning model to obtain spatial features; specifically including:

[0020] The deep temporal features are input into the feature projection module of the similarity network module, and the feature projection is obtained through two fully connected layers. The sample similarity matrix is ​​calculated based on cosine similarity.

[0021] In the dynamic hypergraph construction module, based on the similarity matrix, the k-NN algorithm is used to dynamically select the 8 nearest neighbor nodes of each sample to generate a node-hyperedge association matrix;

[0022] The node-hyperedge correlation matrix is ​​input into the hypergraph convolutional layer with multi-head attention mechanism to obtain spatial features. In the hypergraph convolutional layer with multi-head attention mechanism, information transmission is completed based on the multi-head attention mechanism. The hypergraph Laplacian operator is normalized by the node degree matrix and the hyperedge degree matrix to ensure the scale consistency of feature transmission.

[0023] Furthermore, the temporal features output by the multi-scale feature fusion module and the spatial features output by the dynamic hypergraph learning model are input into the spatiotemporal feature fusion module for concatenation. Cross-domain feature fusion and dimensionality reduction are achieved through a fully connected layer, and a nonlinear transformation is introduced using the ReLU activation function to generate the final feature representation.

[0024] Furthermore, the final feature representation is input into the fully connected classification layer, and the fault type is output through the Softmax activation function.

[0025] Furthermore, the bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion also includes a physical constraint optimization module. By embedding the physical constraint optimization mechanism, a learnable three-dimensional tensor is embedded in the convolutional layer of the model to model the severity of different damage categories. During the model training process, a damage degree constraint term is added, and the constraint function forces adjacent damage degree parameters to maintain an interval of 0.1 to ensure that the diagnosis results conform to the physical laws of equipment degradation.

[0026] Compared with existing technologies, it has the following beneficial effects:

[0027] 1. This invention employs a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion. The bearing fault diagnosis model, which includes a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module, and a fully connected classification layer, performs bearing fault diagnosis. This model concatenates the temporal features output from the multi-scale feature fusion module and the spatial features output from the dynamic hypergraph learning model into the spatiotemporal feature fusion module, achieving coupling of multi-scale spatiotemporal features. This enhances the model's ability to distinguish fault types with similar damage levels, thereby improving the accuracy of bearing fault diagnosis.

[0028] 2. The bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion of the present invention realizes physical constraint embedding optimization through a physical constraint optimization module, introduces damage degree monotonicity constraint in backpropagation, and solves the problem of misjudgment of damage degree classification by constructing a learnable three-dimensional tensor to model the severity of different damage categories, ensuring that the diagnosis results are consistent with the equipment degradation mechanism and improving the accuracy of bearing fault diagnosis. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion provided in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the architecture of a bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion provided in an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of a multi-scale feature fusion module provided in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of a Transformer encoder provided in an embodiment of the present invention.

[0034] Figure 5 This is a schematic diagram of the dynamic hypergraph learning model provided in an embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram of the spatiotemporal feature fusion module provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0038] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0039] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0040] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0041] This invention proposes a bearing fault diagnosis method based on the fusion of dynamic hypergraph convolution and spatiotemporal features. For example... Figure 1 As shown, the bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion includes the following steps S1 to S4.

[0042] S1. Obtain the training dataset; the training dataset includes vibration signals and fault types, wherein the vibration signals are signals obtained from the accelerometer at the drive end.

[0043] Furthermore, vibration signals were acquired from the accelerometer at the drive end, and raw data was collected at a sampling frequency of 12kHz. A sliding window segmentation technique was employed, truncating a signal segment of 2048 points with a step size of 256 points, and cyclic oversampling was performed on minority class samples. Adaptive Gaussian noise of 25-35dB was added to enhance the model's robustness.

[0044] S2. Construct a bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion;

[0045] The bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion includes: a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module, and a fully connected classification layer.

[0046] S3. The bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion is trained based on the training dataset to obtain the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion.

[0047] In one specific implementation, the initial convergence period uses a fixed learning rate to train the base model, the fine adjustment period initiates cosine annealing learning rate scheduling and enhances the physical constraint strength, the stabilization convergence period freezes the feature extraction layer parameters, optimizes the classification head separately and performs hard sample retraining.

[0048] S4. Obtain the vibration signal to be diagnosed, and input the vibration signal to be diagnosed into the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion to obtain the bearing fault type.

[0049] Specifically, the bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion includes: a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module, and a fully connected classification layer; the architecture diagram of the bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion is shown below. Figure 2 As shown.

[0050] 1. Multi-scale feature fusion module

[0051] The multi-scale feature fusion module includes a seven-dimensional signal feature extraction module, a one-dimensional convolutional neural network (1D-CNN) feature extraction module, a linear fully connected layer, and a Transformer encoder; a schematic diagram of the multi-scale feature fusion module is shown below. Figure 3 As shown.

[0052] The vibration signal is input to the multi-scale feature fusion module to obtain temporal features; specifically, this includes: inputting the vibration signal to the seven-dimensional signal feature extraction module and the 1D-CNN feature extraction module respectively to obtain a seven-dimensional physical feature vector and deep temporal features. The seven-dimensional physical feature vector includes: temporal features, frequency domain features, and time-frequency features. The temporal features include kurtosis and impact factor; the frequency domain features include wavelet energy entropy; and the time-frequency features include STFT energy distribution and Hilbert envelope mean. The STFT energy distribution is the time-frequency energy feature E. STFT The first three values ​​of the energy sequence; the seven-dimensional physical feature vector and the deep temporal feature are input into a linear fully connected layer to obtain the fused feature; the fused feature is input into a Transformer encoder to obtain the temporal feature.

[0053] Specifically, the multi-scale feature fusion module constructs a three-level processing chain of "dual-branch feature extraction → heterogeneous fusion → global correlation modeling". It complements the interpretability of the seven-dimensional physical feature vector (kurtosis, impact factor, etc.) extracted by time-frequency domain analysis with the deep expression capability of CNN automatic features. Combined with the long-distance dependency capture of Transformer, it achieves accurate identification of complex bearing faults.

[0054] The vibration signal is input to the seven-dimensional signal feature extraction module to obtain a seven-dimensional physical feature vector. The specific calculation method of the seven-dimensional physical feature vector is as follows:

[0055] The time-domain features include kurtosis and impulse factor. Kurtosis is calculated using the fourth-order central moment to determine the degree to which the signal deviates from a normal distribution. The formula is as follows:

[0056]

[0057] Where N is the number of signal sampling points, and μ and σ are the mean and standard deviation of the signal, respectively.

[0058] The impact factor reflects the ratio between the signal peak value and the average amplitude, and is calculated using the following formula:

[0059]

[0060] Where x is the overall sequence of the original signal, and N is the total number of sampling points of the signal.

[0061] Frequency domain features are then analyzed using Discrete Wavelet Transform (DWT) to decompose the signal into different frequency bands, calculating the energy and energy entropy of each band to effectively characterize the complexity of the signal's frequency domain distribution. The formula for calculating wavelet energy entropy is:

[0062]

[0063] Where M is the number of wavelet decomposition subbands, c k Let be the coefficient of the k-th sub-band.

[0064] Simultaneously, the time-frequency matrix is ​​obtained using the Short-Time Fourier Transform (STFT), and the energy distribution is calculated. The discrete calculation of the Short-Time Fourier Transform can be expressed as:

[0065]

[0066] Where x(n) is a vibration signal of length N, ω(·) is a Hanning window of length L, τ is the time shift index, k is the frequency index, and X(τ,k) is the time-frequency matrix.

[0067] The time-frequency energy distribution is extracted using the sum of squared moduli, focusing on the first three dimensions. The energy characteristic calculation formula is as follows:

[0068]

[0069] Only the time-frequency energy feature E in the τ dimension is taken STFT The first three values ​​of the energy sequence.

[0070] Furthermore, envelope features are extracted using Hilbert transform to comprehensively capture the time-varying characteristics of non-stationary signals. First, an analytic signal is constructed using Hilbert transform:

[0071]

[0072] Where H{x(t)} is the Hilbert transform of x(t), For the analytic signal, its imaginary part is the result of the Hilbert transform. The envelope feature is the magnitude of the analytic signal, and the final mean is taken.

[0073]

[0074] Where A(t) represents the instantaneous amplitude envelope of the signal (the amplitude profile of the fault impact), It reflects the average strength of the envelope (positively correlated with the degree of damage).

[0075] These complementary features together constitute a seven-dimensional feature space, providing a robust feature foundation for subsequent diagnosis.

[0076] Furthermore, the vibration signals are input into the 1D-CNN feature extraction module to obtain deep temporal features. Specifically, a one-dimensional convolutional neural network (1D-CNN) is used to extract deep features from the original vibration signals. The forward propagation of the CNN can be represented as:

[0077] X (l) =σ(W l *X(l-1) +b (l) )

[0078] Among them, X (l) For the output features of the l-th layer, W l b represents the kernel weights. (l) σ is the bias term, and σ is the GELU activation function.

[0079] The seven-dimensional physical feature vector and deep temporal features are input into a linear fully connected layer to obtain fused features, specifically including: the 64-dimensional temporal features extracted by the CNN. With the seven-dimensional physical eigenvector F hand =[Kurtosis,ImpulseFactor,Entropy,...] are concatenated by channel to form a 71-dimensional fusion feature F fusion = The feature fusion and dimension alignment are then achieved through a linear fully connected layer. The parameters of the linear layer are initialized using the Xavier strategy to ensure that the physical interpretability of the handcrafted features is not "drowned out" by the deep learning features after fusion.

[0080] The fused features are input into a Transformer encoder to obtain temporal features. Specifically, the fused features are reshaped into pseudo-sequences, which are then input into a single-layer eight-head attention Transformer encoder. The Transformer encoder is as follows: Figure 4 As shown, this construction enables the Transformer to capture the global correlation of multi-scale features within a sample (such as the potential connection between deep patterns extracted by CNN and the physical semantics of handcrafted features). The eight-head attention mechanism enhances the ability to capture long-distance dependencies by computing attention weights in parallel (such as the difference in correlation patterns between kurtosis and CNN features under different damage levels). Although the dimension of the final output feature vector remains unchanged, it contains the collaborative correlation information of multi-scale features, providing a more robust pattern representation for subsequent fault classification.

[0081] 2. Dynamic Hypergraph Learning Model

[0082] The dynamic hypergraph learning model includes: a similarity network module, a dynamic hypergraph construction module, and a hypergraph convolutional layer that integrates a multi-head attention mechanism;

[0083] The deep temporal features are input into the dynamic hypergraph learning model to obtain spatial features. Specifically, this includes: inputting the deep temporal features into the feature projection module of the similarity network module, calculating the feature projection through two fully connected layers, and calculating the sample similarity matrix based on cosine similarity; inputting the similarity matrix into the dynamic hypergraph construction module, dynamically selecting the 8 nearest neighbor nodes of each sample based on the similarity matrix using the k-NN algorithm to generate a node-hyperedge association matrix; inputting the node-hyperedge association matrix into the hypergraph convolutional layer with a multi-head attention mechanism to obtain spatial features. In the hypergraph convolutional layer with a multi-head attention mechanism, information transmission is completed based on the multi-head attention mechanism, and the hypergraph Laplacian operator is normalized through the node degree matrix and the hyperedge degree matrix to ensure the scale consistency of feature transmission.

[0084] The dynamic hypergraph learning model achieves global modeling of fault features by constructing an adaptive sample association topology. Its complete process is as follows: Figure 5 As shown. The specific process is as follows:

[0085] The deep temporal features are input into the feature projection module of the similarity network module. Feature projection is obtained through two fully connected layers, and the sample similarity matrix is ​​calculated based on cosine similarity. Specifically:

[0086] The deep temporal features are input into the feature projection module of the similarity network module. Dimensionality compression is achieved through two fully connected layers. The first layer uses Xavier initialization for linear transformation and introduces a non-linear mapping with the ReLU activation function. The features output from the second layer are L2 normalized and then used by the cosine similarity calculation module to generate a sample similarity matrix. The definition is as follows:

[0087]

[0088] Where, f: For feature projection networks, time-domain signals are mapped to a semantically richer feature space, S ij Characterizes the directional similarity of feature vectors between samples.

[0089] In the dynamic hypergraph construction module, based on the similarity matrix, the k-NN algorithm is used to dynamically select the 8 nearest neighbor nodes for each sample, generating a node-hyperedge association matrix, specifically:

[0090] The dynamic hypergraph construction module breaks through the limitations of traditional fixed graph structures, dynamically generating topological connections based on the K-NN algorithm with k=8. In the "Top-k selection" process, the 8 nodes with the highest similarity indices are selected from each row of the similarity matrix, and a dynamic node-hyperedge association matrix H is constructed using this matrix.

[0091]

[0092] Incremental hyperedge construction is based on the K-NN algorithm (k=8) to dynamically generate hyperedges. Each hyperedge connects the central sample and its nearest neighbor, forming a node-hyperedge association matrix H∈{0,1}. N×E Where N is the number of nodes and E is the number of hyperedges (set to the same as the number of nodes here), and a value of 1 indicates that the hyperedge contains the center node v. j It also includes its eight nearest neighbors. This dynamic update mechanism allows the topology to adapt to changes in feature distribution during training, reducing redundant connections by 37% compared to a fixed graph structure.

[0093] The node-hyperedge correlation matrix is ​​input into the hypergraph convolutional layer with multi-head attention mechanism to obtain spatial features. In the hypergraph convolutional layer with multi-head attention mechanism, information transmission is completed based on the multi-head attention mechanism. The hypergraph Laplacian operator is normalized by the node degree matrix and the hyperedge degree matrix to ensure the scale consistency of feature transmission.

[0094] Traditional hypergraph convolutions use the Laplacian operator for information transfer, but this invention proposes an improved structure that integrates a multi-head attention mechanism into the hypergraph convolution layer. Its forward propagation formula is as follows:

[0095]

[0096] in For the node-hyperedge incidence matrix, It is a diagonal matrix of node degree. Let W be the hyper-edge degree diagonal matrix, W be the learnable parameters, σ be the ReLU activation function, and λ be the regularization coefficient. Let λ be the node degree diagonal matrix. With hyperdiagonal matrix Generated through the "degree matrix calculation" module, used for normalizing the hypergraph Laplacian operator. Ensure scale consistency in feature propagation. Simultaneously, the four-head attention mechanism computes MultiHeadAttn(X) in parallel. (l) W) captures long-range dependencies, where scaling operations on queries, keys, and values ​​avoid gradient vanishing due to excessively large inner products. Finally, the output X of the hypergraph convolution... (l+1) Composed of a linear combination of topological signal transmission and attention enhancement features, and activated by ReLU to achieve a nonlinear transformation of the feature space, this design improves the training iteration speed of the model on the CWRU dataset by 3.8 times.

[0097] The learnable parameter W in the formula is initialized using Xavier normality to ensure the stability of the feature variance of each layer; the regularization coefficient λ = 0.3 is determined through cross-validation to balance the fusion weights of topological information and attention features. The binary nature of the node-hyperedge association matrix H (containing only 0 and 1) gives the model the advantage of sparse computation. Combined with the dynamic topology update strategy, it effectively reduces computational complexity while improving the flexibility of sample association modeling.

[0098] 3. Spatiotemporal Feature Fusion Module

[0099] The temporal features output by the multi-scale feature fusion module and the spatial features output by the dynamic hypergraph learning model are input into the spatiotemporal feature fusion module for concatenation. Cross-domain feature fusion and dimensionality reduction are achieved through a fully connected layer, and a nonlinear transformation is introduced using the ReLU activation function to generate the final feature representation.

[0100] The spatiotemporal feature fusion module achieves collaborative expression of multi-source features through cross-domain information integration. The module's structure is as follows: Figure 6 As shown.

[0101] 4. Fully connected classification layer

[0102] The final feature representation is input into the fully connected classification layer, and the fault type is output through the Softmax activation function.

[0103] Furthermore, the bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion also includes a physical constraint optimization module. By embedding the physical constraint optimization mechanism, a learnable three-dimensional tensor is embedded in the convolutional layer of the model to model the severity of different damage categories. During the model training process, a damage degree constraint term is added, and the constraint function forces adjacent damage degree parameters to maintain an interval of 0.1 to ensure that the diagnosis results conform to the physical laws of equipment degradation.

[0104] Specifically, to ensure that the diagnostic results conform to the physical laws of equipment degradation, a physical constraint optimization mechanism is embedded in the model. By embedding the physical constraint optimization mechanism, a learnable three-dimensional tensor (severity_params) is embedded in the convolutional layer of the model to model the severity of different damage categories;

[0105] To ensure the physical meaning of the damage parameters, a damage severity constraint term is added during model training. This constraint function forces adjacent damage severity parameters to maintain a 0.1 interval, ensuring that the diagnostic results conform to the physical laws of equipment degradation. The specific constraint term is as follows:

[0106]

[0107] Where δ = 0.1 is the forced interval coefficient. The parameter represents the quantization parameter of the c-th type of fault in the learnable three-dimensional tensor severity_params, where c∈{Ball,Inner,Outer} and d∈{0,1,2} correspond to mild, moderate, and severe, respectively.

[0108] Simultaneously, the network parameters are jointly optimized using a weight of α = 0.2 and a cross-entropy loss:

[0109]

[0110] This invention presents a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion. First, it proposes a dynamic hypergraph learning model, breaking through the limitations of traditional fixed graph structures. Through a real-time hyperedge reconstruction mechanism based on k-NN, the model training speed is increased by 3.8 times, effectively solving the sparsity problem in modeling high-order relationships of fault samples. Second, it achieves physical constraint embedding optimization, pioneering the introduction of damage degree monotonicity constraints in backpropagation. By constructing a learnable 3D tensor to model the severity of different damage categories, it solves the problem of misjudgment of damage degree classification, ensuring that the diagnostic results conform to the equipment degradation mechanism. In addition, it designs a multi-source feature dynamic calibration mechanism, using a gated Transformer to achieve automatic weight allocation of temporal CNN features and frequency domain handmade features, realizing the coupling of multi-scale spatiotemporal features and improving the model's ability to distinguish fault types with similar damage degrees.

[0111] The bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion described above can be implemented as a computer program that can run on a computer device.

[0112] Computer devices include processors, memory, and network interfaces connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0113] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion.

[0114] This processor provides computing and control capabilities to support the operation of the entire computer device.

[0115] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform a bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion.

[0116] This network interface is used for network communication with other devices. Those skilled in the art will understand that the above-described computer device structure is only a partial structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0117] The processor is used to run a computer program stored in the memory, which implements the bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion as described in Embodiment 1.

[0118] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0119] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0120] The present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion as described in Embodiment 1.

[0121] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0122] 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, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. 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 implementations should not be considered beyond the scope of this invention.

[0123] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0124] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0126] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A bearing fault diagnosis method based on dynamic hypergraph convolution and spatiotemporal feature fusion, characterized in that, The method comprises the steps of: S1, obtaining a training data set; the training data set comprises a vibration signal and a fault type, wherein the vibration signal is a signal obtained from a driving end acceleration sensor; S2, constructing a bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion; The bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion comprises a multi-scale feature fusion module, a dynamic hypergraph learning model, a spatiotemporal feature fusion module and a fully connected classification layer; S3, training the bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion based on the training data set to obtain a trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion; S4, obtaining a to-be-diagnosed vibration signal and inputting the to-be-diagnosed vibration signal into the trained bearing fault diagnosis model based on dynamic hypergraph convolution and spatiotemporal feature fusion to obtain a bearing fault type.

2. The method of claim 1, wherein, The multi-scale feature fusion module comprises a seven-dimensional signal feature extraction module, a one-dimensional convolutional neural network (1D-CNN) feature extraction module, a linear fully connected layer and a Transformer encoder; The vibration signal is input into the multi-scale feature fusion module to obtain a time sequence feature; specifically comprising: The vibration signal is input into the seven-dimensional signal feature extraction module and the 1D-CNN feature extraction module respectively to obtain a seven-dimensional physical feature vector and a deep time domain feature, wherein the seven-dimensional physical feature vector comprises time domain features, frequency domain features and time-frequency features, wherein the time domain features comprise kurtosis and impact factor, the frequency domain features comprise wavelet energy entropy, and the time-frequency features comprise STFT energy distribution and Hilbert envelope mean value, the Hilbert envelope mean value is positively correlated with the damage degree, and the STFT energy distribution is a time-frequency energy feature E STFT The first three values of the energy sequence; The seven-dimensional physical feature vector and the deep time domain feature are input into the linear fully connected layer to obtain a fusion feature; The fusion feature is input into the Transformer encoder to obtain a time sequence feature.

3. The method of claim 2, wherein, The dynamic hypergraph learning model comprises a similarity network module, a dynamic hypergraph construction module and a hypergraph convolution layer with a fusion multi-head attention mechanism; The deep time domain feature is input into the dynamic hypergraph learning model to obtain a spatial feature; specifically comprising: The deep time domain feature is input into a feature projection module of the similarity network module, and feature projection is performed through two fully connected layers to calculate a sample similarity matrix based on cosine similarity; In the dynamic hypergraph construction module, based on the similarity matrix, the k-NN algorithm is used to dynamically select eight nearest neighbor nodes of each sample to generate a node-hyperedge association matrix; The node-hyperedge association matrix is input into the hypergraph convolution layer with a fusion multi-head attention mechanism to obtain a spatial feature; in the hypergraph convolution layer with a fusion multi-head attention mechanism, information transmission is completed based on the fusion multi-head attention mechanism, the hypergraph Laplacian operator is normalized through a node degree matrix and a hyperedge degree matrix to ensure the scale consistency of feature transmission.

4. The method of claim 3, wherein, The time sequence feature obtained by the multi-scale feature fusion module and the spatial feature obtained by the dynamic hypergraph learning model are input into the spatiotemporal feature fusion module for splicing, cross-domain feature fusion and dimension reduction are realized through a fully connected layer, a ReLU activation function is used to introduce a nonlinear transformation to generate a final feature representation.

5. The method of claim 4, wherein, The final feature representation is input into the fully connected classification layer, and a fault type is output through a Softmax activation function.

6. The method of claim 1, wherein, The bearing fault diagnosis model based on dynamic hypergraph convolution and spatio-temporal feature fusion further comprises a physical constraint optimization module; by embedding a physical constraint optimization mechanism, a learnable three-dimensional tensor is embedded in the convolution layer of the model, which is used for modeling the severity of different damage categories; during the model training process, a damage degree constraint term is added, and an adjacent damage degree parameter is forced to maintain an interval of 0.1 through a constraint function, so as to ensure that the diagnosis result conforms to the physical law of equipment degradation.

Images