Partial discharge capacity-severity joint estimation method, product and equipment
By using graph-guided residual networks and variational evidence inference methods, the problem of multi-source data fusion and evaluation separation in non-contact partial discharge detection was solved, achieving joint estimation of partial discharge capacity and severity, and improving the accuracy and reliability of detection.
Patent Information
- Application Number
- CN202511538618.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-25
- Publication Date
- 2026-01-27
AI Technical Summary
Existing non-contact partial discharge detection technologies suffer from problems such as difficulty in fusing multi-source data, separation of evaluation methods and lack of physical connection, and difficulty in explaining and evaluating uncertainties by deep learning models, which affect the accuracy and reliability of power equipment.
A physical fusion framework for multimodal data is constructed by employing graph wavelet physical-guided residual networks and variational evidence inference methods. Through graph wavelet modeling and Bayesian uncertainty inference, joint estimation of partial discharge capacity and severity is achieved, combined with continuous signal quantization and discrete level discrimination.
It improves the accuracy and reliability of partial discharge detection, enhances the interpretability and uncertainty assessment capabilities of the model, and is suitable for non-contact discharge detection under complex media conditions.
Smart Images

Figure CN121412914A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power equipment monitoring technology, specifically relating to a method, product, and equipment for joint estimation of partial discharge capacity-severity based on graph wave physical guided residual network and variational evidence inference. Background Technology
[0002] A joint estimation method for partial discharge that integrates graph neural networks, graph wavelets, multimodal physical propagation modeling, and Bayesian uncertainty reasoning has both continuous signal quantization and discrete level discrimination capabilities, making it particularly suitable for non-contact discharge detection scenarios under complex media conditions.
[0003] In recent years, non-contact detection technology has emerged as a research and application hotspot in the industry due to its unique advantages. This type of technology encompasses various methods such as ultrasonic sensing, ultra-high frequency electromagnetic wave detection, and infrared thermal imaging. It can acquire information from multiple dimensions, including sound, light, heat, and electromagnetic fields. It is not only flexible in deployment but also adaptable to complex environments, opening up new avenues for online equipment monitoring.
[0004] However, partial discharge status assessment based on non-contact multimodal sensing still faces many technical challenges. First, multi-source data fusion is difficult. Different types of sensors, such as acoustic microphones, UHF antennas, and infrared cameras, collect data with vastly different formats, scales, physical meanings, and noise characteristics. Currently, there is a lack of a unified fusion framework that conforms to physical laws, making it difficult to comprehensively reflect the true situation of the discharge source within complex equipment structures and propagation paths. This results in ineffective information integration, affecting the accuracy of the assessment. Second, existing assessment methods have shortcomings. Most studies either focus on quantifying discharge physical parameters, such as apparent discharge quantity and repetition rate, or emphasize classifying fault severity levels, failing to effectively combine the two. This separate assessment model cannot establish the intrinsic link between discharge intensity and insulation degradation risk, making it difficult to provide comprehensive risk judgment criteria and limiting the depth of status assessment and early warning effectiveness. Finally, mainstream deep learning models have limitations. Although deep learning performs well in partial discharge detection and classification, its "black box" nature makes prediction results difficult to interpret, and the reliability of the predictions cannot be assessed. In scenarios like power equipment safety monitoring, where the accuracy of decision-making is extremely important, the inability to quantify the uncertainty of model predictions can easily lead to misjudgments or omissions, seriously affecting the reliability of diagnostic results.
[0005] Therefore, there is an urgent need for innovative methods for assessing partial discharge status. New methods must deeply integrate multimodal physical information, organically combining discharge intensity quantification with fault severity assessment, and also be able to evaluate the uncertainty of prediction results. Furthermore, they need a solid theoretical foundation in physics to enhance interpretability, powerful data fusion capabilities to improve accuracy, and a rigorous uncertainty reasoning mechanism to ensure reliability and stability in critical application scenarios, thereby meeting the requirements for accurate equipment diagnosis and risk warning in fields such as smart grids and rail transportation. Summary of the Invention
[0006] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.
[0007] Therefore, the purpose of this invention is to provide a method, product, and device for joint estimation of partial discharge capacity and severity based on graph wave physics-guided residual network and variational evidence inference. It integrates graph neural network, graph wavelet, multimodal physical propagation modeling, and Bayesian uncertainty reasoning, and has both continuous signal quantization and discrete level discrimination capabilities. It is especially suitable for non-contact discharge detection scenarios under complex media conditions.
[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a method for joint estimation of partial discharge capacity and severity, the method comprising: S1. Construct a physical fusion framework for multimodal data: Based on the deployment location of non-contact sensors, construct a partial discharge signal propagation map structure to transform multi-source heterogeneous data into a unified physical map; S2. Establish a physical prior model for partial discharge propagation: Based on the Laplace frequency domain characteristics of the graph, design a multi-scale graph wavelet kernel function with bandpass properties to achieve accurate modeling of the propagation characteristics of partial discharge signals; S3. Compensating for physical model bias: Constructing a GWP-ResNet correction network to compensate for physical model bias and improve estimation accuracy; S4. Perform joint capacity-severity estimation and uncertainty quantification: jointly model continuous discharge quantity and discrete severity level, approximate the true posterior through variational posterior, and output point estimates and confidence information; S5. Risk Assessment and Adaptive Optimization: Based on the results of S4, risk level assessment is performed and adaptive adjustments and optimizations are made when false alarms occur.
[0009] In addition, the partial discharge capacity-severity joint estimation method according to the present invention may also have the following additional technical features: In some implementations, the physical map of step S1 includes: node VEach node corresponds to one sensor installation point. The feature vector includes signal amplitude, signal arrival time, and signal main frequency, which are used to reflect the energy, timing, and spectral characteristics of the partial discharge. Edge set E Characterizes the reachability of signal propagation between sensors; Border rights W : Integrating spatial distance and time delay, signal attenuation is quantified using an exponential decay model; Frequency domain characteristics: Construct the graph Laplace matrix L and perform spectral decomposition to obtain the frequency domain characteristics that reflect the frequency response of the nodal signal coupling.
[0010] In some implementations, the multi-scale graph wavelet kernel function of step S2 is:
[0011] in, These are the Laplace eigenvalues of the graph; It controls the degree of enhancement of low-frequency components; The function controls the attenuation rate of high-frequency components; its construction enables the graph wavelet operator at different scales to have bandpass properties in the spectrum, achieving frequency-selective response for different discharge types.
[0012] In some implementations, the physical prior model for step S2 is constructed by: S Wavelet operators are constructed at each preset scale to form a set of multi-scale frequency domain projection kernel functions. The original discharge power source signal is input into these kernel functions, and by combining wavelet filtering and spatial propagation mechanisms, a physical prior prediction model is obtained.
[0013] In some implementations, the GWP-ResNet correction network in step S3 is a correction network that combines physical priors with residual learning. Its core is to obtain superimposed residuals through a residual learning mechanism and optimize the prediction results of the physical model based on these residuals.
[0014] In some implementations, step S3 includes: Input feature construction: For each node, the physical prediction value, sensor observation value, and local gradient norm of the physical prediction value are fused to form an input vector; Graph convolution residuals: Implemented through several sequentially connected graph convolution residual layers, in the... l The layer updates features by aggregating neighbor node information and connecting it with residuals; Wavelet attention mechanism: By introducing a wavelet attention mechanism into the network, key frequency bands related to specific discharge types are obtained, and weights at different scales are obtained; Residual Output and Correction: A three-layer fully connected network is used to map the weighted features into predicted residuals. It is used to correct the prediction results of the physical model and achieve correction.
[0015] In some of these implementations, the model built in step S4 is a variational evidence inference model. The model describes the joint probability distribution of discharge amount and severity as follows:
[0016]
[0017] in, y For the discharge quantity, it obeys the following: For the mean, The variance follows a normal distribution, which is used to effectively characterize the fluctuation characteristics of the discharge quantity in actual detection. For severity levels, follow the guidelines. For the class distribution of parameters, Dirichlet distribution (by Dirichlet) Generates evidence vectors. To ensure the validity of the distribution, the model can reasonably describe the probability distribution of different severity levels. This model approximates the difficult-to-calculate true posterior distribution using variational posterior parameterization, which modifies the mean separately. ,variance and Dirichlet distribution parameters Modeling is performed by optimizing variational parameters. This makes the variational posterior distribution as close as possible to the true posterior distribution, thereby improving the inference accuracy and reliability of the model.
[0018] In some of these implementations, the loss function L of the variational evidence inference model comprehensively considers regression loss, physical consistency constraints, and KL regularization terms; Among them, regression loss is used to measure the difference between the predicted discharge amount and the actual discharge amount, so as to enable the model to accurately predict the discharge amount; physical consistency constraint ensures that the model prediction results are consistent with the prior estimates based on the physical model, thereby improving the rationality of the prediction; KL regularization term avoids model overfitting by controlling the uncertainty of variational posterior distribution; two control parameters are introduced to control the strength of physical consistency constraint and the degree of uncertainty control, respectively.
[0019] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the partial discharge capacity-severity joint estimation method as described in any of the preceding embodiments.
[0020] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the partial discharge capacity-severity joint estimation method as described in any of the preceding embodiments.
[0021] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the provided partial discharge capacity-severity joint estimation method integrates graph structure physics propagation and residual deep learning model to achieve the unification of physical interpretation and high-precision prediction, thereby improving the reliability of estimation. In this embodiment of the invention, the provided partial discharge capacity-severity joint estimation method introduces a graph wavelet transform mechanism to effectively capture the propagation characteristics of discharge in structural media and enhance the model's adaptability to complex propagation paths. In this embodiment of the invention, the provided partial discharge capacity-severity joint estimation method is based on variational Bayesian inference outputting the joint posterior of continuous and discrete variables, which has the ability to quantify uncertainty and improves the availability and robustness of the model in safety-critical scenarios.
[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0023] Figure 1 This is a flowchart of a method for joint estimation of partial discharge capacity-severity based on graph wave physics-guided residual network and variational evidence inference, disclosed in one embodiment of the present invention. Figure 2 This is a diagram of a graph wavelet kernel function and a graph convolutional attention structure disclosed in one embodiment of the present invention; Figure 3 This is a schematic diagram of the joint posterior estimation result disclosed in an embodiment of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.
[0026] Please see Figure 1As shown, in some embodiments of the present invention, a joint estimation method for partial discharge capacity-severity based on graph wave physics-guided residual networks and variational evidence inference is provided, including the following steps: Step 1: Construct a partial discharge map structure propagation model based on non-contact observation nodes to form a discharge physical response expression of the map structure; When deploying a partial discharge detection system, non-contact sensors such as ultrasonic sensors and UHF antennas are installed at different locations within the equipment. To accurately represent the propagation characteristics of partial discharge signals in complex power equipment structures, a graphical structure is first introduced. To characterize the sensor observation network. The nodes... Corresponding sensor mounting points, edge sets E Characterizing the signal propagation path between sensors, edge weights W This reflects the degree of signal energy attenuation during propagation between different sensors.
[0027] (1) Node
[0028] Each node The feature vector corresponding to the sensor installation location is defined as follows:
[0029] in, It is the signal amplitude, in units of ; It is the signal arrival time, in units of ; It is the main frequency component of the signal, and the unit is . Together, they constitute a preliminary representation of the spatial characteristics of the partial discharge signal, thereby effectively reflecting the local energy, timing, and spectral characteristics of the partial discharge signal.
[0030] (2) Edge set E Edge set in a graph structure E Characterizing whether a valid signal coupling path exists between any two nodes, it is constructed based on the reachability of partial discharge signals in physical space and is defined as:
[0031] Here It is a reachability indicator function used to determine whether an edge can be established between a pair of nodes. Represents a node and There exists an effective signal coupling propagation path. If , then it represents a node and They are physically and temporally inaccessible and do not constitute an edge.
[0032] (3) Edge weight W To fully integrate spatial physical location and signal time-domain propagation characteristics, edge weights W The exponential decay model is used to represent the signal decay process in space and time:
[0033] in, The spatial Euclidean distance between sensor nodes; This represents the time difference between the arrival of the discharge signal at the two nodes; This represents the typical attenuation coefficient of a medium, such as in sulfur hexafluoride gas. =0.5, in transformer oil =1.2; The adjustment parameter for controlling the intensity of the time delay penalty is typically 0.3.
[0034] Based on the graph structure, a graph Laplacian matrix is further constructed. L Its non-normalized and normalized expressions are as follows:
[0035]
[0036] Among them D Let the node degree matrix be... I It is the identity matrix. (Regarding...) L Perform spectral decomposition The eigenvector matrix of the graph can be obtained. U With frequency response eigenvalue matrix These spectral components reflect the energy response capability at different frequencies or propagation modes under the graph structure, that is, the frequency domain representation of the signal coupling relationship between nodes.
[0037] Step 2: Design a multi-scale graph wavelet kernel function to perform frequency domain analysis on signal propagation in the graph structure and construct a physical prior estimation model; Based on the graph Laplace frequency domain representation, a multi-scale graph wavelet kernel function is constructed. This function is used to adjust the signal response in the low-frequency and high-frequency regions to characterize the spatial spread, energy attenuation, and path diversity of partial discharge signals. The function is controlled by parameters... and Controlling the response of each scale to signals in different frequency bands, the first s The wavelet kernel function at each scale is defined as:
[0038] in, These are the Laplace eigenvalues of the graph; It controls the degree of enhancement of low-frequency components; Controlling the attenuation rate of high-frequency components. This function constructs graph wavelet operators at different scales. It has bandpass properties in the spectrum, thus enabling frequency-selective response for different discharge types.
[0039] Through S Wavelet operators are constructed at each preset scale to form a set of multi-scale frequency domain projection kernel functions. The original discharge signal... The input is combined with wavelet filtering and spatial propagation mechanisms to obtain the physical prior prediction results:
[0040] in, This represents the measurement noise introduced by the sensing terminal, typically a zero-mean Gaussian distribution. For its variance, The eigenvector matrix, Eigenvalue diagonal matrix, It is the diagonal operator of the graph wavelet kernel at the s-th scale in the spectral domain. This formula can be understood as convolving the local discharge signal on the graph structure through a multi-scale propagation kernel, and outputting the predicted response value at the sensor node.
[0041] Low-frequency wavelet kernels enhance the response of signals with high global connectivity, long propagation paths, but slow energy attenuation, and can simulate slowly propagating signals such as bubble discharge in oil. High-frequency kernels, on the other hand, respond to sharp, locally coupled pulse-like signals, such as transient high-frequency events like metal tip discharge. By adjusting... and The parameter set can be adapted to the corresponding spectral structure according to different discharge modes to form a physically interpretable response prediction model. To enhance the modeling capability of discharge characteristics under different media, the recommended parameter configurations for common media and discharge types are shown in Table 1.
[0042] Table 1 Parameter Configuration
[0043] Step 3: Construct a graph wavelet physics-guided residual network (GWP-ResNet), which integrates graph wavelet output and node features to learn the nonlinear residuals between the physical model and actual observations; To effectively compensate for biases in physical models during practical applications, a GWP-ResNet correction network was constructed. The core of this network lies in optimizing the prediction results of the physical model through a residual learning mechanism. Its basic principle is based on the formula:
[0044] in, These are the initial predictions from the physical model. The residuals are calculated by the correction network. This is the final prediction result after correction. The correction network corrects the bias of the physical model by superimposing residuals.
[0045] (1) Input feature construction In the input feature construction stage, for each node i , physical prediction value Sensor observations x i and the local gradient norm of the physical prediction value The input vector is fused together. (That is, the initial features) are as follows:
[0046] Among them, physical prediction value The graph wavelet model is for the first i The predicted initial discharge amount at each node provides a preliminary estimate based on theory and prior knowledge; sensor observations x i Recorded nodes i Information including signal amplitude, time of arrival, and dominant frequency components; actual detected signal characteristics; local gradient norm. Nodes in a graph structure i The spatial gradient reflects the rate of change of discharge quantity in the neighborhood, capturing abrupt structural changes. The combination of these three elements provides rich and complementary feature information for subsequent network processing, helping the network to more accurately learn the deviation patterns of the physical model.
[0047] (2) Figure Convolution Residual Layer The graph convolutional residual layer, as a key structure of the network, plays a crucial role in the [missing information - likely a specific function or role]. l The layer updates features by aggregating neighbor node information and using residual connections. For nodes... i Its update formula is:
[0048] in, Let i represent the set of neighboring nodes directly connected to node i. Represents any of the neighboring nodes j ; For the edge The weights; For a degree matrix, diagonal elements Information transmission employs symmetrical normalization: This is to ensure that messages between different nodes are comparable; It is a trainable weight matrix that can adaptively learn the transformation of node features; The GeLU activation function is used to introduce non-linearity into the network and enhance its ability to express complex features. F(·) is a residual skip connection, which directly passes the features of the previous layer to the current layer through the fully connected layer, effectively alleviating the gradient vanishing problem in the training process of deep networks, and enabling the network to learn more complex and abstract feature mapping relationships.
[0049] (3) Wavelet attention mechanism The network also incorporates a wavelet attention mechanism to obtain key frequency bands related to specific discharge types, resulting in different scales. s The weights below :
[0050] in, h (L) The output features of the last layer of the network; It is based on graph Laplace spectral decomposition to realize multi-scale graph wavelet transform; d The feature dimension is defined by the inner product operation and the softmax function. The network then automatically learns the importance of each scale to the current task. When detecting insulator discharge, the network strengthens the scale, highlighting the frequency band features related to the insulator discharge at that scale, thereby achieving accurate identification and feature extraction of specific discharge types.
[0051] (4) Residual output The residual output section utilizes a three-layer fully connected network (MLP) to weight the features. ⊙ h (L) Mapped to prediction residuals It is used to correct the prediction results of the physical model and complete the entire correction process.
[0052]
[0053] Step 4: Establish a variational evidence inference (VEI) model to jointly estimate the discharge amount and severity level, and achieve joint posterior inference and uncertainty estimation; The variational evidence inference model of this invention, through joint modeling of continuous discharge quantity y and discrete severity level c ∈{1,2,... K This allows for a comprehensive and accurate assessment of the partial discharge state, and the joint probability distribution describing the discharge quantity and severity is as follows:
[0054]
[0055] in, It is the probability density of a normal (Gaussian) distribution, and the discharge quantity y Obey For the mean, The variance follows a normal distribution, which effectively characterizes the fluctuation characteristics of discharge quantity in actual detection; severity level. Obey For the class distribution of parameters, It is the probability density of the categorical distribution, and the discrete variable is... The single-order multinomial distribution, Dirichlet distribution (by Dirichlet) ) generated, and Through evidence vectors Ensuring the validity of the distribution ensures that the model can reasonably describe the probability distribution of different severity levels.
[0056] To approximate the true posterior distribution, which is difficult to compute during model training, a variational posterior parameterization method is employed:
[0057] It is a normal distribution. It is a Gamma distribution (this invention uses shape-rate parameterization, shape) ,Rate ), It is the Dirichlet distribution function. yes The variational posterior parameters (mean, variance). yes The variational posterior parameters (shape and rate). yes Variational posterior concentration parameters ( (This represents the set of variational parameters obtained from the network output), and the above uses factorization approximation: This parameterization method facilitates optimization and calculation, and applies parameters to the mean. ,variance and Dirichlet distribution parameters Modeling is performed by optimizing variational parameters. This makes the variational posterior distribution as close as possible to the true posterior distribution, thereby improving the inference accuracy and reliability of the model.
[0058] The model's loss function L comprehensively considers regression loss, physical consistency constraints, and KL regularization:
[0059] Among them, Pphys It is the predicted response of the physical prior model (obtained in step S2), and the regression loss. Used to measure the difference between predicted discharge and actual discharge. These are tradeoff coefficients (default sample values are typically 0.5 or 0.1, which can be calibrated based on data) that enable the model to accurately predict discharge values; physical consistency constraints. To ensure consistency between model predictions and prior estimates based on the physics model, thus improving the reasonableness of the predictions; KL regularization term. By controlling the uncertainty of the variational posterior distribution, overfitting of the model can be avoided and the generalization ability of the model can be enhanced. The strength of the physical consistency constraint and the degree of uncertainty control can be controlled separately, and can be flexibly adjusted according to the actual application scenario.
[0060] The graph shows the wavelet kernel function and the graph convolutional attention structure. Figure 2 As shown.
[0061] Step 5: Output the predicted discharge amount, confidence interval, severity level and its confidence distribution to achieve multi-level risk assessment and early warning; The inference output and risk alarm mechanism constructed in this invention can quantitatively assess the partial discharge state and issue corresponding risk alarms in a timely manner based on the assessment results. It also has adaptive optimization capabilities to adapt to different application scenarios and data changes.
[0062] In terms of uncertainty quantification, the discharge quantity is calculated to measure the range of uncertainty in the prediction. The 95% confidence interval is used to reflect the reliability of the predicted value:
[0063] Figure 3 Figure (a) is a probability density diagram of discharge quantity prediction according to an embodiment of the present invention.
[0064] For severity, the expected information entropy is used to measure the uncertainty of the model:
[0065] in The higher the value, the lower the confidence level of the severity level judgment, providing users with information on the credibility of the severity prediction results, which helps users make reasonable judgments and decisions based on the assessment results. Figure 3 Figure (b) in the figure is a probability distribution diagram of severity levels according to an embodiment of the present invention, showing the probability of three severity levels: defect, degradation and breakdown.
[0066] Risk warning rules are formulated based on three key indicators: discharge amount, severity probability, and uncertainty. Table 2 Risk Alert Rules
[0067] (1) When the discharge amount Severity is the probability of breakdown. b k (Breakdown) > 0.8 and uncertainty When a situation is deemed to be at an emergency risk level, immediate measures must be taken to address it in order to avoid potential serious malfunctions.
[0068] (2) When the discharge amount The probability of severity deterioration b k (Deterioration) > 0.7 and If the situation is deemed high-risk, close monitoring of the equipment's status and further testing are required to promptly identify and address any potential problems.
[0069] (3) When the discharge amount The probability of a defect in terms of severity. b k (Defect) > 0.6 and If the risk level is determined to be medium, the equipment must be included in the routine monitoring scope and its operating status must be continuously tracked.
[0070] The category probability is calculated using variational posterior parameters to ensure the accuracy and reliability of risk assessment. The calculation formula is as follows:
[0071] To enable the model to adapt to different application scenarios and data changes, this invention designs an adaptive adjustment mechanism. When a false alarm occurs, i.e., the actual situation does not match the model prediction, the model parameters are updated according to the following formula:
[0072] Among them, learning rate The value is typically 0.01 to control the step size of parameter updates. The false alarm indicator function I is 1 for false alarms and 0 otherwise, ensuring that parameter adjustments are only made in cases of false alarms. Furthermore, the model dynamically optimizes threshold parameters according to the IEC 60270 standard, further improving its adaptability and accuracy, enabling reliable operation under different conditions and providing stable and accurate support for partial discharge detection and risk assessment.
[0073] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.
[0074] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.
Claims
1. A method for joint estimation of partial discharge capacity and severity, characterized in that, The method includes: S1. Construct a physical fusion framework for multimodal data: Based on the deployment location of non-contact sensors, construct a partial discharge signal propagation map structure to transform multi-source heterogeneous data into a unified physical map; S2. Establish a physical prior model for partial discharge propagation: Based on the Laplace frequency domain characteristics of the graph, design a multi-scale graph wavelet kernel function with bandpass properties to achieve accurate modeling of the propagation characteristics of partial discharge signals; S3. Compensating for physical model bias: Constructing a GWP-ResNet correction network to compensate for physical model bias and improve estimation accuracy; S4. Perform joint capacity-severity estimation and uncertainty quantification: jointly model continuous discharge quantity and discrete severity level, approximate the true posterior through variational posterior, and output point estimates and confidence information; S5. Risk Assessment and Adaptive Optimization: Based on the results of S4, risk level assessment is performed and adaptive adjustment and optimization are carried out when false alarms occur.
2. The partial discharge capacity-severity joint estimation method according to claim 1, characterized in that, The physical map of step S1 includes: node V Each node corresponds to one sensor installation point. The feature vector includes signal amplitude, signal arrival time, and signal main frequency, which are used to reflect the energy, timing, and spectral characteristics of the partial discharge. Edge set E Characterizes the reachability of signal propagation between sensors; Border rights W : Integrating spatial distance and time delay, signal attenuation is quantified using an exponential decay model; Frequency domain characteristics: Construct the graph Laplace matrix L and perform spectral decomposition to obtain the frequency domain characteristics that reflect the frequency response of the nodal signal coupling.
3. The partial discharge capacity-severity joint estimation method according to claim 1, characterized in that, The multi-scale wavelet kernel function for step S2 is: ; in, These are the Laplace eigenvalues of the graph; Control the degree of enhancement of low-frequency components; The function controls the attenuation rate of high-frequency components; its construction enables the graph wavelet operator at different scales to have bandpass properties in the spectrum, achieving frequency-selective response for different discharge types.
4. The partial discharge capacity-severity joint estimation method according to claim 1, characterized in that, The physical prior model in step S2 is constructed by: through S Wavelet operators are constructed at each preset scale to form a set of multi-scale frequency domain projection kernel functions. The original discharge power source signal is input into these kernel functions, and a physical prior prediction model is obtained by combining wavelet filtering and spatial propagation mechanisms.
5. The partial discharge capacity-severity joint estimation method according to claim 1, characterized in that, The GWP-ResNet calibration network in step S3 is a calibration network that combines physical priors with residual learning. Its core is to obtain superimposed residuals through the residual learning mechanism, and then optimize the prediction results of the physical model based on these residuals.
6. The partial discharge capacity-severity joint estimation method according to claim 5, characterized in that, Step S3 includes: Input feature construction: For each node, the physical prediction value, sensor observation value, and local gradient norm of the physical prediction value are fused to form an input vector; Graph convolution residuals: Implemented through several sequentially connected graph convolution residual layers, in the... l The layer updates features by aggregating neighbor node information and connecting it with residuals; Wavelet attention mechanism: By introducing a wavelet attention mechanism into the network, key frequency bands related to specific discharge types are obtained, and weights at different scales are obtained; Residual Output and Correction: A three-layer fully connected network is used to map the weighted features into predicted residuals. It is used to correct the prediction results of the physical model and achieve correction.
7. The partial discharge capacity-severity joint estimation method according to claim 1, characterized in that, The model built in step S4 is a variational evidence inference model; The model describes the joint probability distribution of discharge amount and severity as follows: ; ; in, y For the discharge quantity, it obeys the following: For the mean, The variance follows a normal distribution, which is used to effectively characterize the fluctuation characteristics of the discharge quantity in actual detection. For severity levels, follow the guidelines. For the class distribution of parameters, Dirichlet distribution (by Dirichlet) Generates evidence vectors. To ensure the validity of the distribution, the model can reasonably describe the probability distribution of different severity levels. This model approximates the difficult-to-calculate true posterior distribution using variational posterior parameterization, which modifies the mean separately. ,variance and Dirichlet distribution parameters Modeling is performed by optimizing variational parameters. This makes the variational posterior distribution as close as possible to the true posterior distribution, thereby improving the inference accuracy and reliability of the model.
8. The partial discharge capacity-severity joint estimation method according to claim 7, characterized in that, The loss function L of the variational evidence inference model comprehensively considers regression loss, physical consistency constraints, and KL regularization. Among them, regression loss is used to measure the difference between the predicted discharge amount and the actual discharge amount, so as to enable the model to accurately predict the discharge amount; physical consistency constraint ensures that the model prediction results are consistent with the prior estimates based on the physical model, thereby improving the rationality of the prediction; KL regularization term avoids model overfitting by controlling the uncertainty of variational posterior distribution; two control parameters are introduced to control the strength of physical consistency constraint and the degree of uncertainty control, respectively.
9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the partial discharge capacity-severity joint estimation method according to any one of claims 1-8.
10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the partial discharge capacity-severity joint estimation method according to any one of claims 1-8.
Citation Information
Cited By
Campus abnormal behavior identification method and system and storage medium
CN121723357A
