Robot joint flexible bearing fault diagnosis method driven based on physical priori

Abstract

The present application relates to a kind of robot joint flexible bearing fault diagnosis method based on physical priori driving, comprising: collecting the vibration signal of robot joint flexible bearing under different working conditions, construct the data set including labeled source domain sample set and unlabeled target domain sample set, obtain diagnostic model by training domain adversarial graph convolutional neural network;The target sample to be diagnosed is input into the diagnostic model trained, and the flexible bearing fault class recognition result is output;Domain adversarial graph convolutional neural network introduces kernel selective fusion attention module to adaptively fuse multi-scale feature response, enhances the stability and discriminability of fault feature, constructs sample instance graph and combines multi-receptive field graph convolutional network to enhance the modeling ability of structure relationship between samples;While introducing the physical priori consistency constraint based on fault feature frequency and its harmonic neighborhood energy.The present application improves the accuracy, stability and physical explainability of bearing fault diagnosis under variable working conditions.

Description

Technical Field

[0001] This invention relates to the field of bearing condition monitoring and intelligent fault diagnosis technology, specifically to a method for diagnosing faults in flexible bearings of robot joints based on physical prior driving. Background Technology

[0002] Flexible bearings are key components in robot joint modules, and their operational status directly affects the safety, reliability, and maintenance costs of the equipment. In real-world industrial scenarios, flexible bearings typically operate under complex conditions with continuously changing speeds, loads, and environmental conditions, making their vibration signals prone to significant non-stationarity and distribution drift. Once a flexible bearing fails, it not only reduces robot performance but can also cause downtime and substantial economic losses. Therefore, condition monitoring and fault diagnosis of flexible bearings under varying operating conditions has always been an important research direction in the field of intelligent operation and maintenance. Methods such as convolutional neural networks can directly learn fault characterization features from raw vibration signals, achieving good results in fixed operating conditions or scenarios with slight distribution shifts. However, when operating conditions change significantly, the statistical distribution, time-frequency scale characteristics, and similarity relationships between samples of the vibration signal all change significantly, making existing models prone to feature distortion, decision boundary shifts, and decreased generalization ability.

[0003] To address the diagnostic challenges posed by the lack of labeled samples or even the complete absence of labels in the target operating condition, unsupervised domain adaptive methods have been gradually introduced into bearing cross-operating condition fault diagnosis. Most existing methods align source and target domain features through distribution matching or adversarial learning, but these typically focus more on the consistency of global statistical distributions, neglecting the utilization of inter-sample structural relationships and intra-class topological patterns. Therefore, graph convolutional networks have been further applied to cross-operating condition diagnosis, enhancing structural propagation capabilities by explicitly modeling inter-sample relationships. Nevertheless, existing graph convolutional diagnostic methods still suffer from two shortcomings: First, graph structure generation highly depends on node features extracted by the convolutional network, and under varying operating conditions, these features are susceptible to scale fluctuations and local pattern drift, leading to unstable similarity measurements and distorted adjacency relationships. Second, existing optimization processes primarily rely on data-driven constraints such as classification supervision, domain alignment, and structural alignment, lacking explicit physical consistency constraints directly related to bearing fault mechanisms, thus resulting in deficiencies in diagnostic reliability and physical interpretability.

[0004] Furthermore, the characteristic frequencies of bearing faults and their energy responses in the envelope spectrum contain physical information closely related to the fault location, providing important prior information for fault identification. How to stably and controllably integrate fault mechanism information into the model training process under unlabeled conditions in the target domain, and how to coordinate this with cross-domain alignment and graph structure modeling for optimization, remains a pressing issue in current technology. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for diagnosing faults in flexible bearings of robot joints based on physical priors. This method solves the problems of insufficient stability of feature representation, inadequate modeling of structural relationships between samples, and lack of explicit physical constraints related to the fault mechanism in the fault diagnosis of bearings under varying operating conditions. These problems result in limited accuracy of fault identification, diagnostic stability, and interpretability of results under different operating conditions.

[0006] The technical solution adopted in this invention is as follows: This invention provides a method for diagnosing faults in flexible bearings of robot joints based on physical priors, comprising: collecting vibration signals of flexible bearings of robot joints under different working conditions, constructing a dataset including a labeled source domain sample set and an unlabeled target domain sample set, training a domain adversarial graph convolutional neural network to obtain a diagnostic model; inputting the target sample to be diagnosed into the trained diagnostic model, and outputting the flexible bearing fault category identification result. The training of the domain adversarial graph convolutional neural network includes: After preprocessing the flexible bearing vibration signals of the dataset, the signals are input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation. The enhanced feature representation is mapped to sample node features, and a sample instance graph is constructed using a graph construction layer based on the similarity between samples. The sample instance graph is then input into a multi-receptive field graph convolutional network to obtain the structural enhanced feature representation. The structure-enhanced feature representation is input into the classifier to obtain the model prediction distribution of the target domain samples, and the source domain classification loss is calculated; the structure-enhanced feature representation is input into the domain discriminator to perform domain discrimination on the samples, and the domain alignment loss is calculated; the structure alignment loss is calculated based on the structure-enhanced feature representation. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, a physical prior distribution corresponding to the target domain samples is constructed; based on the physical consistency constraints between the physical prior distribution and the model prediction distribution, the physical consistency loss is calculated. The parameters of the domain adversarial graph convolutional neural network are updated using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model.

[0007] The preferred technical solution is: The calculation of the physical consistency loss includes:

[0008] In the formula, For physical consistency loss, The mathematical expectation of the target domain samples, Indicates sample Gating weights, For Kullback-Leibler divergence, For physical prior distribution, Predict the distribution for the model.

[0009] When calculating the total loss, a delayed linear weighting training strategy is used to adjust the weights of the physical consistency loss, including:

[0010] In the formula, Total loss; , , These are source domain classification loss, domain alignment loss, and structural alignment loss, respectively. and For the weighting factor; The weights of the physical consistency loss, which dynamically change with training rounds, are calculated as follows:

[0011] In the formula, For the current training round, This is the round in which the physical consistency constraints take effect. For linear weighting and continuous rounds, It has the highest weight.

[0012] The physical prior distribution corresponding to the constructed target domain samples includes: For the sample x Bandpass filtering is performed, and the envelope signal is obtained through Hilbert transform. The envelope spectrum is then obtained through fast Fourier transform, and finally the envelope power spectrum is obtained. For each type of basic fault mechanism, a center frequency is defined for the fault characteristic frequency, and upper and lower bounds of the harmonic neighborhood window are constructed. The energy in each harmonic neighborhood is accumulated to obtain a sample. x Physical evidence scores for each fundamental failure mechanism:

[0013] In the formula, For the sample In basic fault mechanism The physical evidence score in the corresponding harmonic neighborhood; For the first Envelope power spectrum values ​​at each frequency point; Representing the Does the frequency point fall on the ? t The first type of fault characteristic frequency h Indicator functions within the harmonic neighborhood window; set of harmonic orders , H , These are the total harmonic order and the total number of basic fault mechanisms, respectively. Introducing a category mechanism indicator matrix C represents the total number of fault categories, yielding the original mechanism score:

[0014] In the formula, For the sample x Belongs to the The original mechanism score of the fault class; Indicates the first Class-based faults include basic fault mechanisms , Indicates the first Class-specific faults do not include basic fault mechanisms. ; Normalizing the original mechanism fractions yields the physical prior distribution:

[0015] In the formula, For the sample x Belongs to the Physical prior distribution of fault types; The temperature parameter is introduced to adjust the smoothness of the prior distribution. It is a numerically stable term; j Index for fault categories.

[0016] The calculation of the indicator function includes , In the formula: For the first The actual frequency of each frequency point; , in, For the first t The first type of fault characteristic frequency The center frequency of the first harmonic, , These are the upper and lower bounds of the corresponding harmonic neighborhood window, respectively; The relative width coefficient of the harmonic neighborhood window is used to determine the harmonic neighborhood window relative to the center frequency. The bandwidth size; The sampling frequency.

[0017] The definition of the center frequency includes: , in, , For the first t Fault characteristic frequencies.

[0018] The structural alignment loss is calculated based on the maximum mean difference and includes:

[0019] In the formula, For structural alignment loss, These are labeled source domain sample sets and unlabeled target domain sample sets, respectively. Represents a nonlinear mapping function; , The first Source domain samples With the Target domain samples Structural enhancement feature representation; Represents the regenerated nucleus Hilbert space The square norm in .

[0020] The present invention also provides a system for performing the method, comprising: The acquisition module is used to collect vibration signals of the flexible bearings of robot joints under different working conditions and construct a dataset including labeled source domain sample sets and unlabeled target domain sample sets. The training module is used to train the domain adversarial graph convolutional neural network using the dataset to obtain a diagnostic model; The training of the domain adversarial graph convolutional neural network includes: After preprocessing the flexible bearing vibration signals of the dataset, the signals are input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation. The enhanced feature representation is mapped to sample node features, and a sample instance graph is constructed using a graph construction layer based on the similarity between samples. The sample instance graph is then input into a multi-receptive field graph convolutional network to obtain the structural enhanced feature representation. The structure-enhanced feature representation is input into the classifier to obtain the model prediction distribution of the target domain samples, and the source domain classification loss is calculated; the structure-enhanced feature representation is input into the domain discriminator to perform domain discrimination on the samples, and the domain alignment loss is calculated; the structure alignment loss is calculated based on the structure-enhanced feature representation. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, a physical prior distribution corresponding to the target domain samples is constructed; based on the physical consistency constraints between the physical prior distribution and the model prediction distribution, the physical consistency loss is calculated. The parameters of the domain adversarial graph convolutional neural network are updated using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model. The identification module is used to input the target sample to be diagnosed into the trained diagnostic model and output the identification result of the flexible bearing fault category.

[0021] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0022] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when loaded by a processor, is capable of executing the steps of the method.

[0023] The technical solution of the present invention can achieve at least some of the following beneficial effects: This invention combines kernel-selective fusion attention-based multi-scale feature enhancement, graph structure stability modeling, and sample-level physical prior constraints. Through the synergistic effect of these three elements, it can effectively enhance the stability of feature representation under varying operating conditions, the reliability of graph structure propagation, and the accuracy and physical consistency of cross-domain diagnostic results. Specific advantages include: Within the joint learning framework of labeled source domain samples and unlabeled target domain samples, vibration signal fault characterization features are obtained through an improved convolutional feature extraction network. Kernel-selective fusion attention modules are used to adaptively enhance multi-scale feature responses, which improves the stability and discriminativeness of fault features, thereby increasing the reliability of node representation and similarity measurement during sample instance graph construction. Instance graphs are constructed based on inter-sample similarity, and structure propagation is performed using a multi-receptive-field graph convolutional network, enhancing the modeling ability of topological relationships between samples under cross-working conditions. Sample-level physical prior distributions are constructed based on envelope power spectra, fault feature frequencies, and their harmonic neighborhood energies. Physical consistency constraints are introduced into the joint optimization process of the model, improving the accuracy, stability, and physical interpretability of bearing fault diagnosis under varying working conditions.

[0024] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the training process of a domain adversarial graph convolutional neural network according to an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the kernel-selective fusion attention module according to an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of the fault characteristic frequency and its harmonic neighborhood in an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram of the distribution results of bearing fault characteristics across operating conditions obtained in an embodiment of the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0030] This embodiment provides a method for diagnosing faults in flexible bearings of robot joints based on physical prior actuation, including: Vibration signals of flexible bearings in robot joints under different working conditions are collected, and a dataset including labeled source domain sample sets and unlabeled target domain sample sets is constructed. A domain adversarial graph convolutional neural network is trained to obtain a diagnostic model. The target sample to be diagnosed is input into the trained diagnostic model, and the flexible bearing fault category identification result is output.

[0031] In one specific implementation, a labeled source domain sample set is constructed. With unlabeled target domain sample set for: ,

[0032] In the formula, Indicates the number of samples in the source domain. Indicates the number of samples in the target domain. Indicates the first One source domain sample, Indicates the corresponding label, Indicates the first There are 1 target domain sample; the source domain and the target domain have the same label space, but the sample distribution is different.

[0033] See Figure 1 The structure of the domain adversarial graph convolutional neural network specifically includes: an improved convolutional feature extraction network, a graph construction layer, a multi-receptive-field graph convolutional layer, a domain discriminator, and a classifier; the training of the domain adversarial graph convolutional neural network includes: S1. After preprocessing the flexible bearing vibration signal of the dataset, the signal is input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation.

[0034] As a preferred embodiment, the improved convolutional feature extraction network includes convolutional layers, pooling layers, a kernel-selective fusion attention module, and a fully connected layer. Specifically, the improved convolutional feature extraction network in this embodiment introduces the kernel-selective fusion attention module between the last feature map output of a traditional convolutional neural network and the fully connected layer. The kernel-selective fusion attention module extracts multi-scale features from the fault representation feature map using two deep convolutional branches, and obtains two candidate feature responses through pointwise convolution mapping. Then, it adaptively fuses the candidate feature responses through temporal selection and channel selection to obtain the enhanced feature representation.

[0035] In one specific implementation, the training process of the improved convolutional feature extraction network includes: S11. Standardize and segment the bearing vibration signals collected under different operating conditions, input them into an improved convolutional feature extraction network, and obtain the fault characterization feature map corresponding to the sample, including:

[0036] In the formula, the input vibration signal sample , This represents a convolutional feature extraction network. The extracted fault characterization feature map.

[0037] See Figure 2 Kernel-selective fusion attention module for fault characterization feature map As input, the specific training process includes the following S12 to S17: S12. Fault characterization feature map Two depthwise convolutional branches are used to extract multi-scale features, and then pointwise convolutional mapping is used to obtain two candidate feature responses:

[0038] In the formula, , Indicates batch size. Indicates the number of channels. Indicates timing length; This represents a depthwise convolution operation with a kernel length of 3. This represents a depthwise convolution operation with a kernel length of 5 and an expansion rate of 2. and This indicates a pointwise convolution mapping operation.

[0039] S13. Concatenate the two candidate feature responses along the channel dimension to obtain the fused feature representation:

[0040] The fusion feature representation Perform average pooling and max pooling along the channel dimension to obtain time series description information:

[0041] Time series description information , The layers are stitched together and a 1x1 convolutional layer is used. Generate a temporal attention map:

[0042] In the formula, This represents the time-series description information consisting of average pooling branches and max pooling branches; The corresponding branch timing selection mask is obtained through sigmoid activation. ,in .

[0043] S14. Fusion Features Perform global average pooling over the time dimension to obtain the channel description:

[0044] A bottleneck-based channel selection structure is constructed using a two-layer fully connected mapping. The channel description is compressed and then mapped back to the original channel space to obtain the channel response.

[0045] In the formula, For a compact representation, d Indicates the compressed channel dimension; The channel response after mapping back to the original channel space; and These represent two fully connected layers, respectively.

[0046] S15. Will Rearranged as That is, the original channel dimension is divided into two groups of lengths. The branch responses are calculated, and softmax is applied to the branch dimension to obtain the channel selection weights for the two branches:

[0047] In the formula, express The Middle Attention value for each channel location, ,as well as and These represent the features after rearrangement. On the two branches The response value at each location.

[0048] S16. Adjust channel selection weights Timing selection mask After broadcasting by dimension, element-wise multiplication is performed to obtain the joint weights:

[0049] S17. Characteristics of each branch Perform weighted summation, after Obtain attention features:

[0050] The attention gating mask is obtained by applying the sigmoid function to the attention features:

[0051] For the original input feature map Element-wise recalibration is performed to obtain the enhanced feature map:

[0052] In the formula, This represents element-wise multiplication. This represents the feature map enhanced by the kernel-selective fusion attention module, i.e., the enhanced feature representation.

[0053] S2. Map the enhanced feature representation to sample node features, and use the graph construction layer to construct a sample instance graph based on the similarity between samples, determining the corresponding graph adjacency relationships. Preferably, the graph construction layer maps the sample node features and calculates similarity, and combines normalization processing with a sparsity strategy that retains the first k rows to achieve adaptive learning of the graph structure. In a specific embodiment, step S2 includes the following steps: S21. Enhance feature representation After flattening and fully connected mapping, sample-level embeddings are obtained. ; S22. Using a multilayer perceptron (MLP) to... Perform mapping to obtain the mapped node features:

[0054] S23. Calculate the similarity matrix based on the inner product of the features of the mapped nodes, and then normalize it to obtain:

[0055] For similarity matrix Perform by row k The sparsity reduction process, which preserves the terms, yields the weighted adjacency matrix of the instance graph:

[0056] In the formula, This represents the normalized similarity matrix.

[0057] S3. Input the sample instance map into a multi-receptive-field convolutional network, and use the multi-receptive-field convolutional network to perform structure propagation on the constructed instance map. Obtain structure-enhanced feature representations through structure propagation under different receptive fields, thereby improving the ability to learn the correlation between samples and fault features.

[0058] As a preferred approach, at least two layers of multi-receptive-field graph convolutional networks are stacked on the sample instance image to obtain a structure-enhanced feature representation. Each layer of the multi-receptive-field graph convolutional network has multiple graph convolutional channels with different receptive fields on the sample instance image, which perform neighborhood aggregation on the features of the sample nodes respectively. The graph convolutional output features obtained under different receptive fields are then concatenated to obtain the multi-receptive-field graph convolutional output features.

[0059] Specifically, the single-layer multi-receptive-field convolution operation is represented as follows:

[0060] In the formula, Let Laplacian eigenvalues ​​be the diagonal matrix. Used to control the range of neighborhood aggregation For learnable parameters, and Indicates the number of receptive fields.

[0061] S4. Input the structure-enhanced feature representation into the classifier to obtain the model prediction distribution of the target domain samples, and calculate the source domain classification loss; input the structure-enhanced feature representation into the domain discriminator to perform domain discrimination on the samples, and calculate the domain alignment loss; calculate the structure alignment loss based on the structure-enhanced feature representation. In a specific embodiment, the following steps are included: S41. Classifier output unnormalized score representation Then through softmax The model prediction distribution is obtained, which is the predicted probability distribution of the target sample for each fault category:

[0062] In the formula: For the model to sample x Category c The predicted probability; , This represents the output of the multi-receptive-field convolutional network for sample x, i.e., the structure-enhanced feature representation. Represents a label classifier; Source domain classification loss The calculations include:

[0063] in Represents a label classifier. Represents cross-entropy loss, Represents the mathematical expectation; Represents source domain samples The structure-enhanced feature representation obtained after graph structure propagation, i.e. .

[0064] S42. Domain Alignment Loss The calculations include:

[0065] in, Representation domain discriminator; Represents target domain samples The structure-enhanced feature representation obtained after graph structure propagation, i.e. .

[0066] S43. Structural alignment loss Calculations based on the maximum mean difference include:

[0067] In the formula, Represents a nonlinear mapping function. Represents the regenerated nucleus Hilbert space. Represents the regenerated nucleus Hilbert space The square norm in .

[0068] S5. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, construct the physical prior distribution corresponding to the target domain samples. In a preferred embodiment, this includes the following steps: S51. Regarding the sample x Bandpass filtering is performed, and the envelope signal is obtained by Hilbert transform. The envelope spectrum was obtained through Fast Fourier Transform:

[0069] The corresponding envelope power spectrum is expressed as:

[0070] Where: frequency index in single-sided spectrum , For floor operations, The number of points in the Fourier transform; For the first Envelope power spectrum values ​​at each frequency point; Furthermore, the first The actual frequency corresponding to each frequency point Represented as:

[0071] In the formula, The sampling frequency.

[0072] S52. Define a center frequency for the fault characteristic frequency of each type of basic fault mechanism, and construct upper and lower bounds for the harmonic neighborhood window. In a preferred embodiment, the following steps are included: Since each fault category includes at least one basic fault mechanism, this embodiment sets the number of basic fault mechanisms to be [number missing]. K Basic Fault Mechanism t The corresponding fault characteristic frequency is denoted as , This embodiment is preferred. K =3, corresponding to the inner ring fault characteristic frequency. Outer ring fault characteristic frequency and rolling element failure characteristic frequency Under the condition of a deep groove ball bearing with an approximately zero contact angle, the characteristic frequencies of inner ring failure, outer ring failure, and rolling element failure are expressed as follows:

[0073]

[0074]

[0075] In the formula, For frequency conversion, The number of rolling elements. The diameter of the rolling element, Given the pitch circle diameter, the basic failure mechanism is... t Corresponding fault characteristic frequency ; For any basic fault mechanism and its first First harmonic, Define the center frequency And satisfy: Construct the upper bound of the harmonic neighborhood window Lower Boundary :

[0076] in, The relative width coefficient of the harmonic neighborhood window is used to determine the harmonic neighborhood window relative to the center frequency. The bandwidth size; Construct a rectangular indicator window:

[0077] In the formula, Representing the Does the frequency point fall on the ? t The first type of fault characteristic frequency h Indicator functions within the harmonic neighborhood window.

[0078] The fault characteristic frequency and its harmonic neighborhood in this embodiment are as follows: Figure 3 As shown, Figure 3 The red curve represents the single-sided envelope power spectrum of the sample signal; the blue, orange and green dashed lines represent the inner ring fault characteristic frequency, outer ring fault characteristic frequency and rolling element fault characteristic frequency and the center position of each harmonic; the shaded area of ​​the corresponding color represents the harmonic neighborhood window constructed with each characteristic frequency and its harmonic as the center, which is used to accumulate and statistically analyze the energy of the envelope power spectrum within the window, thereby extracting the physical evidence of the corresponding basic fault mechanism.

[0079] S53. Accumulate the energy within each harmonic neighborhood to obtain a sample. x Physical evidence scores for each fundamental failure mechanism:

[0080] In the formula, For the sample In basic fault mechanism The physical evidence score in the corresponding harmonic neighborhood; for any sample The evidence vector for its basic failure mechanism can be obtained as follows:

[0081] S54. Introducing the Category Mechanism Indicator Matrix C represents the total number of fault categories. Indicates the first Class-based faults include basic fault mechanisms , Indicates the first Class-specific faults do not include basic fault mechanisms. For a single fault category, there is a corresponding basic fault mechanism; for a composite fault category, there is a combination of multiple basic fault mechanisms. Based on this mapping relationship, the original mechanism score is obtained:

[0082] In the formula, For the sample x Belongs to the The original mechanism score of the fault class; Normalizing the original mechanism fractions yields the physical prior distribution:

[0083] In the formula, For the sample x Belongs to the Physical prior distribution of fault types; The temperature parameter is introduced to adjust the smoothness of the prior distribution. It is a numerically stable term; j Index for fault categories. When When the size is smaller, the physical prior distribution is sharper, emphasizing the dominant mechanism category. When When the size is large, the distribution is smoother, which helps to mitigate the adverse effects of noise physical evidence on training.

[0084] S6. Based on the physical consistency constraint between the physical prior distribution and the model prediction distribution, calculate the physical consistency loss. ,include:

[0085] In the formula, The mathematical expectation of the target domain samples, Indicates sample The gating weights are determined when the target sample simultaneously satisfies the confidence gating, energy gating, and frequency validity gating conditions. ,otherwise ; For Kullback-Leibler divergence, For sample-level physical prior distribution Predict the distribution for the model; when the denominator is zero, let .

[0086] Specifically, physical consistency constraints are applied only to target domain samples that meet reliability conditions. The reliability conditions refer to the following conditions: the physical prior confidence of the sample exceeds a preset threshold, the energy of the fault feature frequency neighborhood meets a preset threshold, and the corresponding fault feature frequency and its harmonic neighborhood window can be effectively constructed.

[0087] S7. Update the parameters of the domain adversarial graph convolutional neural network using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model.

[0088] In a preferred embodiment, the calculation of the weighted total loss includes: The weights of the physical consistency loss are adjusted using a delayed linear weighting training strategy: , In the formula, The weights for the physical consistency loss, which change dynamically with each training epoch. For the current training round, This is the round in which the physical consistency constraints take effect. For linear weighting and continuous rounds, It has the highest weight.

[0089] Based on physical consistency loss Source domain classification loss Domain alignment loss and structural alignment loss Calculate the total loss :

[0090] In the formula, and This is a weighting factor.

[0091] Specifically, the network parameters of each module in the domain adversarial graph convolutional neural network are jointly updated by minimizing the total loss function to obtain the trained diagnostic model. After the target sample to be diagnosed is input into the trained variable-condition bearing fault diagnosis model, the diagnostic model outputs the predicted probability distribution of the target sample belonging to each fault category, and the category with the highest predicted probability is taken as the final fault category identification result.

[0092] This embodiment also provides a system for performing the aforementioned method for diagnosing faults in flexible bearings of robot joints based on physical prior actuation, comprising: The acquisition module is used to collect vibration signals of the flexible bearings of robot joints under different working conditions and construct a dataset including labeled source domain sample sets and unlabeled target domain sample sets. The training module is used to train the domain adversarial graph convolutional neural network using the dataset to obtain a diagnostic model; The training of the domain adversarial graph convolutional neural network includes: After preprocessing the flexible bearing vibration signals of the dataset, the signals are input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation. The enhanced feature representation is mapped to sample node features, and a sample instance graph is constructed using a graph construction layer based on the similarity between samples. The sample instance graph is then input into a multi-receptive field graph convolutional network to obtain the structural enhanced feature representation. The structure-enhanced feature representation is input into the classifier to obtain the model prediction distribution of the target domain samples, and the source domain classification loss is calculated; the structure-enhanced feature representation is input into the domain discriminator to perform domain discrimination on the samples, and the domain alignment loss is calculated; the structure alignment loss is calculated based on the structure-enhanced feature representation. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, a physical prior distribution corresponding to the target domain samples is constructed; based on the physical consistency constraints between the physical prior distribution and the model prediction distribution, the physical consistency loss is calculated. The parameters of the domain adversarial graph convolutional neural network are updated using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model. The identification module is used to input the target sample to be diagnosed into the trained diagnostic model and output the identification result of the flexible bearing fault category.

[0093] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0094] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when loaded by a processor, is capable of executing the steps of the method.

[0095] The following specific example verifies the effectiveness of the method in rolling bearing fault diagnosis.

[0096] This example uses bearing vibration signals from the HUST bearing dataset as experimental data, including six different types of fault vibration signals: inner ring fault, outer ring fault, rolling element fault, combined inner and outer ring fault, combined inner and rolling element fault, and combined outer ring and rolling element fault. Three load conditions (0W, 200W, and 400W) are set to characterize the bearing vibration characteristics under different operating conditions. The data sampling frequency is 51.2kHz. Each original vibration signal segment is normalized using the Z-score normalization method, and the signal is segmented into fixed-length segments using a sliding window with a length and step size of 1024. Based on this, each sample is labeled to construct a fault diagnosis dataset. The experimental data used are shown in Table 1. Table 1. Number of original signals corresponding to each fault category in the example.

[0097] The samples under different working conditions in the dataset are respectively regarded as different domains, and divided into training set and test set in a ratio of 8:2. Among them, the source domain training samples are labeled training samples, the target domain training samples are unlabeled transfer samples, and the target domain test samples are test evaluation samples. The source domain samples and target domain samples are input into the domain adversarial graph convolutional neural network. At the same time, the physical prior distribution is obtained using the target domain samples. The physical prior distribution is used to impose physical constraints on the target domain samples. The total loss is calculated based on the physical consistency loss, source domain classification loss, domain alignment loss and structural alignment loss. The network is trained by minimizing the total loss function, and finally the optimized diagnostic model is obtained. The diagnostic model is used to identify the fault category of the target domain test samples, thereby realizing bearing fault diagnosis under variable working conditions.

[0098] Figure 4 The results show the distribution of bearing fault characteristics across operating conditions obtained using this example. Figure 4 In the diagram, the horizontal axis represents the first-dimensional feature after dimensionality reduction using t-SNE (t-Distributed Stochastic Neighbor Embedding), and the vertical axis represents the second-dimensional feature after t-SNE dimensionality reduction. S1-S6 and T1-T6 represent the feature distributions of the six fault categories in the source and target domains, respectively. Fault categories 1-6 correspond to inner race faults, outer race faults, rolling element faults, combined inner and outer race faults, combined inner race and rolling element faults, and combined outer race and rolling element faults, respectively. Categories with the same number correspond to the same fault category.

[0099] Depend on Figure 4 As can be seen, the features learned in this example exhibit clear distribution boundaries and good inter-class separability across different fault categories. The source and target domains also achieve sufficient alignment and overlap for features within the same fault category, resulting in a compact cluster structure. Notably, for the sixth type of fault—the combined outer ring and rolling element fault—both models present two relatively independent subclusters, but each subcluster contains sufficient overlap between source and target domain samples. This suggests that this phenomenon is more likely to reflect the intra-class multimodal distribution characteristics of the same fault under different operating conditions, rather than being caused by insufficient cross-domain alignment. Overall, the results verify that the method in this embodiment achieves more stable cross-domain feature alignment while maintaining intra-class structural information. Furthermore, the samples of each fault category exhibit good intra-class compactness and inter-class separability in the feature space, and the source and target domain samples of the same type can achieve effective aggregation and alignment. This indicates that the method in this embodiment can maintain relatively stable feature extraction capabilities and strong fault identification capabilities under varying operating conditions, demonstrating good performance in the diagnosis of bearing faults under varying operating conditions.

[0100] In summary, this invention combines kernel-selective fusion attention multi-scale feature enhancement, graph structure stability modeling, and sample-level physical prior constraints. Through the synergistic effect of these three elements, it can effectively enhance the stability of feature representation under varying operating conditions, the reliability of graph structure propagation, and the accuracy and physical consistency of cross-domain diagnostic results.

[0101] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of flexible bearings in robot joints based on physical prior actuation, characterized in that, include: Vibration signals of flexible bearings in robot joints under different working conditions are collected, and a dataset including labeled source domain sample sets and unlabeled target domain sample sets is constructed. A domain adversarial graph convolutional neural network is trained to obtain a diagnostic model. The target sample to be diagnosed is input into the trained diagnostic model, and the flexible bearing fault category identification result is output. The training of the domain adversarial graph convolutional neural network includes: After preprocessing the flexible bearing vibration signals of the dataset, the signals are input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation. The enhanced feature representation is mapped to sample node features, and a sample instance graph is constructed using a graph construction layer based on the similarity between samples. The sample instance graph is then input into a multi-receptive field graph convolutional network to obtain the structural enhanced feature representation. The structure-enhanced feature representation is input into the classifier to obtain the model prediction distribution of the target domain samples, and the source domain classification loss is calculated; the structure-enhanced feature representation is input into the domain discriminator to perform domain discrimination on the samples, and the domain alignment loss is calculated; the structure alignment loss is calculated based on the structure-enhanced feature representation. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, a physical prior distribution corresponding to the target domain samples is constructed; based on the physical consistency constraints between the physical prior distribution and the model prediction distribution, the physical consistency loss is calculated. The parameters of the domain adversarial graph convolutional neural network are updated using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model. The physical prior distribution corresponding to the constructed target domain samples includes: The sample is bandpass filtered, and the envelope signal is obtained by Hilbert transform. The envelope spectrum is obtained by fast Fourier transform, and then the envelope power spectrum is obtained. For each type of basic fault mechanism, a center frequency is defined for the fault characteristic frequency, and upper and lower bounds of the harmonic neighborhood window are constructed. The energy in each harmonic neighborhood is accumulated to obtain the physical evidence score of the sample on each basic fault mechanism: By introducing the categorical mechanism indicator matrix, the original mechanism scores are obtained: The physical prior distribution is obtained by normalizing the original mechanism fraction.

2. The method according to claim 1, characterized in that, The calculation of the physical consistency loss includes: ， In the formula, For physical consistency loss, The mathematical expectation of the target domain samples, Indicates sample Gating weights, For Kullback-Leibler divergence, For physical prior distribution, Predict the distribution for the model.

3. The method according to claim 2, characterized in that, When calculating the total loss, a delayed linear weighting training strategy is used to adjust the weights of the physical consistency loss, including: ， In the formula, Total loss; , , These are source domain classification loss, domain alignment loss, and structural alignment loss, respectively. and For the weighting factor; The weights of the physical consistency loss, which dynamically change with training rounds, are calculated as follows: ， In the formula, For the current training round, This is the round in which the physical consistency constraints take effect. For linear weighting and continuous rounds, It has the highest weight.

4. The method according to claim 1, characterized in that, The accumulation of energy within each harmonic neighborhood yields the physical evidence score of the sample for each fundamental fault mechanism, including: ， In the formula, For the sample In basic fault mechanism The physical evidence score in the corresponding harmonic neighborhood; For the first Envelope power spectrum values ​​at each frequency point; Representing the Does the frequency point fall on the ? t The first type of fault characteristic frequency h Indicator functions within the harmonic neighborhood window; set of harmonic orders , H , These are the total harmonic order and the total number of basic fault mechanisms, respectively. Introducing a category mechanism indicator matrix C represents the total number of fault categories, yielding the original mechanism score: ， In the formula, For the sample x Belongs to the The original mechanism score of the fault class; Indicates the first Class-based faults include basic fault mechanisms , Indicates the first Class-specific faults do not include basic fault mechanisms. ; The normalization of the original mechanism score to obtain the physical prior distribution includes: ， In the formula, For the sample x Belongs to the Physical prior distribution of fault types; The temperature parameter is introduced to adjust the smoothness of the prior distribution. It is a numerically stable term; j Index for fault categories.

5. The method according to claim 4, characterized in that, The calculation of the indicator function includes: ， In the formula: For the first The actual frequency of each frequency point; ， in, For the first t The first type of fault characteristic frequency The center frequency of the first harmonic, , These are the upper and lower bounds of the corresponding harmonic neighborhood window, respectively; The relative width coefficient of the harmonic neighborhood window is used to determine the harmonic neighborhood window relative to the center frequency. The bandwidth size; The sampling frequency.

6. The method according to claim 5, characterized in that, The definition of the center frequency includes: ， in, , For the first t Fault characteristic frequencies.

7. The method according to claim 1, characterized in that, The structural alignment loss is calculated based on the maximum mean difference, including: ， In the formula, For structural alignment loss, These are labeled source domain sample sets and unlabeled target domain sample sets, respectively. Represents a nonlinear mapping function; , The first Source domain samples With the Target domain samples Structural enhancement feature representation; Represents the regenerated nucleus Hilbert space The square norm in .

8. A system for performing the method according to any one of claims 1-7, characterized in that, include: The acquisition module is used to collect vibration signals of the flexible bearings of robot joints under different working conditions and construct a dataset including labeled source domain sample sets and unlabeled target domain sample sets. The training module is used to train the domain adversarial graph convolutional neural network using the dataset to obtain a diagnostic model; The training of the domain adversarial graph convolutional neural network includes: After preprocessing the flexible bearing vibration signals of the dataset, the signals are input into an improved convolutional feature extraction network to obtain a fault characterization feature map. The improved convolutional feature extraction network introduces a kernel-selective fusion attention module, which takes the fault characterization feature map as input and outputs an enhanced feature representation. The enhanced feature representation is mapped to sample node features, and a sample instance graph is constructed using a graph construction layer based on the similarity between samples. The sample instance graph is then input into a multi-receptive field graph convolutional network to obtain the structural enhanced feature representation. The structure-enhanced feature representation is input into the classifier to obtain the model prediction distribution of the target domain samples, and the source domain classification loss is calculated; the structure-enhanced feature representation is input into the domain discriminator to perform domain discrimination on the samples, and the domain alignment loss is calculated; the structure alignment loss is calculated based on the structure-enhanced feature representation. Based on the fault characteristic frequencies and their harmonic neighborhood energy distributions in the envelope power spectrum of the target domain samples, a physical prior distribution corresponding to the target domain samples is constructed; based on the physical consistency constraints between the physical prior distribution and the model prediction distribution, the physical consistency loss is calculated. The parameters of the domain adversarial graph convolutional neural network are updated using the weighted total loss of the physical consistency loss and other losses to obtain the trained diagnostic model. The identification module is used to input the target sample to be diagnosed into the trained diagnostic model and output the identification result of the flexible bearing fault category.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by a processor, it is able to execute the steps of the method according to any one of claims 1 to 7.

Images