Substation equipment fault diagnosis method based on physical constraint bayesian deep learning

Abstract

The application relates to a power transformation equipment fault diagnosis method based on physical constraint Bayesian deep learning, which comprises the following steps: obtaining multi-modal input features by respectively adopting synchronous compression wavelet transform and empirical mode decomposition to process original signals; extracting different dimension features from the multi-modal input features by a double-branch feature extraction backbone network to obtain fusion features, generating a reconstructed signal by input signal reconstruction decoding branch, constructing a physical consistency loss, inputting a Bayesian classifier to learn a variational approximate posterior distribution of network weights through variational inference and outputting a category probability distribution; constructing a total optimization objective function based on the physical consistency loss; constructing a comprehensive uncertainty measurement index by adopting classification prediction entropy and reconstruction quality, sorting sample prediction reliability, and realizing reliability evaluation and risk control of fault diagnosis results through a high-uncertainty sample rejection mechanism. The fault recognition precision and the reliability of diagnosis results under strong noise and complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring and health assessment technology for power equipment, specifically to a fault diagnosis method for power equipment based on physical constraint Bayesian deep learning. Background Technology

[0002] As a critical primary equipment in the power transmission and transformation process, the health status of substation equipment directly affects the safety, stability, and reliability of the power grid operation. Against the backdrop of rapid development in industrial big data, intelligent sensing, and deep learning technologies, data-driven intelligent fault diagnosis methods have made significant progress in the field of substation equipment condition monitoring. Current substation equipment fault diagnosis technologies are mainly divided into two categories: signal analysis methods and deep learning methods. Signal analysis methods rely on manual extraction of time-domain and frequency-domain features such as kurtosis and spectrum. The feature extraction process is cumbersome, prone to losing key information from vibration signals, and its diagnostic accuracy is difficult to guarantee under varying operating conditions. Deep learning methods can autonomously mine deep features from data, eliminating reliance on manual feature engineering. However, their output is only a single, definitive classification label, lacking a supporting confidence assessment mechanism. That is, traditional deep learning models typically only provide deterministic point estimates and cannot measure the reliability of prediction results. When faced with industrial noise, operating condition fluctuations, or out-of-domain anomalies, the model is prone to overconfident and erroneous predictions about data samples with unknown distributions. Furthermore, they generally suffer from insufficient model interpretability.

[0003] Therefore, developing uncertainty quantification methods for fault diagnosis is of significant theoretical and engineering application value for improving the reliability of diagnostic results and mitigating potential safety risks. In recent years, a large amount of research has been conducted on the uncertainty quantification problem in fault diagnosis. In Bayesian deep learning, researchers have introduced Monte Carlo random deactivation sampling and variational inference into deep learning models for uncertainty quantification assessment. In physical information fusion, researchers have proposed Bayesian physical information neural networks, which embed partial differential equation constraints into the loss function and combine functional tensor Monte Carlo and variational inference to achieve reasonable uncertainty quantification. However, these methods mainly focus on regression problems in continuous physical fields and are difficult to directly apply to discrete classification fault diagnosis scenarios. In uncertainty decomposition, researchers have achieved the separation and quantification of cognitive uncertainty and accidental uncertainty in fault diagnosis through expected calibration error and mixed density calibration networks. However, these methods lack explicit constraints on the physical mechanisms of the fault. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method for power equipment based on physically constrained Bayesian deep learning, which solves the technical problem that existing methods are unable to meet the requirements for fault identification accuracy and reliability of diagnosis results under strong noise and complex operating conditions.

[0005] The technical solution adopted in this invention is as follows: This invention provides a method for fault diagnosis of substation equipment based on physically constrained Bayesian deep learning, comprising: Vibration signals of power equipment under different operating conditions are collected, and synchronous compressed wavelet transform and empirical mode decomposition are used to process the vibration signals to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. The dual-branch feature extraction backbone network extracts features of different dimensions from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components, respectively. After being spliced ​​and fused by the fusion layer, the fused feature vector is output. The fused feature vector is input to the signal and the decoding branch is reconstructed to generate a temporal reconstructed signal, and a physical consistency loss is constructed. Simultaneously, the fused feature vector is input to a Bayesian classifier, and the variational approximate posterior distribution of the Bayesian layer network weights is obtained through variational inference learning, and the fault category probability distribution is output. The variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two. We jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss to perform iterative training and parameter updates of the model. The failure category probability distribution of the Bayesian classifier is obtained by random deactivation using Monte Carlo methods. The uncertainty of the failure category probability distribution is quantified by classification prediction entropy, and the total uncertainty is divided into cognitive uncertainty and accidental uncertainty. The average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal is used to define the reconstruction quality. The classification prediction entropy and reconstruction quality are combined to construct a comprehensive uncertainty metric. The reliability of sample predictions is ranked using the comprehensive uncertainty metric, and the credibility assessment and risk control of fault diagnosis results are achieved through a high-uncertainty sample rejection mechanism.

[0006] The preferred technical solution is: The construction of the physical consistency loss includes: Perform a Fast Fourier Transform on the reconstructed time-domain signal to obtain the amplitude spectrum, and calculate the signal energy within the preset fault characteristic frequency band. ; Calculate the total energy of the time-domain reconstructed signal across the entire frequency band. ; The ratio of the energy in the preset fault frequency band to the total energy across the entire frequency band is defined as the physical fit index. :

[0007] in, To prevent extremely small positive smoothing constants with a denominator of zero; The physical consistency loss is constructed by using the energy concentration of a preset fault characteristic frequency band as a physical constraint.

[0008] In the formula, S represents the set of sample indices belonging to the non-healthy category in the current batch. Let be the physical fit index for the s-th non-healthy category sample.

[0009] The construction of the comprehensive uncertainty metric includes:

[0010] In the formula, Let x be the classification prediction entropy for predicting class y given input x. To reconstruct quality, The hyperparameters for balancing the weights.

[0011] The construction of the KL divergence regularization loss includes:

[0012] In the formula, The loss is the KL divergence regularization loss. To constrain the weights of the Bayesian layer network variational approximate posterior distribution Compared with the true posterior distribution The KL divergence regularization term of the difference between them, where, An approximation is made using a standard Gaussian prior distribution.

[0013] The signal reconstruction loss is the mean square error between the original signal and the time-domain reconstructed signal.

[0014] The cognitive uncertainty is measured using mutual information to describe the uncertainty caused by large differences between predictions obtained from multiple samplings when the sample coverage is insufficient or the test samples deviate from the training distribution; the random uncertainty is measured using expected entropy to describe the uncertainty caused by the overlap of random noise in the data itself and fault features.

[0015] The dual-branch feature extraction backbone network includes: A deep separable convolutional network for extracting spatial texture features from synchronous compressed wavelet transform time-frequency maps; Long Short-Term Memory Network used for extracting temporal dynamic features from the intrinsic mode function components.

[0016] The present invention also provides a fault diagnosis device for substation equipment based on physical constraint Bayesian deep learning, comprising: The multimodal feature construction module is used to collect vibration signals of power equipment under different operating conditions, and process the vibration signals by synchronous compressed wavelet transform and empirical mode decomposition respectively to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. The fusion feature construction module is used to extract features of different dimensions from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components through a dual-branch feature extraction backbone network, and output a fusion feature vector after splicing and fusing through a fusion layer. The signal reconstruction and Bayesian classification module is used to map the fused feature vector into the signal reconstruction decoding branch to generate a temporal reconstructed signal and construct a physical consistency loss. Simultaneously, the fused feature vector is input into a Bayesian classifier, and variational inference learning is used to obtain the variational approximate posterior distribution of the Bayesian layer network weights, outputting a fault category probability distribution. This variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two. The joint optimization module is used to jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss, and to perform iterative training and parameter updates of the model. A comprehensive uncertainty metric construction module is used to obtain the fault category probability distribution of the Bayesian classifier through Monte Carlo random deactivation, quantify the uncertainty of the fault category probability distribution using classification prediction entropy, and divide the total uncertainty into cognitive uncertainty and accidental uncertainty; the average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal is used to define the reconstruction quality; and the comprehensive uncertainty metric is constructed by fusing the classification prediction entropy and the reconstruction quality. The credibility assessment and risk control module is used to rank the reliability of sample predictions using the comprehensive uncertainty metric, and to achieve credibility assessment and risk control of fault diagnosis results through a high uncertainty sample rejection mechanism.

[0017] The present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method.

[0018] The present invention also provides a computer device, the computer device including a processor and a memory, the memory storing at least one instruction, the instruction being loaded and executed by the processor to implement the method.

[0019] The technical solution of the present invention can achieve at least some of the following beneficial effects: This invention proposes a reliable fault diagnosis method for substation equipment that integrates physical consistency constraints and probabilistic Bayesian representation learning, while also possessing uncertainty quantification capabilities. Through multimodal signal feature extraction, a dual-branch parallel network structure, and a multi-loss joint training mechanism, it mines the inherent fault mechanism features of vibration signals. Simultaneously, it quantifies the classification and reconstruction uncertainties during the prediction process, relying on the rejection of high-uncertainty samples to achieve fault diagnosis reliability assessment, effectively improving the accuracy of fault identification and the reliability of diagnostic results under strong noise and complex operating conditions. Specifically, it has the following advantages: This invention essentially constructs a physically consistent Bayesian probabilistic representation learning framework (PCP-BNN). By reconstructing latent features into vibration signals through a decoder and using the energy concentration of fault feature frequency bands as a physical constraint embedding loss function, it achieves a deep fusion of fault prior knowledge and data-driven models in discrete classification scenarios.

[0020] This invention essentially proposes an uncertainty quantification criterion based on signal reconstruction quality. Building upon Monte Carlo random deactivation sampling's decoupling of cognitive and accidental uncertainties, it introduces the average mean square error between the reconstructed and original signals as a measure of reconstruction quality. This error is then fused with prediction entropy to form a comprehensive uncertainty metric, enhancing risk perception and credibility assessment of unreliable prediction samples. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the construction process of the synchronous compressed wavelet transform time-frequency diagram according to an embodiment of the present invention.

[0023] Figure 3 These are the time-domain and frequency-domain diagrams of the original signal and the corresponding reconstructed signal in the embodiments of the present invention.

[0024] Figure 4 This is a diagram showing the classification results obtained by the method in an embodiment of the present invention.

[0025] Figure 5 This is a diagram showing the total uncertainty distribution of the method in the embodiments of the present invention.

[0026] Figure 6 This is a reliability calibration curve of the model in an embodiment of the present invention.

[0027] Figure 7 for Figure 6 An enlarged view of the curve within the dashed box. Detailed Implementation

[0028] The specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0029] See Figure 1 This embodiment provides a fault diagnosis method for substation equipment based on physically constrained Bayesian deep learning, including: S1. Collect vibration signals of the power equipment under different operating conditions, and process the vibration signals using synchronous compressed wavelet transform and empirical mode decomposition respectively to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. Specifically, this includes the following steps: S11. Acquire the vibration signal using an accelerometer; S12. By synchronously compressing and rearranging the time-frequency coefficients of the continuous wavelet transform along the frequency direction, that is:

[0030] In the formula, Indicates the time position Target angular frequency The redistribution time-frequency coefficients of the synchronous compressed wavelet transform. For frequency unit width, Used for amplitude normalization after frequency redistribution For the first on the scale axis A discrete scale point, The sampling interval between adjacent scales. The scale normalization factor, For the input signal at scale Time and location The continuous wavelet transform coefficients at the point, This represents the instantaneous angular frequency estimated from the wavelet coefficients.

[0031] Multi-scale time-frequency features are constructed by changing the scale resolution parameter (i.e., the number of scale samples) of the synchronous compressed wavelet transform. Three types of time-frequency spectra are generated, namely high resolution (scale resolution parameter = 64), medium resolution (scale resolution parameter = 32), and low resolution (scale resolution parameter = 16). After logarithmic normalization and size unification, they are stacked along the channel dimension to form a 3×64×64 three-channel tensor, which is the synchronous compressed wavelet transform time-frequency map as the input feature of the subsequent network.

[0032] The process of constructing the time-frequency plot of the synchronous compressed wavelet transform is described in [reference needed]. Figure 2 , Figure 2 From left to right, the images show the original partial discharge fault signal, the corresponding high, medium, and low resolution time-frequency graphs, and the final fused image.

[0033] S13. The empirical mode decomposition adaptively decomposes the original vibration signal into several mode functions with intrinsic scale characteristics and a residual component. In this embodiment, the first 5 mode functions of the empirical mode decomposition are selected as the feature inputs of the subsequent network.

[0034] S2. By using a dual-branch feature extraction backbone network, features of different dimensions are extracted from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components, respectively. After being spliced ​​and fused by the fusion layer, a fused feature vector is output.

[0035] As a preferred embodiment, the dual-branch feature extraction backbone network includes: a deep separable convolutional neural network (DSCNN) for extracting spatial texture features from the synchronous compressed wavelet transform time-frequency map; and a long short-term memory network (LSTM) for extracting temporal dynamic features from the intrinsic mode function components.

[0036] S3. The fused feature vector is input to the signal and the decoding branch is reconstructed to generate a temporal reconstructed signal, and a physical consistency loss is constructed. At the same time, the fused feature vector is input to the Bayesian classifier, and the variational approximate posterior distribution of the Bayesian layer network weights is obtained through variational inference learning, and the fault category probability distribution is output. The variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two.

[0037] In this embodiment, the Bayesian classification branch and the signal reconstruction decoding branch are set in parallel after the fusion layer. The fused feature vector is then split into multiple paths. As a preferred method, this includes the following steps: S31. Input the fused feature vector into a Bayesian classifier, obtain the variational approximate posterior distribution of the Bayesian layer network weights through variational inference learning, and output the fault category probability distribution, including: Due to the posterior distribution of the weight parameters in a Bayesian neural network It is usually difficult to solve analytically directly, so variational inference techniques are used to approximate it. Specifically, this is achieved by introducing an optimizable variational approximation posterior distribution. To approximate the true posterior distribution and minimize the KL divergence between them, the weight parameter distribution is iteratively updated. The mathematical expression is:

[0038] In the formula, Indicates the weights of the Bayesian layer network The variational approximation of the posterior distribution, In order to be in Training dataset The true posterior distribution, represents the learnable parameters of the variational distribution.

[0039] Due to the true posterior distribution Since the KL divergence cannot be directly calculated, it is difficult to optimize directly. Therefore, the posterior optimization problem is transformed into finding the maximum value of the lower bound of evidence. The formula for the lower bound of evidence is as follows:

[0040] In the formula, the first term is the expected log-likelihood, used to measure the... The following model is used on the training dataset Fitting ability Indicates in Below Calculate the expected value; the second term is the KL divergence regularization term, used to constrain... Compared with the true posterior distribution The differences between them can suppress model overfitting and improve generalization performance.

[0041] As a specific method, an independent Gaussian distribution is used. Perform approximate modeling:

[0042] in, Indicates weight The variational approximation of the posterior distribution, Represents the weight matrix Follow the mean With variance Univariate normal distribution, mean with standard deviation These are the learnable variational parameters. The subscripts i and j are the row and column indices of the elements in the weight matrix, respectively. This represents the random weight variable in the i-th row and j-th column of the Bayesian neural network weight matrix (i.e., the weights of the Bayesian layer network). To ensure that the random sampling process can effectively participate in backpropagation, a reparameterization technique is used to complete the parameter sampling:

[0043] in, To introduce the Gaussian noise term, this operation separates the randomness of the sampling process from the computation graph. While ensuring total differential optimization of the parameter chain, a closed analytical regularization term can be obtained by combining it with the standard Gaussian prior.

[0044] The classification results obtained by the method in this embodiment can be found in [reference]. Figure 4 .

[0045] S32. The fused feature vector input signal is reconstructed and decoded to generate a temporal reconstructed signal, and a physical consistency loss is constructed, including: S321. Using a fully connected decoder consisting of a 3-layer fully connected network as the signal reconstruction decoding branch, the fused feature vector is mapped to a time-domain reconstructed signal, and the reconstructed signal is used as the carrier of physical constraints. Its mathematical expression is:

[0046] Specifically, For time-domain reconstructed signals, It is the ReLU activation function. , These are the weight matrices and bias vectors for each layer, respectively. To fuse feature vectors.

[0047] Figure 3 In the image, (a) and (b) are the time-domain plots of the original signal and the corresponding reconstructed signal, respectively. Figure 3 In the diagrams (c) and (d), the frequency domain diagrams of the original signal and the corresponding reconstructed signal are shown, respectively.

[0048] S322. Constructing physical consistency loss includes: When a transformer experiences a specific fault, its electromagnetic force, magnetostriction, and mechanical coupling response will cause significant energy accumulation in the vibration signal near the corresponding fault characteristic frequency and its harmonics. Based on this constraint, a fast Fourier transform is performed on the reconstructed signal output by the fully connected decoder to obtain the amplitude spectrum. The degree of fit between the reconstructed signal and the physical laws of the fault is quantified by the frequency band energy ratio. The specific calculation steps are as follows: Calculate the signal energy within the preset fault characteristic frequency band. :

[0049] In the formula, Represents the time-domain reconstruction signal After Fast Fourier Transform, at frequency The amplitude in the frequency domain at that point, These are the lower and upper limits of the preset fault characteristic frequency band, respectively; Calculate the total energy of the time-domain reconstructed signal across the entire frequency band. :

[0050] In the formula, Indicates the sampling frequency of the vibration signal. It is the Nyquist frequency, which is the highest effective frequency of a single-sided spectrum.

[0051] The ratio of the energy in the preset fault frequency band to the total energy across the entire frequency band is defined as the physical fit index. :

[0052] In the formula, This represents a minimal positive smoothing constant to prevent the denominator from being zero, used to ensure the numerical stability of ratio calculations.

[0053] The physical consistency loss is constructed by using the energy concentration of a preset fault characteristic frequency band as a physical constraint.

[0054] Where S represents the set of sample indexes belonging to the non-healthy category in the current batch, which is obtained by filtering samples in the batch according to the real category labels, and only applies frequency band energy physical constraints to faulty samples; Let be the physical fit index for the s-th non-healthy category sample.

[0055] S4. Jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss to perform iterative training and parameter updates of the model.

[0056] Specifically, the total loss function is constructed using the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss. :

[0057] In the formula, Cross-entropy is used for classification loss; The KL divergence regularization loss is specifically the KL divergence regularization term between the variational approximation posterior of the Bayesian layer weights and the standard Gaussian prior distribution. ; , , These are the weighting coefficients for each loss term; The signal reconstruction loss is preferably defined as the mean square error between the original signal and the reconstructed signal:

[0058] In the formula, N is the number of samples involved in the calculation of repeated loss; and They represent the first The original vibration signal and its reconstructed signal.

[0059] S5. To quantify the uncertainty of the model's prediction results, a comprehensive uncertainty metric is constructed, including the following steps: S51. Obtain the fault category probability distribution of the Bayesian classifier through Monte Carlo random deactivation:

[0060] in, Indicates the first One input sample, This indicates the class that the model predicts for the sample. Represents the training dataset. Indicates the first The network weights are then sampled from the Bayesian weight approximation posterior distribution. Indicates the number of random inactivations in Monte Carlo. Indicates the first Model for samples under group weight parameters Output the probability distribution of fault categories.

[0061] The uncertainty of the probability distribution of the fault categories is quantified using classification prediction entropy:

[0062] in, Samples in the Bayesian fault category probability distribution Predicted as The average predicted probability of the class Total number of fault categories; Let be the classification prediction entropy for a given input x that predicts class y.

[0063] S52. The total uncertainty is divided into cognitive uncertainty and accidental uncertainty. Preferably, mutual information is used to measure the cognitive uncertainty to describe the uncertainty caused by large differences between prediction results obtained by the model from multiple samplings when the sample coverage is insufficient or the test samples deviate from the training distribution; expected entropy is used to measure the accidental uncertainty to describe the uncertainty caused by the overlap of random noise in the data itself and fault features.

[0064] The mathematical expression for mutual information is as follows:

[0065] in, Indicates the first Classification prediction entropy under sub-model sampling conditions; The mathematical expression for expected entropy:

[0066] Furthermore, the classification prediction entropy, mutual information, and expected entropy satisfy the following relationship:

[0067] S53. The average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal ( Define reconstruction quality:

[0068] In the formula, The original input signal, For the first The reconstructed signal obtained by Monte Carlo sampling.

[0069] S54. Integrate the classification prediction entropy and reconstruction quality to construct a comprehensive uncertainty metric. :

[0070] In the formula, The hyperparameters for balancing the weights.

[0071] S6. The reliability of sample predictions is ranked using the comprehensive uncertainty metric, and the credibility assessment and risk control of fault diagnosis results are achieved through a high uncertainty sample rejection mechanism.

[0072] The total uncertainty distribution obtained by the method in this embodiment is as follows: Figure 5 As shown. By Figure 5 It can be seen that the total uncertainty of the medium- and low-temperature overheating fault samples is significantly higher than that of the other four types of operating conditions, reflecting that the characteristics of this type of fault are weak and the identification is low. The model has the lowest confidence in fault identification and signal reconstruction for this type of fault and the largest fluctuation in the results. The uncertainties of the healthy state, high-temperature overheating, partial discharge and high-energy discharge are all maintained at a low level and are concentrated in distribution. The model's diagnosis and reconstruction results for the above operating conditions are stable and reliable.

[0073] To quantitatively verify the reliability of the model's output probability in this embodiment, industry-standard expected calibration error and negative log-likelihood are used as quantitative evaluation metrics. The expected calibration error quantifies the deviation between the model's prediction confidence and the actual classification accuracy, while the negative log-likelihood evaluates the goodness of fit between the model's predicted distribution and the actual fault labels. Verification results can be found in [link to verification results]. Figure 6 and Figure 7 The results show that the model's reliability calibration curve fits the ideal reference line well. This indicates that the model's predicted fault category probabilities are highly consistent with the actual classification results, and the prediction confidence accurately reflects the reliability of the actual diagnosis. While maintaining excellent fault identification accuracy, this model can output probability prediction results with good calibration performance, providing reliable support for uncertainty measurement, high-uncertainty sample identification, and fault diagnosis risk control.

[0074] This embodiment also provides a fault diagnosis device for substation equipment based on physically constrained Bayesian deep learning, including: The multimodal feature construction module is used to collect vibration signals of power equipment under different operating conditions, and process the vibration signals by synchronous compressed wavelet transform and empirical mode decomposition respectively to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. The fusion feature construction module is used to extract features of different dimensions from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components through a dual-branch feature extraction backbone network, and output a fusion feature vector after splicing and fusing through a fusion layer. The signal reconstruction and Bayesian classification module is used to map the fused feature vector into the signal reconstruction decoding branch to generate a temporal reconstructed signal and construct a physical consistency loss. Simultaneously, the fused feature vector is input into a Bayesian classifier, and variational inference learning is used to obtain the variational approximate posterior distribution of the Bayesian layer network weights, outputting a fault category probability distribution. This variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two. The joint optimization module is used to jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss, and to perform iterative training and parameter updates of the model. A comprehensive uncertainty metric construction module is used to obtain the fault category probability distribution of the Bayesian classifier through Monte Carlo random deactivation, quantify the uncertainty of the fault category probability distribution using classification prediction entropy, and divide the total uncertainty into cognitive uncertainty and accidental uncertainty; the average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal is used to define the reconstruction quality; and the comprehensive uncertainty metric is constructed by fusing the classification prediction entropy and the reconstruction quality. The credibility assessment and risk control module is used to rank the reliability of sample predictions using the comprehensive uncertainty metric, and to achieve credibility assessment and risk control of fault diagnosis results through a high uncertainty sample rejection mechanism.

[0075] This embodiment also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method described therein.

[0076] This embodiment also provides a computer device, which includes a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the method described.

[0077] 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 substation equipment based on physically constrained Bayesian deep learning, characterized in that, include: Vibration signals of power equipment under different operating conditions are collected, and synchronous compressed wavelet transform and empirical mode decomposition are used to process the vibration signals to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. The dual-branch feature extraction backbone network extracts features of different dimensions from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components, respectively. After being spliced ​​and fused by the fusion layer, the fused feature vector is output. The fused feature vector is input to the signal and the decoding branch is reconstructed to generate a temporal reconstructed signal, and a physical consistency loss is constructed. Simultaneously, the fused feature vector is input to a Bayesian classifier, and the variational approximate posterior distribution of the Bayesian layer network weights is obtained through variational inference learning, and the fault category probability distribution is output. The variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two. We jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss to perform iterative training and parameter updates of the model. The failure category probability distribution of the Bayesian classifier is obtained by random deactivation using Monte Carlo methods. The uncertainty of the failure category probability distribution is quantified by classification prediction entropy, and the total uncertainty is divided into cognitive uncertainty and accidental uncertainty. The average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal is used to define the reconstruction quality. The classification prediction entropy and reconstruction quality are combined to construct a comprehensive uncertainty metric. The reliability of sample predictions is ranked using the comprehensive uncertainty metric, and the credibility assessment and risk control of fault diagnosis results are achieved through a high-uncertainty sample rejection mechanism.

2. The method according to claim 1, characterized in that, The construction of the physical consistency loss includes: Perform a Fast Fourier Transform on the reconstructed time-domain signal to obtain the amplitude spectrum, and calculate the signal energy within the preset fault characteristic frequency band. ; Calculate the total energy of the time-domain reconstructed signal across the entire frequency band. ; The ratio of the energy in the preset fault frequency band to the total energy across the entire frequency band is defined as the physical fit index. : in, To prevent extremely small positive smoothing constants with a denominator of zero; The physical consistency loss is constructed by using the energy concentration of a preset fault characteristic frequency band as a physical constraint. In the formula, S represents the set of sample indices belonging to the non-healthy category in the current batch. Let be the physical fit index for the s-th non-healthy category sample.

3. The method according to claim 1, characterized in that, The construction of the comprehensive uncertainty measurement index includes: In the formula, Let x be the classification prediction entropy for predicting class y given input x. To reconstruct quality, The hyperparameters for balancing the weights.

4. The method according to claim 1, characterized in that, The construction of the KL divergence regularization loss includes: In the formula, For KL divergence regularization loss, To constrain the weights of the Bayesian layer network Variational approximate posterior distribution Compared with the true posterior distribution The KL divergence regularization term of the difference between them, where, An approximation is made using a standard Gaussian prior distribution.

5. The method according to claim 1, characterized in that, The signal reconstruction loss is the mean square error between the original signal and the time-domain reconstructed signal.

6. The method according to claim 1, characterized in that, The cognitive uncertainty is measured using mutual information to describe the uncertainty caused by large differences between predictions obtained from multiple samplings when the sample coverage is insufficient or the test samples deviate from the training distribution; the random uncertainty is measured using expected entropy to describe the uncertainty caused by the overlap of random noise in the data itself and fault features.

7. The method according to claim 1, characterized in that, The dual-branch feature extraction backbone network includes: A deep separable convolutional network for extracting spatial texture features from synchronous compressed wavelet transform time-frequency maps; Long Short-Term Memory Network used for extracting temporal dynamic features from the intrinsic mode function components.

8. A fault diagnosis device for substation equipment based on physically constrained Bayesian deep learning according to any one of claims 1-7, characterized in that, include: The multimodal feature construction module is used to collect vibration signals of power equipment under different operating conditions, and process the vibration signals by synchronous compressed wavelet transform and empirical mode decomposition respectively to obtain the synchronous compressed wavelet transform time-frequency diagram and intrinsic mode function components. The fusion feature construction module is used to extract features of different dimensions from the synchronous compressed wavelet transform time-frequency map and intrinsic mode function components through a dual-branch feature extraction backbone network, and output a fusion feature vector after splicing and fusing through a fusion layer. The signal reconstruction and Bayesian classification module is used to map the fused feature vector into the signal reconstruction decoding branch to generate a temporal reconstructed signal and construct a physical consistency loss. Simultaneously, the fused feature vector is input into a Bayesian classifier, and variational inference learning is used to obtain the variational approximate posterior distribution of the Bayesian layer network weights, outputting a fault category probability distribution. This variational approximate posterior distribution is used to approximate the true posterior distribution, and the Bayesian layer network weight distribution is iteratively updated by minimizing the KL divergence between the two. The joint optimization module is used to jointly optimize the classification cross-entropy loss, KL divergence regularization loss, signal reconstruction loss, and physical consistency loss, and to perform iterative training and parameter updates of the model. A comprehensive uncertainty metric construction module is used to obtain the fault category probability distribution of the Bayesian classifier through Monte Carlo random deactivation, quantify the uncertainty of the fault category probability distribution using classification prediction entropy, and divide the total uncertainty into cognitive uncertainty and accidental uncertainty; the average mean square error between the reconstructed signal obtained by Monte Carlo sampling and the original signal is used to define the reconstruction quality; and the comprehensive uncertainty metric is constructed by fusing the classification prediction entropy and the reconstruction quality. The credibility assessment and risk control module is used to rank the reliability of sample predictions using the comprehensive uncertainty metric, and to achieve credibility assessment and risk control of fault diagnosis results through a high uncertainty sample rejection mechanism.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction that is loaded and executed by the processor to implement the method as described in any one of claims 1 to 7.

Images