Insulation state on-line monitoring method and system for power transmission and transformation equipment in new energy station

Abstract

This invention provides a method and system for online monitoring of the insulation status of power transmission and transformation equipment in new energy power plants. The method includes first collecting the equipment's operating sequence data, obtaining multi-resolution components through wavelet packet decomposition, calculating the covariance matrix through a covariance attention module, mapping it to a Riemannian manifold to form a covariance manifold representation, and simultaneously constructing a prior knowledge graph that integrates physical adjacency and electromagnetic coupling relationships. Graph regularization constraints are input during attention training to obtain a structured attention weight matrix. The multi-resolution components are weighted and fused to obtain deep fusion features, which are then parameterized using a Dirichlet distribution. An uncertainty metric is calculated. If the uncertainty is higher than a threshold, the importance of the components is evaluated based on the weights, and the wavelet packet decomposition parameters are adjusted for re-decomposition. If the uncertainty is lower than or equal to the threshold, the multi-class probability of the insulation status is output based on the expected distribution, thus achieving online insulation status monitoring.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to an online monitoring method and system for the insulation status of power transmission and transformation equipment in new energy power plants. Background Technology

[0002] A large number of new power transmission and transformation equipment have been connected to the power grid. The health status of these devices, especially their insulation systems, is the cornerstone of ensuring the safe and stable operation of the entire new energy power plant and even the power grid. Traditional periodic inspections or offline testing methods can no longer meet the needs of new energy power plants for real-time and accurate monitoring of equipment status. Therefore, the development of online monitoring and fault early warning technologies has become an urgent need in the industry. However, the operating environment of new energy power plants is extremely complex. The intermittency and fluctuation of power generation make the operating conditions of transmission and transformation equipment highly variable, and the monitoring data exhibits typical non-stationary and high-noise characteristics. Existing online monitoring methods mostly rely on threshold judgment or shallow statistical models, which cannot extract weak signals indicating early insulation degradation from strong interference backgrounds, often resulting in false alarms or missed alarms, and failing to achieve true predictive maintenance.

[0003] Utilizing data-driven methods such as deep learning can improve the performance of monitoring models. By automatically learning complex dependencies in time-series data using models such as recurrent neural networks and attention mechanisms, the accuracy of state assessment has been improved to some extent. However, standard attention mechanisms, when assessing the importance of different features, often fail to consider the high-order statistical information inherent in multi-source monitoring signals and are highly sensitive to outliers and noise in the data, easily leading to biases in the allocation of attention weights and affecting model judgment. Most models are purely data-driven black boxes, failing to incorporate prior knowledge such as the objective physical connections and electromagnetic couplings between components of power transmission and transformation equipment. This results in the learned feature correlations potentially contradicting reality, reducing the model's generalization ability and interpretability. Existing methods mostly employ fixed feature extraction frameworks and lack mechanisms for evaluating the uncertainty of their own prediction results. They cannot adjust the front-end signal processing strategy based on the model's reliability, and their adaptability is severely tested when faced with unfamiliar operating condition data. Summary of the Invention

[0004] To address the issues that existing technologies fail to incorporate the objective prior knowledge between components of power transmission and transformation equipment, lack an assessment mechanism for the uncertainty of their own prediction results, and are unable to adjust the front-end signal processing strategy based on the model's credibility.

[0005] In one aspect, this disclosure provides a method for online monitoring of the insulation status of power transmission and transformation equipment in new energy power plants, including: The operation monitoring time series data of the power transmission and transformation equipment of the new energy power station is acquired, and the time series data is decomposed by wavelet packet to generate a set of multi-resolution components. The multi-resolution components are used as features and input to the covariance attention module to calculate the covariance matrix between features. The covariance matrix is ​​then mapped to the tangent space of the Riemann manifold through the log-Euclidean metric to generate the covariance manifold representation of the features. A prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components is constructed. During the training process of the covariance attention module, a graph regularization term is constructed using the prior knowledge graph to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge. The structured attention weight matrix is ​​applied to the multi-resolution components for weighted fusion to obtain deep fusion features. A Dirichlet distribution is parameterized using the deep fusion features, and a model uncertainty metric is calculated based on the Dirichlet distribution. If the uncertainty metric is higher than a threshold, the importance of each multi-resolution component is evaluated according to the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results. Then, the wavelet packet decomposition step is re-executed. Otherwise, the expected value of the Dirichlet distribution is calculated as the multi-class probability output of the device insulation state.

[0006] Optionally, the step of performing wavelet packet decomposition on the time-series data to generate a set of multi-resolution components includes: The original signal is completely decomposed to the preset maximum number of layers by pre-setting wavelet basis functions. Then, the tree is traversed from bottom to top. The information entropy criterion of node energy or Shannon entropy is used to determine whether to prune, thereby constructing the optimal wavelet packet decomposition tree. The multi-resolution components corresponding to the leaf nodes of the tree are used as output.

[0007] Optionally, the step of inputting the multi-resolution components as features into the covariance attention module and calculating the covariance matrix between features includes: A subset containing a specific proportion of samples is randomly initialized from all observed samples; The covariance matrix and Mahalanobis distance between the multi-resolution component features are calculated based on the subset, and the samples are filtered according to the Mahalanobis distance to update the subset; The update process is iteratively executed until the subset is stable, and the covariance matrix calculated from the stable subset is used as the result covariance matrix.

[0008] Optionally, the construction of a prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components includes: Each multi-resolution component is abstracted as a node in the graph; If the components monitored by the monitoring channels of two multi-resolution components have a physical adjacency relationship in the equipment structure, then an edge with a first preset weight is established between the corresponding nodes. If the electromagnetic coupling coefficient between the components monitored by the monitoring channels of two multi-resolution components is greater than the preset coupling threshold calculated by finite element simulation, then an edge with a weight equal to the coupling coefficient is established between the corresponding nodes.

[0009] Optionally, during the training process of the covariance attention module, a graph regularization term is constructed using the prior knowledge graph to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge, including: The Laplacian matrix of the prior knowledge graph is embedded into the loss function of the covariance attention module; The regularization term is in the form of ,in λ is the normalized Laplacian matrix, W is the attention weight matrix to be learned with the same number of rows as the dimension of the Laplacian matrix, and λ is the regularization hyperparameter used to ensure that the features corresponding to the connected nodes in the graph receive similar attention weights.

[0010] Optionally, parameterizing a Dirichlet distribution using the deep fusion features includes: The deep fusion features are input into a fully connected neural network layer, and the number of output nodes of the network layer is equal to the preset number N of insulation state categories; After applying a non-negative activation function to the output of the network layer, a bias greater than zero is added to ensure that all outputs are greater than zero, resulting in N concentration parameters. This constitutes the parameter vector α of the Dirichlet distribution, which is composed of the concentration parameter. to composition.

[0011] Optionally, if the uncertainty metric is higher than a threshold, the importance of each multi-resolution component is evaluated based on the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results, including: The threshold is the statistical value of uncertainty measurement over the most recent monitoring periods; When the uncertainty measure of the current period exceeds the threshold, the structured attention weight matrix is ​​summed by rows or columns to obtain the cumulative attention weight of each multi-resolution component. Components with cumulative weights higher than preset quantile values ​​are selected, and the time-frequency region features corresponding to the components are analyzed. Based on the features, one or more decomposition parameters of the subsequent wavelet packet decomposition are adjusted. The decomposition parameters include the number of decomposition layers or node selection strategies.

[0012] Optionally, calculating the expected value of the Dirichlet distribution as the multi-class probability output of the equipment insulation state includes: Using concentration parameters to The parameter vector α is used to calculate the total concentration. The total concentration For all concentration parameters The sum; For the k-th insulation state category, the output probability is obtained by passing the concentration parameter of the category. Divide by total concentration Calculations were performed to obtain a set of normalized probability values ​​representing the possibilities of different insulation states.

[0013] On the other hand, this disclosure also provides an online monitoring system for the insulation status of power transmission and transformation equipment in new energy power plants, including the following modules: The generation module is used to acquire the operation monitoring time series data of the power transmission and transformation equipment of the new energy power station, perform wavelet packet decomposition on the time series data to generate a set of multi-resolution components; input the multi-resolution components as features to the covariance attention module, calculate the covariance matrix between features, and map the covariance matrix to the tangent space of the Riemann manifold through the log-Euclidean metric to generate the covariance manifold representation of the features. A construction module is used to construct a prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components. During the training process of the covariance attention module, the prior knowledge graph is used to construct a graph regularization term to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge. The calculation module is used to apply the structured attention weight matrix to the multi-resolution components for weighted fusion to obtain deep fusion features; parameterize a Dirichlet distribution using the deep fusion features, and calculate the model uncertainty measure based on the Dirichlet distribution; if the uncertainty measure is higher than a threshold, evaluate the importance of each multi-resolution component according to the structured attention weight matrix, adjust the decomposition parameters of the wavelet packet decomposition based on the importance evaluation result, and then return to re-execute the wavelet packet decomposition step; otherwise, calculate the expected value of the Dirichlet distribution as the multi-class probability output of the device insulation state.

[0014] Further, the wavelet packet decomposition of the time-series data to generate a set of multi-resolution components includes: The original signal is completely decomposed to the preset maximum number of layers by pre-setting wavelet basis functions. Then, the tree is traversed from bottom to top. The information entropy criterion of node energy or Shannon entropy is used to determine whether to prune, thereby constructing the optimal wavelet packet decomposition tree. The multi-resolution components corresponding to the leaf nodes of the tree are used as output.

[0015] Further, the step of inputting the multi-resolution components as features into the covariance attention module and calculating the covariance matrix between features includes: A subset containing a specific proportion of samples is randomly initialized from all observed samples; The covariance matrix and Mahalanobis distance between the multi-resolution component features are calculated based on the subset, and the samples are filtered according to the Mahalanobis distance to update the subset; The update process is iteratively executed until the subset is stable, and the covariance matrix calculated from the stable subset is used as the result covariance matrix.

[0016] Furthermore, the construction of the prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components includes: Each multi-resolution component is abstracted as a node in the graph; If the components monitored by the monitoring channels of two multi-resolution components have a physical adjacency relationship in the equipment structure, then an edge with a first preset weight is established between the corresponding nodes. If the electromagnetic coupling coefficient between the components monitored by the monitoring channels of two multi-resolution components is greater than the preset coupling threshold calculated by finite element simulation, then an edge with a weight equal to the coupling coefficient is established between the corresponding nodes.

[0017] Furthermore, during the training process of the covariance attention module, a graph regularization term is constructed using the prior knowledge graph to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge, including: The Laplacian matrix of the prior knowledge graph is embedded into the loss function of the covariance attention module; The regularization term is in the form of ,in λ is the normalized Laplacian matrix, W is the attention weight matrix to be learned with the same number of rows as the dimension of the Laplacian matrix, and λ is the regularization hyperparameter used to ensure that the features corresponding to the connected nodes in the graph receive similar attention weights.

[0018] Further, the parameterization of a Dirichlet distribution using the deep fusion features includes: The deep fusion features are input into a fully connected neural network layer, and the number of output nodes of the network layer is equal to the preset number N of insulation state categories; After applying a non-negative activation function to the output of the network layer, a bias greater than zero is added to ensure that all outputs are greater than zero, resulting in N concentration parameters. This constitutes the parameter vector α of the Dirichlet distribution, which is composed of the concentration parameter. to composition.

[0019] Further, if the uncertainty metric is higher than a threshold, the importance of each multi-resolution component is evaluated based on the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results, including: The threshold is the statistical value of uncertainty measurement over the most recent monitoring periods; When the uncertainty measure of the current period exceeds the threshold, the structured attention weight matrix is ​​summed by rows or columns to obtain the cumulative attention weight of each multi-resolution component. Components with cumulative weights higher than preset quantile values ​​are selected, and the time-frequency region features corresponding to the components are analyzed. Based on the features, one or more decomposition parameters of the subsequent wavelet packet decomposition are adjusted. The decomposition parameters include the number of decomposition layers or node selection strategies.

[0020] Further, the calculation of the expected value of the Dirichlet distribution, as the multi-class probability output of the equipment insulation state, includes: Using concentration parameters to The parameter vector α is used to calculate the total concentration. The total concentration For all concentration parameters The sum; For the k-th insulation state category, the output probability is obtained by passing the concentration parameter of the category. Divide by total concentration Calculations were performed to obtain a set of normalized probability values ​​representing the possibilities of different insulation states.

[0021] This invention suppresses noise and outlier data interference by performing multi-scale decomposition of monitoring data and constructing a covariance manifold representation, accurately identifying deep nonlinear correlations between features and improving the stability and reliability of the model. By utilizing a prior knowledge graph of the equipment structure and electromagnetic coupling relationship, the learning process of attention weights is structurally constrained, making model decisions more interpretable and avoiding spurious correlations that may arise from relying solely on data-driven approaches, thus enhancing the model's generalization ability. A closed-loop adjustment mechanism for uncertainty representation and feature extraction is established, which can optimize the front-end signal processing based on the confidence level of the model output, ensuring the accuracy and reliability of monitoring results under complex operating conditions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the hardware device structure; Figure 2 This is a flowchart of an online monitoring method for the insulation status of power transmission and transformation equipment in new energy power plants. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0024] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0025] Hardware device structure such as Figure 1 As shown, it consists of three main parts: the transformer itself, signal sensors, and a data processing platform cluster. Firstly, this invention proposes an online monitoring method for the insulation status of transmission and transformation equipment in new energy power plants, such as... Figure 2 As shown, it includes the following steps: S1. Obtain the operation monitoring time series data of the power transmission and transformation equipment of the new energy power station, perform wavelet packet decomposition on the time series data to generate a set of multi-resolution components; input the multi-resolution components as features to the covariance attention module, calculate the covariance matrix between features, and map the covariance matrix to the tangent space of the Riemann manifold through the log-Euclidean metric to generate the covariance manifold representation of the features. Multi-source time-series signals, such as partial discharge and equipment temperature, are collected using sensors including high-frequency current transformers, ultra-high-frequency sensors, and infrared thermal imagers. These signals are then aggregated and stored as CSV or Parquet files using a SCADA system. The raw data is normalized using z-score standardization. For each normalized time-series data stream, the WaveletPacket function is called with an initial maximum decomposition level (e.g., 5 levels) and a mother wavelet function (e.g., db4) to perform full wavelet packet decomposition. For each node in the decomposition tree, the Shannon entropy algorithm is used to calculate the information entropy as an energy cost. The decomposition tree is traversed from bottom to top. If the information entropy of a parent node is less than or equal to the sum of the information entropies of its two child nodes, node pruning is performed to obtain the optimal wavelet packet decomposition tree. The reconstructed signals corresponding to each leaf node of the tree are the multi-resolution components. All multi-resolution components are combined into a feature matrix. The MinCovDet function from the Scikit-learn library is used to fit this feature matrix, resulting in a covariance matrix that is resistant to the influence of outliers. To process this symmetric positive definite covariance matrix, the linalg.eigh function from the SciPy library is used to perform eigenvalue decomposition on the matrix, obtaining the eigenvalue diagonal matrix D and the eigenvector matrix U. The log function from the NumPy library is then used to take the natural logarithm of all diagonal elements (eigenvalues) in D to obtain log(D), and finally, matrix multiplication is performed. The logarithm of the covariance matrix is ​​calculated, and this result is the vectorized representation in the tangent space of the Riemannian manifold. The Log-Euclidean metric is a Riemannian metric defined on a symmetric positive definite matrix manifold. It establishes an isomorphism between the SPD matrix space and the Euclidean space through the matrix logarithm, thereby simplifying geometric operations.

[0026] In one optional embodiment, the step of performing wavelet packet decomposition on the time-series data to generate a set of multi-resolution components includes: The original signal is completely decomposed to the preset maximum number of layers by pre-setting wavelet basis functions. Then, the tree is traversed from bottom to top. The information entropy criterion of node energy or Shannon entropy is used to determine whether to prune, thereby constructing the optimal wavelet packet decomposition tree. The multi-resolution components corresponding to the leaf nodes of the tree are used as output.

[0027] Choose wavelet basis functions suitable for transient signal analysis, preferably from the Daubechies series (db4 or db8) or the Symlets series (sym8) with good time-frequency localization characteristics. Set the maximum number of decomposition levels. The decomposition layer values ​​are typically based on the signal length N and the sampling frequency. Determined, an empirical choice is For example, for a signal with a length of 4096, it can be set The value is 10. First, the original signal is completely decomposed into the 10th order. The first layer yields a complete wavelet packet binary tree; then from the second layer... The pruning decision is made layer by layer from bottom to top: For each node (j,k) in the j-th layer, if its entropy E(j,k) is less than or equal to the sum of the entropies of its two child nodes (j+1,2k) and (j+1,2k+1), i.e., E(j,k)≤E(j+1,2k)+E(j+1,2k+1), then the parent node is retained as a leaf node and its child nodes are pruned; otherwise, the child nodes are retained, and the parent node is no longer a leaf node. In each layer j, j starts from... Decrementing from -1 down to 0, calculate the Shannon entropy for each node (j,k). The specific steps are: first, calculate the total energy of the wavelet packet coefficients of that node. Then calculate the normalized probability. Finally, Shannon entropy was obtained. The bottom-up traversal process is performed recursively until the root node is reached. The set of all unpruned leaf nodes constitutes the optimal wavelet packet decomposition tree, and the signal components corresponding to the decomposition tree, i.e., the multi-resolution components, serve as feature inputs for subsequent processing.

[0028] In one possible embodiment, the step of inputting the multi-resolution components as features into the covariance attention module and calculating the covariance matrix between features includes: A subset containing a specific proportion of samples is randomly initialized from all observed samples; The covariance matrix and Mahalanobis distance between the multi-resolution component features are calculated based on the subset, and the samples are filtered according to the Mahalanobis distance to update the subset; The update process is iteratively executed until the subset is stable, and the covariance matrix calculated from the stable subset is used as the result covariance matrix.

[0029] The minimum covariance determinant (MCD) algorithm is employed. There are n observation samples, each containing p multi-resolution component features. The subset size h is set, preferably within the range [n / 2, 3n / 4]. A commonly used setting is... For example, for n=1000 samples and p=8 features, then h=504. H samples are randomly selected from the n samples to form an initial subset H0. Then, the iteration proceeds to step C: in the k-th iteration, based on the current subset... Calculate the sample mean Covariance Matrix Using the mean and covariance, calculate the Mahalanobis distance for all n samples. Sort all samples in ascending order of Mahalanobis distance, and select the h samples with the smallest distance to form a new subset. The stopping condition for iteration is that the subset no longer changes or the determinant of the covariance matrix converges, i.e. To prevent infinite loops, a maximum number of iterations is set, such as 100. After the iterations are complete, a stable subset is used as the basis for the iterations. The calculated covariance matrix , which is the covariance matrix we are looking for.

[0030] S2, Construct a prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components; During the training process of the covariance attention module, use the prior knowledge graph to construct a graph regularization term, apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge; Based on equipment planning drawings and electromagnetic simulation analysis results, key components such as transformer windings, bushings, iron cores, and circuit breaker contacts are identified. Each multi-resolution component is mapped to the component corresponding to its monitoring channel as a graph node. If the components corresponding to two multi-resolution components are physically connected or the spatial distance is less than a preset threshold, a physical adjacency edge is established between these two component nodes. If the transient process of one component will induce a voltage exceeding a threshold on another component through an electromagnetic field, an electromagnetic coupling edge is established between the corresponding component nodes. The graph structure is constructed using the NetworkX library, and the normalized graph Laplacian matrix L is calculated. In the self-attention mechanism training of the covariance attention module, the covariance manifold representation is used as input. The query matrix Q, key matrix K, and value matrix V are generated through a fully connected layer. The attention weight matrix W is calculated using the formula... Received, among which Let W be the dimension of the query / key vector, where W is a V×V square matrix and V is the number of multi-resolution components. Construct a graph regularization term, mathematically represented as weight hyperparameters. Multiply by the transpose of W, multiply by the graph Laplacian matrix L, and then multiply by the trace of W, i.e. During the backpropagation process of model training, this graph regularization term is added to the main classification loss function, such as the cross-entropy loss function, to form a new total loss function. The Adam optimizer is used to minimize the total loss function and iteratively update the network parameters, so that the attention weights corresponding to device components with adjacency or coupling relationships become more similar after training.

[0031] In one possible embodiment, the construction of a priori knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components includes: Each multi-resolution component is abstracted as a node in the graph; If the components monitored by the monitoring channels of two multi-resolution components have a physical adjacency relationship in the equipment structure, then an edge with a first preset weight is established between the corresponding nodes. If the electromagnetic coupling coefficient between the components monitored by the monitoring channels of two multi-resolution components is greater than the preset coupling threshold calculated by finite element simulation, then an edge with a weight equal to the coupling coefficient is established between the corresponding nodes.

[0032] The number of nodes V in the graph is equal to the number of multi-resolution components output by wavelet packet decomposition. Assuming there are 8 components, the graph contains 8 nodes. Based on the equipment's planning drawings or 3D model, identify direct physical contact or adjacent relationships between components. For example, in a transformer, the high-voltage winding and low-voltage winding, and the winding and core are physically adjacent. If the monitoring channels corresponding to nodes i and j monitor the high-voltage winding and low-voltage winding respectively, then an edge is established between them. First preset weight. This can be set to a fixed value, such as 1.0, indicating strong correlation. An electromagnetic field model of the power transmission and transformation equipment is established using finite element analysis software such as ANSYS Maxwell or COMSOL. The electromagnetic coupling coefficient matrix between different monitoring points is obtained through simulation calculations. A coupling threshold is then set. The preferred range is [0.05, 0.2], for example. =0.1. If the coupling coefficient between the monitoring points corresponding to node i and node j is... Greater than Then an edge is established between them, with weight... That is, the coupling coefficient. If two nodes simultaneously satisfy both physical adjacency and electromagnetic coupling conditions, then the edge weight is the maximum of the two. The generated graph is represented by a V×V adjacency matrix A, where .

[0033] In one possible embodiment, during the training process of the covariance attention module, the construction of a graph regularization term using the prior knowledge graph applies a structured constraint to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge, including: The Laplacian matrix of the prior knowledge graph is embedded into the loss function of the covariance attention module; The regularization term is in the form of ,in λ is the normalized Laplacian matrix, W is the attention weight matrix to be learned with the same number of rows as the dimension of the Laplacian matrix, and λ is the regularization hyperparameter used to ensure that the features corresponding to the connected nodes in the graph receive similar attention weights.

[0034] The graph Laplacian matrix L is calculated based on the adjacency matrix A obtained in the previous embodiment. A normalized Laplacian matrix is ​​typically used. Where D is the degree matrix and the diagonal matrix. I is the identity matrix. This form can better handle nodes with uneven degrees. The attention weight matrix W to be learned is a V×V matrix, where V is the number of features, i.e., multi-resolution components. The i-th row of W... Let represent the attention weight vector for the i-th feature. The graph regularization term is defined as... The mathematical essence of this item is... It penalizes the difference in weight vectors between node pairs (i,j) connected by edges; the closer the connection, the more severe the penalty. The larger the value, the stronger the penalty. Regularization hyperparameter. The strength of the graph constraint is controlled by the parameter value, which is determined on the validation set through grid search or Bayesian optimization; for example, it can be initially set to 0.01. This regularization term is added to the model's main loss function, such as cross-entropy loss. In this process, the total loss function is formed. Optimization is performed during model training.

[0035] S3, apply the structured attention weight matrix to the multi-resolution components for weighted fusion to obtain deep fusion features; parameterize a Dirichlet distribution using the deep fusion features, and calculate the model uncertainty measure based on the Dirichlet distribution; if the uncertainty measure is higher than a threshold, evaluate the importance of each multi-resolution component according to the structured attention weight matrix, adjust the decomposition parameters of the wavelet packet decomposition based on the importance evaluation result, and then return to re-execute the wavelet packet decomposition step; otherwise, calculate the expected value of the Dirichlet distribution as the multi-class probability output of the device insulation state.

[0036] The weights in the structured attention weight matrix are element-wise multiplied with their corresponding multi-resolution components and then summed to achieve weighted fusion, generating a single deep fusion feature vector. This deep fusion feature vector is then input into a fully connected layer with a Softplus activation function, outputting N positive values, which serve as the concentration parameters for an N-class Dirichlet distribution. ; Calculate the total concentration For all The sum, according to the formula Calculate the model uncertainty metric U for the current sample, where N is the number of insulation state categories. Set an uncertainty threshold, which is twice the moving average of the uncertainty metric U over the most recent M time steps plus the moving standard deviation. If the currently calculated U is higher than this threshold, extract the sum of each row or column of the structured attention weight matrix as the importance score for each multi-resolution component. Analyze whether the high importance scores are distributed in high-frequency or low-frequency components. If the proportion of high-frequency components exceeds a set ratio, such as 60%, increase the decomposition level of the wavelet packet decomposition by 1 to enhance the resolution of high-frequency details, and then return to the decomposition step using the new decomposition level. If U is lower than or equal to the threshold, calculate the state probability of each class according to the formula, i.e., using the concentration parameter of that class. Divide by total concentration This yields an N-dimensional probability vector, which is output as a multi-classification result, such as normal, minor defects, and severe defects.

[0037] The Dirichlet distribution is a continuous multivariate probability distribution defined on the simplex. It is the conjugate prior of the multivariate distribution and is used to represent the uncertainty of a set of probability vectors.

[0038] In one possible embodiment, parameterizing a Dirichlet distribution using the deep fusion features includes: The deep fusion features are input into a fully connected neural network layer, and the number of output nodes of the network layer is equal to the preset number N of insulation state categories; After applying a non-negative activation function to the output of the network layer, a bias greater than zero is added to ensure that all outputs are greater than zero, resulting in N concentration parameters. This constitutes the parameter vector α of the Dirichlet distribution, which is composed of the concentration parameter. to composition.

[0039] Assuming the deep fusion feature is a dimensional feature... The vector z is fed into a fully connected layer, also known as the evidence layer, whose weight matrix has dimensions of . The bias vector has dimension N, where N is the total number of insulation state categories. For example, N=4 corresponds to normal, minor defect, severe defect, and emergency alarm. The output of this layer is an N-dimensional vector. This is called the evidence vector. To ensure the concentration parameters of the Dirichlet distribution... Strictly positive for each element of the evidence vector e Apply the Softplus activation function, in the form of: Because it is smooth and always positive. Add a bias of 1 to each activated value, i.e. The increment operation ensures that even if the evidence output by the neural network is zero, the concentration parameter is at least 1, thus avoiding the problems of zero probability and numerical instability. The resulting N-dimensional vector... This is the Dirichlet distribution. The parameters are given by p, where p is the probability vector representing each category.

[0040] In one possible embodiment, if the uncertainty metric is higher than a threshold, the importance of each multi-resolution component is evaluated according to the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results, including: The threshold is the statistical value of uncertainty measurement over the most recent monitoring periods; When the uncertainty measure of the current period exceeds the threshold, the structured attention weight matrix is ​​summed by rows or columns to obtain the cumulative attention weight of each multi-resolution component. Components with cumulative weights higher than preset quantile values ​​are selected, and the time-frequency region features corresponding to the components are analyzed. Based on the features, one or more decomposition parameters of the subsequent wavelet packet decomposition are adjusted. The decomposition parameters include the number of decomposition layers or node selection strategies.

[0041] The uncertainty measure U of the model is obtained through the total concentration. Calculations show that , where N is the number of categories. Maintain a sliding window containing uncertainty values ​​from the most recent M monitoring periods, for example, M=200. Set the threshold... Set as the mean of the uncertainty values ​​within this window. Add k times the standard deviation ,Right now Where k is typically 2 or 3, corresponding to a 95% or 99.7% confidence interval. When the uncertainty of the current monitoring period... When this occurs, feedback adjustment is triggered. At this point, the importance of the V×V structured attention weight matrix W is evaluated by calculating the sum of each row. To obtain the total output attention for the i-th multi-resolution component, calculate the importance scores for all components. The p-th quantile value is determined, where p is preferably 90 or 95. All components with scores higher than this quantile value are identified as key components. The positions of these key components in the wavelet packet decomposition tree are analyzed to determine the time-frequency regions they cover. If the key components are mainly concentrated in a certain high-frequency region, it indicates that the current decomposition has insufficient resolution for that region; therefore, the maximum decomposition level for the next wavelet packet decomposition is increased. Increase by 1, for example, from 5 to 6. If the critical components are widely distributed, consider adjusting the node selection strategy, such as assigning higher weights to the node entropy values ​​in the frequency band where the critical components are located in the Shannon entropy criterion, to guide the decomposition tree to deepen into the aforementioned regions.

[0042] In one possible embodiment, calculating the expected value of the Dirichlet distribution as a multi-class probability output of the device insulation state includes: Using concentration parameters to The parameter vector α is used to calculate the total concentration. The total concentration For all concentration parameters The sum; For the k-th insulation state category, the output probability is obtained by passing the concentration parameter of the category. Divide by total concentration Calculations were performed to obtain a set of normalized probability values ​​representing the possibilities of different insulation states.

[0043] In obtaining the parameter vector of the Dirichlet distribution The next step is to calculate the total concentration. Also known as the Dirichlet intensity. The calculation formula is: The value represents the total amount of evidence the model has collected for the current prediction. The larger the value, the higher the confidence level of the model. Calculate the predicted probability for each category k. According to the properties of the Dirichlet distribution, the expected value, i.e., the most likely probability distribution, is: This calculation process is performed once for all N categories, resulting in an N-dimensional probability vector. The sum of the elements in this vector is 1, representing the probability that the device is in various insulation states. For example, for N=4 categories, if the concentration parameter output by the model is... =[50.5,3.0,1.5,1.0], then the total concentration =56.0. The corresponding probability output is P≈[0.902,0.054,0.027,0.017]. The probability vector is the multi-class probability output of the equipment insulation status.

[0044] In one possible embodiment, a diagnostic model based on graph regularized attention and evidence deep learning is employed, which includes a graph regularized covariance attention module and an evidence network.

[0045] The diagnostic model receives multi-resolution components as input and outputs an evidence vector for parameterizing the Dirichlet distribution.

[0046] The core of the graph-regularized covariance attention module is a structured attention weight matrix W to be learned. The dimension of this matrix is ​​the same as the number of multi-resolution components, V (V×V). This module takes a feature set consisting of V multi-resolution components as input and performs weighted fusion through an attention mechanism to generate deep fused features. During training, the loss function of this module includes a graph regularization term. ,in For the normalized Laplace matrix of the prior knowledge graph, This is a hyperparameter. This regularization term uses prior knowledge to guide the learning of attention weights, so that features that are strongly correlated in physics or electromagnetism can receive similar attention.

[0047] The evidence network is a fully connected neural network layer used to map deep fusion features to the evidence space. The input to this network is the deep fusion feature vector z output by the graph regularization covariance attention module. The number of output nodes of this fully connected layer is the preset number of insulation state categories N. The output result e is processed by an activation function to ensure non-negativity and then a bias greater than zero is added to obtain N concentration parameters. The specific calculation is as follows: ,in It is the output of the fully connected layer for the k-th class, and Softplus is the activation function, in the form of: .Depend on to The vector formed This is the parameter vector of the Dirichlet distribution, which serves as the output of the entire diagnostic model.

[0048] The effectiveness of each core module in the ablation protocol was verified by setting up a series of ablation experiments.

[0049] The experimental conditions were as follows: the dataset contained 5000 transient signal samples collected from a 500 kV substation, covering four insulation states: normal, minor, severe, and emergency, and was divided into training and test sets in a 7:3 ratio; the hardware environment consisted of an NVIDIA RTX 3080 GPU and 32GB of RAM. The baseline model used standard three-layer discrete wavelet transform (DWT) to extract features and employed a deep neural network based on an attention mechanism combined with Softmax for classification. The baseline model achieved an average accuracy of 93.5% and an F1 score of 0.932 on the test set. The complete model proposed in this invention, combining wavelet packet decomposition (WPD), covariance modulo decomposition (MCD), graph regularized attention, and an evidence network, achieved an average accuracy of 98.2% and an F1 score of 0.981. Compared to traditional methods, feature extraction is more targeted, and the attention mechanism guided by statistics and prior knowledge can more accurately detect deep dependencies related to insulation states, thus achieving more reliable diagnosis.

[0050] To analyze the specific contributions of graph regularization and covariance matrix, module ablation was performed based on the complete model.

[0051] The experimental conditions remained constant. Two variant models were set up for comparison: Model A was the complete model, Model B had the graph regularization term removed (i.e., the regularization hyperparameter was set to zero), and Model C replaced the covariance MCD with the standard sample covariance (SCM). The experimental data recorded the diagnostic accuracy of each model on the same test set: Model A was 98.2%, Model B was 96.5%, and Model C was 97.1%. The results showed that removing the graph regularization term resulted in a 1.7 percentage point decrease in accuracy, making it the most affected module; removing the covariance matrix resulted in a 1.1 percentage point decrease in accuracy. The graph regularization term, by utilizing the prior physical and electromagnetic correlations between device components, constrained the learning space of attention weights, focusing the learning space on physically meaningful feature combinations and avoiding spurious correlations; the covariance, through the MCD algorithm, successfully identified and reduced the weights of abnormal samples or noise in the data, enhancing the stability of feature representation and the model's generalization ability.

[0052] To verify the advantages of Evidence-Based Deep Learning (EDL) and the uncertainty feedback mechanism, the model of this invention was compared with a similarly structured model using a standard Softmax output layer. Experimental conditions were identical except for the evaluation metrics. Evaluation metrics included the model's accuracy on 200 specially selected fuzzy boundary samples and the Expected Calibration Error (ECE), representing the confidence level of the calibration; a lower ECE was better. Specific experimental data are as follows: the model using EDL achieved an accuracy of 91.0% on fuzzy samples with an ECE of 0.015; while the model using Softmax achieved an accuracy of 87.5% with an ECE of 0.048. Furthermore, in a simulated continuous diagnostic task spanning 500 monitoring cycles, the complete model with the uncertainty feedback adjustment mechanism maintained a stable accuracy of 97.5%, while the accuracy of the EDL model without the feedback mechanism slowly decreased to 95.8%. EDL can more accurately represent uncertainty, providing low-confidence judgments for fuzzy or unknown samples, thereby improving the reliability of decision-making; the feedback mechanism based on this uncertainty metric can adjust the feature extraction strategy to adapt to the slow changes in signal characteristics, demonstrating the model's capabilities in long-term monitoring applications.

[0053] Secondly, the present invention also provides an online monitoring system for the insulation status of power transmission and transformation equipment in new energy power plants, comprising the following modules: The generation module is used to acquire the operation monitoring time series data of the power transmission and transformation equipment of the new energy power station, perform wavelet packet decomposition on the time series data to generate a set of multi-resolution components; input the multi-resolution components as features to the covariance attention module, calculate the covariance matrix between features, and map the covariance matrix to the tangent space of the Riemann manifold through the log-Euclidean metric to generate the covariance manifold representation of the features. A construction module is used to construct a prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components. During the training process of the covariance attention module, the prior knowledge graph is used to construct a graph regularization term to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge. The calculation module is used to apply the structured attention weight matrix to the multi-resolution components for weighted fusion to obtain deep fusion features; parameterize a Dirichlet distribution using the deep fusion features, and calculate the model uncertainty measure based on the Dirichlet distribution; if the uncertainty measure is higher than a threshold, evaluate the importance of each multi-resolution component according to the structured attention weight matrix, adjust the decomposition parameters of the wavelet packet decomposition based on the importance evaluation result, and then return to re-execute the wavelet packet decomposition step; otherwise, calculate the expected value of the Dirichlet distribution as the multi-class probability output of the device insulation state.

[0054] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A method for online monitoring of the insulation status of power transmission and transformation equipment in new energy power plants, characterized in that, Includes the following steps: Acquire the operation monitoring time series data of the power transmission and transformation equipment of the new energy power station, perform wavelet packet decomposition on the time series data, and generate a set of multi-resolution components; The multi-resolution components are input as features to the covariance attention module to calculate the covariance matrix between features. The covariance matrix is ​​then mapped to the tangent space of the Riemannian manifold using a log-Euclidean metric to generate the covariance manifold representation of the features. A prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components is constructed. During the training process of the covariance attention module, a graph regularization term is constructed using the prior knowledge graph to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge. The structured attention weight matrix is ​​applied to the multi-resolution components for weighted fusion to obtain deep fusion features; a Dirichlet distribution is parameterized using the deep fusion features, and a model uncertainty measure is calculated based on the Dirichlet distribution; If the uncertainty metric is higher than the threshold, the importance of each multi-resolution component is evaluated according to the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results. Then, the wavelet packet decomposition step is re-executed. Otherwise, the expected value of the Dirichlet distribution is calculated as the multi-class probability output of the equipment insulation state.

2. The method according to claim 1, characterized in that, The wavelet packet decomposition of the time-series data generates a set of multi-resolution components, including: The original signal is completely decomposed to the preset maximum number of layers by pre-setting wavelet basis functions. Then, the tree is traversed from bottom to top. The information entropy criterion of node energy or Shannon entropy is used to determine whether to prune, thereby constructing the optimal wavelet packet decomposition tree. The multi-resolution components corresponding to the leaf nodes of the tree are used as output.

3. The method according to claim 1, characterized in that, The step of inputting the multi-resolution components as features into the covariance attention module and calculating the covariance matrix between features includes: A subset containing a specific proportion of samples is randomly initialized from all observed samples; The covariance matrix and Mahalanobis distance between the multi-resolution component features are calculated based on the subset, and the samples are filtered according to the Mahalanobis distance to update the subset; The update process is iteratively executed until the subset is stable, and the covariance matrix calculated from the stable subset is used as the result covariance matrix.

4. The method according to claim 1, characterized in that, The construction of the prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components includes: Each multi-resolution component is abstracted as a node in the graph; If the components monitored by the monitoring channels of two multi-resolution components have a physical adjacency relationship in the equipment structure, then an edge with a first preset weight is established between the corresponding nodes. If the electromagnetic coupling coefficient between the components monitored by the monitoring channels of two multi-resolution components is greater than the preset coupling threshold calculated by finite element simulation, then an edge with a weight equal to the coupling coefficient is established between the corresponding nodes.

5. The method according to claim 3, characterized in that, During the training process of the covariance attention module, a graph regularization term is constructed using the prior knowledge graph to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge, including: The Laplacian matrix of the prior knowledge graph is embedded into the loss function of the covariance attention module; The regularization term is in the form of ,in λ is the normalized Laplacian matrix, W is the attention weight matrix to be learned with the same number of rows as the dimension of the Laplacian matrix, and λ is the regularization hyperparameter used to ensure that the features corresponding to the connected nodes in the graph receive similar attention weights.

6. The method according to claim 1, characterized in that, The process of parameterizing a Dirichlet distribution using the deep fusion features includes: The deep fusion features are input into a fully connected neural network layer, and the number of output nodes of the network layer is equal to the preset number N of insulation state categories; After applying a non-negative activation function to the output of the network layer, a bias greater than zero is added to ensure that all outputs are greater than zero, resulting in N concentration parameters. This constitutes the parameter vector α of the Dirichlet distribution, which is composed of the concentration parameter. to composition.

7. The method according to claim 1, characterized in that, If the uncertainty metric is higher than a threshold, the importance of each multi-resolution component is evaluated based on the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results, including: The threshold is the statistical value of uncertainty measurement over the most recent monitoring periods; When the uncertainty measure of the current period exceeds the threshold, the structured attention weight matrix is ​​summed by rows or columns to obtain the cumulative attention weight of each multi-resolution component. Components with cumulative weights higher than preset quantile values ​​are selected, and the time-frequency region features corresponding to the components are analyzed. Based on the features, one or more decomposition parameters of the subsequent wavelet packet decomposition are adjusted. The decomposition parameters include the number of decomposition layers or node selection strategies.

8. The method according to claim 1, characterized in that, The calculation of the expected value of the Dirichlet distribution, as the multi-class probability output of the equipment insulation state, includes: Using concentration parameters to The parameter vector α is used to calculate the total concentration. The total concentration For all concentration parameters The sum; For the k-th insulation state category, the output probability is obtained by passing the concentration parameter of the category. Divide by total concentration Calculations were performed to obtain a set of normalized probability values ​​representing the possibilities of different insulation states.

9. An online monitoring system for the insulation status of power transmission and transformation equipment in new energy power plants, characterized in that, Includes the following modules: The generation module is used to acquire the operation monitoring time series data of the power transmission and transformation equipment of the new energy power station, perform wavelet packet decomposition on the time series data, and generate a set of multi-resolution components. The multi-resolution components are input as features to the covariance attention module to calculate the covariance matrix between features. The covariance matrix is ​​then mapped to the tangent space of the Riemannian manifold using a log-Euclidean metric to generate the covariance manifold representation of the features. A construction module is used to construct a prior knowledge graph based on the physical adjacency and electromagnetic coupling relationships of device components. During the training process of the covariance attention module, the prior knowledge graph is used to construct a graph regularization term to apply structured constraints to the process of learning attention weights based on the covariance manifold representation, thereby obtaining a structured attention weight matrix guided by graph knowledge. The calculation module is used to apply the structured attention weight matrix to the multi-resolution components for weighted fusion to obtain deep fusion features; A Dirichlet distribution is parameterized using the deep fusion features, and a model uncertainty measure is calculated based on the Dirichlet distribution; If the uncertainty metric is higher than the threshold, the importance of each multi-resolution component is evaluated according to the structured attention weight matrix, and the decomposition parameters of the wavelet packet decomposition are adjusted based on the importance evaluation results. Then, the wavelet packet decomposition step is re-executed. Otherwise, the expected value of the Dirichlet distribution is calculated as the multi-class probability output of the equipment insulation state.

10. The system according to claim 9, characterized in that, The wavelet packet decomposition of the time-series data generates a set of multi-resolution components, including: The original signal is completely decomposed to the preset maximum number of layers by pre-setting wavelet basis functions. Then, the tree is traversed from bottom to top. The information entropy criterion of node energy or Shannon entropy is used to determine whether to prune, thereby constructing the optimal wavelet packet decomposition tree. The multi-resolution components corresponding to the leaf nodes of the tree are used as output.

Images