Nonlinear feature compression, preliminary screening and hierarchical evaluation off-line diagnosis method for sampling channel of transformer substation voltage transformer

Abstract

This invention discloses an offline diagnostic method for nonlinear feature compression, initial screening, and hierarchical evaluation of sampling channels of substation voltage transformers. Addressing the offline risk assessment needs of substation voltage transformer sampling channels, this method first constructs a unified output health discriminant factor by combining kernel principal component analysis and entropy weighting, reducing redundancy and enhancing cross-site robustness. An extreme learning machine is introduced to perform offline initial screening of sample vectors, focusing on abnormal sampling channel segments that meet the anomaly judgment criteria to improve batch processing efficiency. Based on fuzzy membership functions and a monotonic mapping relationship between predefined risk levels and scores, multi-level risk classification of initial screening candidate samples is achieved. Furthermore, trend quantities are combined to correct risk evolution for risk decision-making, ensuring that the score remains stable near the boundary zone and is more sensitive to abrupt changes. This method effectively improves the offline data fault diagnosis and location capabilities of voltage measurement links under conditions of multiple disturbances and non-stationary operation.

Description

Technical Field

[0001] This invention relates to the technical field of voltage measurement link fault diagnosis, and in particular to an offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of voltage transformers in substations. Background Technology

[0002] A substation voltage measurement link typically consists of a voltage transformer or voltage sensor, a sampling and synchronization unit, a merging unit for sampling and message transmission, and a protection and control device for receiving and calculating data. With the increasing adoption of digital substations, electronic voltage transformers and their sampling channels are widely used in process-level sample value transmission and relay protection, measurement, and metering operations. Over long-term operation, these channels may experience issues such as ratio and phase shift deviations, zero drift and temperature drift, dynamic response degradation, time synchronization errors and jitter, frame drops, and time delay fluctuations, leading to increased phasor calculation errors or a higher risk of protection maloperation or failure to operate. Against this backdrop, substation voltage transformers, as critical measurement and monitoring devices, directly affect the accuracy of power quality assessment, reactive power control, and protection actions, making them a core component for ensuring system safety and sustainable operation.

[0003] Existing research on fault diagnosis of voltage measurement links can be broadly categorized into three types: mechanism-oriented, data-driven, and collaborative optimization. Mechanism-oriented methods typically construct residuals or observations based on voltage transformers or their equivalent measurement models, and combine statistical tests or observers to achieve anomaly identification. This type of method has a certain degree of interpretability, but it is quite sensitive to model accuracy, parameter calibration, and consistency of operating conditions. In substation electronic voltage transformer sampling channels, due to equipment differences, noise interference, and discrete offline data windows, the residual statistical characteristics are prone to drift, and misjudgment or missed detection may occur after threshold shifting. Data-driven methods are based on historical sampling values ​​or waveform data, and use machine learning or deep learning to directly learn anomaly representations, which can improve the sample recognition capability under complex patterns. However, in offline batch processing scenarios, they are limited by the influence of constant sample imbalance, uneven operating condition coverage, and cross-batch distribution changes, resulting in insufficient generalization and reproducibility. Moreover, the output often remains at the level of anomaly discrimination, making it difficult to form stable risk classification and interpretable localization. Collaborative optimization methods further integrate mechanistic residuals and data-driven discrimination to achieve integrated detection, isolation, and estimation, or enhanced robustness. However, the algorithm structure and tuning complexity are high, and when the joint characterization of link factors such as time quality and frame loss in offline data is insufficient, it is still difficult to output cross-batch comparable health scores and consistent degradation trend conclusions. Regarding the aforementioned related technologies, the inventors believe that they suffer from insufficient capability for offline data fault diagnosis and location of voltage measurement links under conditions of multiple disturbances and non-stationary operation. Summary of the Invention

[0004] To address the shortcomings of existing technologies in offline data fault diagnosis and location capabilities for voltage measurement links under conditions of multiple disturbances and non-stationary operation, this application provides an offline diagnostic method for nonlinear feature compression, initial screening, and graded evaluation of sampling channels for substation voltage transformers. This method achieves a balance between accuracy, interpretability, and offline processing efficiency, alleviating problems such as difficulty in detecting early-stage gradual degradation, high threshold sensitivity, performance fluctuations caused by sample imbalance and cross-batch distribution variations, and providing a reliable basis for risk classification, priority ranking, and maintenance decisions for sampling channels.

[0005] Firstly, the above-mentioned inventive objective of this application is achieved through the following technical solution: An offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of voltage transformers in substations, the method comprising: Acquire multi-source offline data from the sampling channels of voltage transformers in substations, extract multi-source features from the multi-source offline data to form sample vectors, preprocess the sample vectors, and generate a unified input health discrimination factor; The health discrimination factors are trained using an extreme learning machine. Based on the training results, the sampling channels corresponding to health discrimination factors that exceed the preset threshold are batch-screened and abnormal sampling channels are marked to obtain preliminary candidate samples. Each of the initial screening candidate samples is subjected to fuzzy weighted scoring, and the score value is matched with a predefined risk level. The initial screening candidate samples are then classified and quantified based on the matching results. Based on the risk level quantification results of each preliminary screening candidate sample, the abnormal sampling channels corresponding to the preliminary screening candidate samples are ranked by channel risk, and maintenance decision processing is carried out based on the ranking results.

[0006] In a preferred embodiment, this application can be further configured as follows: The acquisition of multi-source offline data from the sampling channels of substation voltage transformers, extraction of multi-source features from the multi-source offline data to form sample vectors, and preprocessing of the sample vectors to generate a unified input health discrimination factor specifically includes: The offline exported data of the voltage transformer sampling channel in the substation is obtained. Multi-source features of the offline exported data within the same sampling time window are extracted according to the sampling time to form a sample vector, and the sample vector is preprocessed. Construct a kernel matrix of the preprocessed sample vectors, and perform eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector in the principal component direction; Calculate the information entropy of the sample vector and the difference degree of the information entropy. Normalize the corresponding sample vector according to the difference degree to obtain the normalized weight value corresponding to the sample vector. The projected representation of each sample vector is weighted with its corresponding normalized weight value to generate a health discriminant factor with the weighted representation as a unified input.

[0007] In a preferred embodiment, this application can be further configured as follows: constructing a kernel matrix of the preprocessed sample vectors, and performing eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector in the principal component direction, specifically including: The projection expression of the sample vector onto the principal component direction is shown below: (1) in, Indicates the first The projection of the i-th sample vector onto the principal component direction, i.e., the i-th sample vector... Principal component vectors after dimensionality reduction of each sample vector Indicates the first The nth sample vector The eigenvector coefficients of the corresponding principal components of each feature. Indicates the first The number of multi-source features per sample vector The Gaussian radial basis function kernel is expressed as follows: (2) in, Representing vectors , The square Euclidean distance between them , Indicates the first The, the A sample vector, This represents a predefined parameter whose value depends on the training dataset. This represents a free parameter.

[0008] In a preferred embodiment, this application can be further configured as follows: calculating the information entropy of the sample vector and calculating the difference degree of the information entropy, and normalizing the corresponding sample vector according to the difference degree to obtain the normalized weight value corresponding to the sample vector, including: The expression for calculating the information entropy of the sample vector is as follows: (3) in, Indicates the first The th sample vector Indicates the first The number of multi-source features in the data. Indicates the normalized i-th The feature in the first The proportion in each sample vector; The difference degree of the information entropy is calculated, and the expression for the difference degree is as follows: (4) in, Indicates the first The th sample vector The degree of difference of each feature; The corresponding sample vectors are normalized based on the difference to obtain the normalized weights of the sample vectors. The calculation expression for the normalized weights is as follows: (5) in, Indicates the first The th sample vector Normalized weights of each feature, This represents the number of principal components retained by the KPCA algorithm.

[0009] In a preferred embodiment, this application can be further configured as follows: the step of weighting the projected representation of each sample vector with its corresponding normalized weight value to generate a health discrimination factor with the weighted representation as a unified input includes: The weighted representation expression between the projected representation of each sample vector and its corresponding normalized weight value is shown below: (6) in, This represents the weighted representation of the principal component projection of the corresponding sample vector, where normalized weight values ​​are applied. This represents the health discriminant factor with a weighted representation as a unified input. They represent the first to the last. The normalized weights of the principal components retained by the KPCA algorithm. Indicates the first The projection representation of each sample vector onto the principal component direction.

[0010] In a preferred embodiment, this application can be further configured as follows: training the health discriminant factors using an extreme learning machine, and batch-filtering sampling channels corresponding to health discriminant factors exceeding a preset threshold based on the training results and marking abnormal sampling channels to obtain preliminary candidate samples, including: The health discriminant factors are converted into a unified matrix representation for extreme learning machine training. The expression for the unified matrix representation is as follows: (6) in, This represents the hidden layer output matrix of the Extreme Learning Machine. This represents the output weight matrix of the extreme learning machine. This represents the expected output of the Extreme Learning Machine; The hidden layer output matrix expression is as follows: (7) in, This represents the activation function of the Extreme Learning Machine. Indicates the first The input weights and biases of the hidden nodes, The input samples of the Extreme Learning Machine are the health discriminant factors. This represents the number of samples input to the Extreme Learning Machine. Indicates the number of neurons in the hidden layer; The expression for the output weight matrix is ​​as follows: (8) in, Represent the Moore–Penrose generalized inverse matrix, when When it is reversible, we have: (9) Based on the training results, the initial screening output of the health discriminant factor is calculated using an extreme learning machine. The expression for the initial screening output is as follows: (10) in, The input to the extreme learning machine represents the first... A sample vector, Indicates the first The weights of each hidden node from the hidden layer to the output layer. Indicates the first a sample vector The initial screening output is processed for anomaly detection and sample labeling based on a preset anomaly detection expression to obtain initial screening candidate samples. The anomaly detection expression is shown below: (11) in, This indicates the preset abnormal threshold.

[0011] In a preferred embodiment, this application can be further configured as follows: training the health discrimination factors using an extreme learning machine, batch-filtering sampling channels corresponding to health factors exceeding a preset threshold based on the training results, and marking abnormal sampling channels to obtain preliminary candidate samples, further includes: The Moore–Penrose generalized inverse matrix is ​​optimized by introducing a regularization parameter. The regularization expression for the Moore–Penrose generalized inverse matrix is ​​as follows: (12) in, This indicates the introduced regularization parameter. The total number of sample vectors in the sample set composed of the offline exported data from the sampling channels.

[0012] In a preferred embodiment, this application can be further configured as follows: performing fuzzy weighted scoring on each of the initial screening candidate samples, matching the score values ​​with predefined risk levels, and performing graded quantification on the initial screening candidate samples based on the matching results, including: The fuzzy membership degree of each of the initial screening candidate samples is calculated, and the functional expression of the fuzzy membership degree is as follows: (13) in, Indicates the first The first preliminary screening candidate sample Diagnostic value corresponding to each feature In the Membership degree of risk level , They represent the first The boundary values ​​of the scoring intervals for each feature dimension at different risk levels; The fuzzy weighted score for each of the initial candidate samples is calculated, and the fuzzy weighted score expression is as follows: (14) in, Indicates the first Fuzzy weighted scores of initial screening candidate samples Indicates the first The first preliminary screening candidate sample Normalized weights of each feature, Indicates a predefined risk level Fixed grade coefficient, These represent predefined risk levels; By using the predefined risk level and fuzzy weighted score mapping relationship, the preliminary screening candidate samples are mapped to the corresponding risk level for graded quantification based on the fuzzy weighted score calculation results.

[0013] In a preferred embodiment, this application can be further configured as follows: performing fuzzy weighted scoring on each of the initial screening candidate samples, matching the score values ​​with predefined risk levels, and performing graded quantification on the initial screening candidate samples based on the matching results, further comprising: The fuzzy weighted score is normalized, and the function expression for fuzzy weighted score normalization is shown below: (15) in, This represents the sum of fuzzy responses across all risk levels.

[0014] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions: An offline diagnostic system for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of voltage transformers in substations, the system comprising: The data preprocessing module is used to acquire multi-source offline data from the sampling channels of voltage transformers in substations, extract multi-source features from the multi-source offline data to form sample vectors, and preprocess the sample vectors to generate a unified input health discrimination factor. The data filtering module is used to train the health discrimination factors using an extreme learning machine, and to batch filter the sampling channels corresponding to health discrimination factors that exceed a preset threshold based on the training results and mark abnormal sampling channels to obtain preliminary candidate samples. The risk grading module is used to perform fuzzy weighted scoring on each of the initial screening candidate samples, match the score value with a predefined risk level, and perform grading and quantification on the initial screening candidate samples based on the matching result. The risk diagnosis module is used to rank the abnormal sampling channels corresponding to each preliminary screening candidate sample according to the risk level quantification result, and to make maintenance decisions based on the ranking result.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs a compact health factor using kernel principal component analysis and entropy weighting, reducing redundancy and enhancing cross-site robustness. A lightweight learning model is then introduced for offline initial screening, focusing on suspected abnormal segments to improve batch processing efficiency. Furthermore, a four-level risk classification is implemented based on fuzzy membership functions, and trend quantities are combined to correct risk evolution, ensuring the score remains stable near the boundary and is more sensitive to sudden changes. Validation on actual sampled channel data demonstrates that this invention maintains high anomaly detection capability under conditions of sample imbalance and background disturbance, and outputs graded risk results suitable for channel ranking and maintenance decisions. Compared to existing methods, the technical advantages of this invention are mainly reflected in: firstly, balancing feature compactness and discriminative power through "kernel manifold compression + objective weighting"; secondly, improving offline diagnostic efficiency and detection quality through a cascade mechanism of "lightweight initial screening - fine classification"; and thirdly, compensating for static scores with interpretable trend quantities, making the comprehensive risk smoother and more reproducible in the threshold neighborhood, and possessing dynamic characterization capabilities for degradation evolution and sudden anomalies. 2. This invention no longer relies on a single threshold or single-domain feature for discrimination. Instead, it targets offline data from the sampling channel of electronic voltage transformers, first performs nonlinear dimensionality reduction using KPCA to reduce redundancy and noise, and then uses the entropy weight method to objectively assign weights to the principal components, forming a compact and highly discriminative health discrimination factor, thereby improving the comparability and stability of cross-batch and cross-operating condition analysis results. 3. Unlike existing methods that directly conduct full-scale detailed evaluation or rely on manual segment-by-segment screening, this invention introduces an ELM initial screening module, which quickly completes training and scoring through closed-loop solution, prioritizes screening out suspected abnormal candidate sets, and then conducts subsequent hierarchical evaluation of candidate samples, realizing the concentrated investment of computing resources and manual review resources, improving offline batch processing efficiency and reducing invalid analysis. 4. Existing methods often output binary alarms or show significant score fluctuations near the grade boundaries, making it difficult to directly connect with maintenance priority ranking. This invention adopts fuzzy membership modeling based on four-level risk semantics and forms a stable static risk score through membership degree normalization. It establishes a monotonic mapping relationship between the score and the grade, making the output near the boundary area smoother and more consistent. At the same time, it provides interpretable and reproducible quantitative basis for channel risk ranking and maintenance decisions. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 This is a flowchart illustrating the implementation of the offline diagnostic method for nonlinear feature compression, initial screening, and graded evaluation in this embodiment.

[0018] Figure 2 This is a flowchart illustrating the implementation of step S10 of the offline diagnostic method for nonlinear feature compression, initial screening, and graded evaluation in this embodiment.

[0019] Figure 3 This is a schematic diagram of the Extreme Learning Machine structure in this embodiment.

[0020] Figure 4 This is a flowchart illustrating the implementation of step S20 of the offline diagnostic method for nonlinear feature compression, initial screening, and graded evaluation in this embodiment.

[0021] Figure 5 This is a flowchart illustrating the implementation of step S30 of the offline diagnostic method for nonlinear feature compression, initial screening, and graded evaluation in this embodiment.

[0022] Figure 6 This is a schematic diagram of the PR curve effect of the simulation verification test in this embodiment.

[0023] Figure 7 This is a schematic diagram of the DET curve effect of the simulation verification test in this embodiment.

[0024] Figure 8 This is a structural block diagram of the offline diagnostic system for nonlinear feature compression, initial screening, and graded evaluation in this embodiment. Detailed Implementation

[0025] 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.

[0026] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] In one embodiment, such as Figure 1 As shown, this application discloses an offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers, specifically including the following steps: S10: Obtain multi-source offline data from the sampling channels of the substation voltage transformers, extract multi-source features from the multi-source offline data to form sample vectors, preprocess the sample vectors, and generate a unified input health discrimination factor.

[0030] Specifically, such as Figure 2 As shown, step S10 includes: S101: Obtain the offline exported data of the voltage transformer sampling channel in the substation, extract the multi-source features of the offline exported data within the same sampling time window according to the sampling time to form a sample vector, and preprocess the sample vector.

[0031] Specifically, offline exported data from the sampling channels of substation voltage transformers is used to form a sample set. ,in, This indicates the number of sample vectors in the sample set. Represents a sample vector. Indicates the first Each sample vector in The observations for each diagnostic quantity are multi-source feature observations. Offline exported data from the same sampling time window for each sampling channel are aggregated according to the sampling time, and the corresponding multi-source features are obtained by extracting data for the corresponding indicators according to preset diagnostic indicators. These preset diagnostic indicators include voltage amplitude / fluctuation, harmonic content or THD, frequency and phase deviation, zero drift and bias, dynamic response delay, as well as offline available data quality information and environmental temperature and humidity.

[0032] Furthermore, considering the non-stationary and non-Gaussian distribution characteristics of offline data across batches and operating conditions, this step performs robust preprocessing on each dimension of features, including but not limited to gentle upper and lower truncation, unit interval mapping, and missing value imputation, thereby avoiding the dominant influence of a small number of outliers on subsequent dimensionality reduction and weighting, and improving the comparability of analysis results from different periods.

[0033] S102: Construct the kernel matrix of the preprocessed sample vectors, and perform eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector in the principal component direction.

[0034] Specifically, in this embodiment, for the potential nonlinear coupling relationship between multi-source features, kernel principal component analysis is used to perform nonlinear dimensionality reduction on the sample set. First, a kernel function is constructed using a Gaussian radial basis kernel function. The kernel matrix of the preprocessed sample vectors is formed, and the expression of the Gaussian radial basis function is shown below: (2) in, Representing vectors , The square Euclidean distance between them , Indicates the first The, the A sample vector, This represents a predefined parameter whose value depends on the training dataset. This represents a free parameter.

[0035] Eigenvalue decomposition of the kernel matrix yields the projection representation of the corresponding sample vectors along the principal component directions, as shown in the following expression: (1) in, Indicates the first The projection of the i-th sample vector onto the principal component direction, i.e., the i-th sample vector... Principal component vectors after dimensionality reduction of each sample vector Indicates the first The nth sample vector The eigenvector coefficients of the corresponding principal components of each feature. Indicates the first The number of multi-source features in a sample vector.

[0036] In this embodiment, the kernel matrix is ​​centered according to a standard to eliminate mean bias.

[0037] S103: Calculate the information entropy of the sample vector and the difference of the information entropy. Normalize the corresponding sample vector according to the difference to obtain the normalized weight value of the sample vector.

[0038] Specifically, the expression for calculating the information entropy of a sample vector is as follows: (3) in, Indicates the first The th sample vector Information entropy of each feature Indicates the first The number of multi-source features in a sample vector Indicates the normalized i-th The feature in the first The proportion in each sample vector.

[0039] The difference degree of information entropy is calculated, and the expression for the difference degree is as follows: (4) in, Indicates the first The th sample vector The degree of difference of each feature.

[0040] The corresponding sample vectors are normalized based on the degree of difference to obtain the normalized weights of the sample vectors. The calculation expression for the normalized weights is as follows: (5) in, Indicates the first The th sample vector Normalized weights of each feature, This represents the number of principal components retained by the KPCA algorithm.

[0041] S104: Weight the projected representation of each sample vector with the corresponding normalized weight value to generate a health discrimination factor with the weighted representation as the unified input.

[0042] In this embodiment, the health discrimination factor refers to a compact representation obtained by KPCA dimensionality reduction and entropy weight assignment, which is used as a unified input for subsequent initial screening and grading assessment.

[0043] Specifically, the weighted representation expression between the projected representation of each sample vector and its corresponding normalized weight value is shown below: (6) in, This represents the weighted representation of the principal component projection of the corresponding sample vector, where normalized weight values ​​are applied. This represents the health discriminant factor with a weighted representation as a unified input. They represent the first to the last. The normalized weights of the principal components retained by the KPCA algorithm. Indicates the first The projection representation of each sample vector onto the principal component direction.

[0044] In this embodiment, the original high-dimensional observations are compressed into a small number of highly discriminative "health factors" using the KPCA algorithm, providing a compact input for subsequent weighting and discrimination.

[0045] S20: Train the health discrimination factors using an extreme learning machine, and based on the training results, batch filter the sampling channels corresponding to health discrimination factors that exceed the preset threshold and mark the abnormal sampling channels to obtain preliminary candidate samples.

[0046] Specifically, the schematic diagram of the Extreme Learning Machine structure in this embodiment is as follows: Figure 3 As shown, the health discrimination factors are trained and scored using the Extreme Learning Machine algorithm in a closed-form solution, prioritizing the screening of suspected abnormal sampling channel segments to provide a high-confidence candidate set for subsequent fine-grained classification and trend assessment. Figure 4 As shown, step S20 specifically includes: S201: Transform the health discriminant factors into a unified matrix representation for the extreme learning machine for data training. The expression for the unified matrix representation is shown below: (6) in, This represents the hidden layer output matrix of the Extreme Learning Machine. This represents the output weight matrix of the extreme learning machine. This represents the expected data of the Extreme Learning Machine.

[0047] The hidden layer output matrix expression is shown below: (7) in, This represents the activation function of the Extreme Learning Machine. , Indicates the first The input weights and biases of the hidden nodes, The input samples of the Extreme Learning Machine are the health discriminant factors. Input the number of samples for the Extreme Learning Machine. This indicates the number of neurons in the hidden layer.

[0048] The expression for the output weight matrix is ​​as follows: (8) in, Represent the Moore–Penrose generalized inverse matrix, when When it is reversible, we have: (9) S202: Based on the training results, the initial screening output of the health discriminant factor is calculated using an extreme learning machine. The expression for the initial screening output is as follows: (10) in, The input to the extreme learning machine represents the first... A sample vector, Indicates the first The weights of each hidden node from the hidden layer to the output layer. Indicates the first a sample vector S203: Perform anomaly detection and sample labeling on the initial screening output according to the preset anomaly detection expression to obtain the initial screening candidate samples. The anomaly detection expression is shown below: (11) in, This indicates the preset abnormal threshold.

[0049] Specifically, sample vectors that meet the anomaly judgment expression, i.e., are greater than or equal to a preset anomaly threshold, are marked as suspected anomaly candidate samples and enter the subsequent risk classification; if they are less than the preset anomaly threshold, they are marked as regular samples, so as to achieve rapid initial screening and reduction of manual review volume in offline batch processing scenarios.

[0050] This embodiment also includes: The Moore–Penrose generalized inverse matrix is ​​optimized by introducing a regularization parameter. The regularization expression for the Moore–Penrose generalized inverse matrix is ​​as follows: (12) in, This indicates the introduced regularization parameter. This represents the total number of sample vectors in the sample set comprised of the offline exported data from the sampling channels. A regularization parameter is introduced to improve generalization ability and suppress overfitting.

[0051] S30: Perform fuzzy weighted scoring on each initial screening candidate sample, match the score with a predefined risk level, and perform hierarchical quantification on the initial screening candidate samples based on the matching results.

[0052] Specifically, after the initial screening of samples, only samples from sampling channels that are identified as suspected abnormalities or potential risks undergo refined risk quantification. To achieve a monotonic mapping between the semantics of risk levels and numerical scores, four risk levels are predefined: Acceptable (A), Minor (Mi), Major (Ma), and Critical (C), with fixed level coefficients assigned to each level. .like Figure 5 As shown, step S30 specifically includes: S301: Calculate the fuzzy membership degree of each initial candidate sample. The functional expression for the fuzzy membership degree is shown below: (13) in, Indicates the first The first preliminary screening candidate sample Each feature corresponds to a diagnostic quantity In the Membership degree of risk level , They represent the first The boundary values ​​of the scoring intervals for each feature dimension at different risk levels.

[0053] S302: Calculate the fuzzy weighted score for each initial candidate sample. The expression for the fuzzy weighted score is as follows: (14) in, Indicates the first Fuzzy weighted scoring of initial candidate samples Indicates the first The first screening of candidate samples Normalized weights of each feature, Indicates a predefined risk level Fixed grade coefficient, These represent predefined risk levels.

[0054] S303: By using the predefined risk level and fuzzy weighted score mapping relationship, the preliminary candidate samples are mapped to the corresponding risk level for graded quantification based on the fuzzy weighted score calculation results.

[0055] Specifically, a mapping relationship between each risk level and its corresponding fuzzy weighted score is pre-defined, and the fuzzy weighted score calculation result of each sample vector is assigned to the corresponding risk level according to the mapping relationship, thereby completing the risk classification of the initial screening candidate samples. The risk level is quantitatively divided by the fuzzy weighted score of the batch of initial screening candidate samples.

[0056] This embodiment also considers that the membership functions of different risk levels often overlap, which may lead to large fluctuations in the calculated scores within the boundary region, thereby reducing the stability of risk characterization. To avoid this problem, this embodiment introduces a membership normalization term to normalize the fuzzy weighted score based on formula (14). The function expression for the fuzzy weighted score normalization is as follows: (15) Among them, the denominator This represents the sum of fuzzy responses under all risk levels, ensuring that the scoring function remains normalized under the superposition of multiple levels of fuzziness, and improving the stability of fuzzy scoring results in the fuzzy interval.

[0057] S40: Based on the risk level quantification results of each initial screening candidate sample, the corresponding abnormal sampling channels are ranked by channel risk, and maintenance decisions are made based on the ranking results.

[0058] Specifically, based on the risk level quantification results of each initial screening candidate sample, the corresponding abnormal sampling channels are prioritized for maintenance. The higher the risk level, the higher the priority, and the more priority they need to be maintained. The relevant data of the corresponding abnormal sampling channels are then sent to the management terminal for maintenance decision-making according to the ranking results.

[0059] In one embodiment, to verify the model's predictive performance, this embodiment uses PR (Precision-Recall) curves and DET (Detection Error Tradeoff) curves to examine the diagnostic effectiveness. The PR curve depicts the trade-off between precision and recall (also known as TPR) at different discrimination thresholds, and is particularly suitable for imbalanced scenarios where the proportion of anomalous samples is low. In such scenarios, relying solely on an indicator with the false positive rate on the horizontal axis may lead to overly optimistic conclusions, while the PR curve can more directly reflect the effectiveness of "detecting anomalous samples." An ideal model should maintain high precision at a high recall rate, making the PR curve as close to the upper right corner as possible. Furthermore, the overall performance can be quantified using AUPR (Area Under the PR Curve); a higher AUPR indicates better overall performance.

[0060] like Figure 6-7 As shown, the larger the area under the PR curve (AUC), the better the prediction accuracy of the model. This model is compared with commonly used SVM and simplified models without initial screening (Fuzzy Entropy-weighted Risk Assessment, FERA). It is evident that this model has high accuracy in predicting the risks of photovoltaic power plants under multi-source bias data environments. For the DET curve, the model used in this paper (Hybrid-Feature and Trend-aware Risk Assessment, HFTRA) significantly outperforms the compared models.

[0061] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0062] In one embodiment, an offline diagnostic system for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers is provided. This system corresponds to the offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers described in the above embodiment. Figure 8 As shown, the system includes a data preprocessing module, a data filtering module, a risk classification module, and a risk diagnosis module. Detailed descriptions of each functional module are as follows: The data preprocessing module is used to acquire multi-source offline data from the sampling channels of voltage transformers in substations, extract multi-source features from the multi-source offline data to form sample vectors, preprocess the sample vectors, and generate a unified input health discrimination factor.

[0063] The data filtering module is used to train the health discrimination factors through extreme learning machine, and to batch filter the sampling channels corresponding to the health discrimination factors that exceed the preset threshold based on the training results and mark the abnormal sampling channels to obtain the initial candidate samples.

[0064] The risk grading module is used to perform fuzzy weighted scoring on each initial screening candidate sample, match the score value with a predefined risk level, and perform grading and quantification of the initial screening candidate samples based on the matching results.

[0065] The risk diagnosis module is used to rank the abnormal sampling channels corresponding to each initial screening candidate sample based on the risk level quantification results, and to make maintenance decisions based on the ranking results.

[0066] Preferably, the data preprocessing module specifically includes: The sample preprocessing submodule is used to acquire offline exported data from the sampling channels of substation voltage transformers, extract multi-source features of the offline exported data within the same sampling time window according to the sampling time to form a sample vector, and preprocess the sample vector.

[0067] The principal component projection submodule is used to construct the kernel matrix of the preprocessed sample vectors and perform eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector in the principal component direction.

[0068] The weight calculation submodule is used to calculate the information entropy of the sample vector and the difference of the information entropy. Based on the difference, the corresponding sample vector is normalized to obtain the normalized weight value of the sample vector.

[0069] The weighted representation submodule is used to weight the projected representation of each sample vector with the corresponding normalized weight value to generate a health discrimination factor with the weighted representation as the unified input.

[0070] Preferably, the principal component projection submodule specifically includes: The projection of the sample vector onto the principal component direction is expressed as follows: (1) in, Indicates the first The projection of the i-th sample vector onto the principal component direction, i.e., the i-th sample vector... Principal component vectors after dimensionality reduction of each sample vector Indicates the first The nth sample vector The eigenvector coefficients of the corresponding principal components of each feature. Indicates the first The number of multi-source features in a sample vector The Gaussian radial basis function kernel is expressed as follows: (2) in, Representing vectors , The square Euclidean distance between them , Indicates the first The, the A sample vector, This represents a predefined parameter whose value depends on the training dataset. This represents a free parameter.

[0071] Preferably, the weight calculation submodule includes: The expression for calculating the information entropy of a sample vector is shown below: (3) in, Indicates the first The th sample vector Information entropy of each feature Indicates the first The number of multi-source features in a sample vector Indicates the normalized i-th The feature in the first The proportion in each sample vector.

[0072] The difference degree of information entropy is calculated, and the expression for the difference degree is as follows: (4) in, Indicates the first The th sample vector The degree of difference of each feature.

[0073] The corresponding sample vectors are normalized based on the degree of difference to obtain the normalized weights of the sample vectors. The calculation expression for the normalized weights is as follows: (5) in, Indicates the first The th sample vector Normalized weights of each feature, This represents the number of principal components retained by the KPCA algorithm.

[0074] Preferably, the weighted representation submodule includes: The weighted representation expression between the projected representation of each sample vector and its corresponding normalized weight value is shown below: (6) in, This represents the weighted representation of the principal component projection of the corresponding sample vector, where normalized weight values ​​are applied. This represents the health discriminant factor with a weighted representation as a unified input. They represent the first to the last. The normalized weights of the principal components retained by the KPCA algorithm. Indicates the first The projection representation of each sample vector onto the principal component direction.

[0075] Preferably, the data filtering module includes: The data training submodule is used to convert health discriminant factors into a unified matrix representation for extreme learning machine training. The expression for the unified matrix representation is shown below: (6) in, This represents the hidden layer output matrix of the Extreme Learning Machine. This represents the output weight matrix of the extreme learning machine. This represents the expected output of the Extreme Learning Machine.

[0076] The hidden layer output matrix expression is shown below: (7) in, This represents the activation function of the Extreme Learning Machine. , Indicates the first The input weights and biases of the hidden nodes, The input samples of the Extreme Learning Machine are the health discriminant factors. This represents the number of samples input to the Extreme Learning Machine. This indicates the number of neurons in the hidden layer.

[0077] The expression for the output weight matrix is ​​as follows: (8) in, Represent the Moore–Penrose generalized inverse matrix, when When it is reversible, we have: (9) The initial screening output submodule is used to calculate the initial screening output of the health discriminant factor using the extreme learning machine based on the training results. The expression for the initial screening output is as follows: (10) in, The input to the extreme learning machine represents the first... A sample vector, Indicates the first The weights of each hidden layer node from the hidden layer to the output layer. Indicates the first Activation function for each sample vector.

[0078] The anomaly detection submodule is used to perform anomaly detection and sample labeling on the initial screening output according to a preset anomaly detection expression, to obtain initial screening candidate samples. The anomaly detection expression is shown below: (11) in, This indicates the preset abnormal threshold.

[0079] Preferably, the data filtering module also includes: The regularization optimization submodule is used to introduce regularization parameters to optimize the Moore–Penrose generalized inverse matrix. The regularization expression for the Moore–Penrose generalized inverse matrix is ​​shown below: (12) in, This indicates the introduced regularization parameter. This represents the total number of sample vectors in the sample set composed of the offline exported data from the sampling channel.

[0080] Preferably, the risk classification module includes: The membership calculation submodule is used to calculate the fuzzy membership degree of each initially screened candidate sample. The function expression for the fuzzy membership degree is shown below: (13) in, Indicates the first The first preliminary screening candidate sample Diagnostic value corresponding to each feature In the Membership degree of risk level , They represent the first The boundary values ​​of the scoring intervals for each feature dimension at different risk levels.

[0081] The weighted scoring submodule is used to calculate the fuzzy weighted score for each initially screened candidate sample. The fuzzy weighted scoring expression is shown below: (14) in, Indicates the first Fuzzy weighted scores of initial screening candidate samples Indicates the first The first preliminary screening candidate sample Normalized weights of each feature, Indicates a predefined risk level Fixed grade coefficient, These represent predefined risk levels.

[0082] The hierarchical quantification submodule is used to map the initial screening candidate samples to the corresponding risk levels for hierarchical quantification based on the fuzzy weighted score calculation results, according to the predefined risk level and fuzzy weighted score level coefficient mapping relationship.

[0083] Preferably, the risk classification module also includes: The scoring normalization submodule is used to normalize the fuzzy weighted scores. The function expression for fuzzy weighted score normalization is shown below: (15) in, This represents the sum of fuzzy responses across all risk levels.

[0084] Specific limitations regarding the offline diagnostic system for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels can be found in the above-mentioned limitations on the offline diagnostic methods for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels, and will not be repeated here. Each module in the aforementioned offline diagnostic system for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0085] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application of the technical solution and the constraints involved. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0086] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0087] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers, characterized in that, The method includes: Acquire multi-source offline data from the sampling channels of voltage transformers in substations, extract multi-source features from the multi-source offline data to form sample vectors, preprocess the sample vectors, and generate a unified input health discrimination factor; The health discrimination factors are trained using an extreme learning machine. Based on the training results, the sampling channels corresponding to health discrimination factors that exceed the preset threshold are batch-screened and abnormal sampling channels are marked to obtain preliminary candidate samples. Each of the initial screening candidate samples is subjected to fuzzy weighted scoring, and the score value is matched with a predefined risk level. The initial screening candidate samples are then classified and quantified based on the matching results. Based on the risk level quantification results of each preliminary screening candidate sample, the abnormal sampling channels corresponding to the preliminary screening candidate samples are ranked by channel risk, and maintenance decision processing is carried out based on the ranking results.

2. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels according to claim 1, characterized in that, The process of acquiring multi-source offline data from the sampling channels of substation voltage transformers, extracting multi-source features from the multi-source offline data to form sample vectors, preprocessing the sample vectors, and generating a unified input health discrimination factor specifically includes: The offline exported data of the voltage transformer sampling channel in the substation is obtained. Multi-source features of the offline exported data within the same sampling time window are extracted according to the sampling time to form a sample vector, and the sample vector is preprocessed. Construct a kernel matrix of the preprocessed sample vectors, and perform eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector in the principal component direction; Calculate the information entropy of the sample vector and the difference degree of the information entropy. Normalize the corresponding sample vector according to the difference degree to obtain the normalized weight value corresponding to the sample vector. The projected representation of each sample vector is weighted with its corresponding normalized weight value to generate a health discriminant factor with the weighted representation as a unified input.

3. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels according to claim 2, characterized in that, The process of constructing a kernel matrix for the preprocessed sample vectors and performing eigenvalue decomposition based on the kernel matrix to obtain the projection representation of each sample vector along the principal component direction specifically includes: The projection expression of the sample vector onto the principal component direction is shown below: （1） in, Indicates the first The projection of the i-th sample vector onto the principal component direction, i.e., the i-th sample vector... Principal component vectors after dimensionality reduction of each sample vector Indicates the first The nth sample vector The eigenvector coefficients of the corresponding principal components of each feature. Indicates the first The number of multi-source features in a sample vector The Gaussian radial basis function kernel is expressed as follows: （2） in, Representing vectors , The square Euclidean distance between them , Indicates the first The, the A sample vector, This represents a predefined parameter whose value depends on the training dataset. This represents a free parameter.

4. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels according to claim 2, characterized in that, The process of calculating the information entropy of the sample vector and the difference degree of the information entropy, and then normalizing the corresponding sample vector based on the difference degree to obtain the normalized weight value corresponding to the sample vector includes: The expression for calculating the information entropy of the sample vector is as follows: （3） in, Indicates the first The th sample vector Information entropy of each feature Indicates the first The number of multi-source features in a sample vector Indicates the normalized i-th The feature in the first The proportion in each sample vector; The difference degree of the information entropy is calculated, and the expression for the difference degree is as follows: （4） in, Indicates the first The th sample vector The degree of difference of each feature; The corresponding sample vectors are normalized based on the difference to obtain the normalized weights of the sample vectors. The calculation expression for the normalized weights is as follows: （5） in, Indicates the first The th sample vector Normalized weights of each feature, This represents the number of principal components retained by the KPCA algorithm.

5. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of substation voltage transformer sampling channels according to claim 2, characterized in that, The step of weighting the projected representation of each sample vector with its corresponding normalized weight value to generate a health discrimination factor with the weighted representation as a unified input includes: The weighted representation expression between the projected representation of each sample vector and its corresponding normalized weight value is shown below: （6） in, This represents the weighted representation of the principal component projection of the corresponding sample vector, where normalized weight values ​​are applied. This represents the health discriminant factor with a weighted representation as a unified input. They represent the first to the last. The normalized weights of the principal components retained by the KPCA algorithm. Indicates the first The projection representation of each sample vector onto the principal component direction.

6. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers according to claim 1, characterized in that, The process of training the health discrimination factors using an extreme learning machine, and then batch-filtering sampling channels corresponding to health discrimination factors exceeding a preset threshold based on the training results and marking abnormal sampling channels to obtain preliminary candidate samples includes: The health discriminant factors are converted into a unified matrix representation for extreme learning machine training. The expression for the unified matrix representation is as follows: （6） in, This represents the hidden layer output matrix of the Extreme Learning Machine. This represents the output weight matrix of the extreme learning machine. This represents the expected output of the Extreme Learning Machine; The hidden layer output matrix expression is as follows: （7） in, This represents the activation function of the Extreme Learning Machine. , Indicates the first The input weights and biases of the hidden nodes, The input samples of the Extreme Learning Machine are the health discriminant factors. This represents the number of samples input to the Extreme Learning Machine. Indicates the number of neurons in the hidden layer; The expression for the output weight matrix is ​​as follows: （8） in, Represent the Moore–Penrose generalized inverse matrix, when When it is reversible, we have: （9） Based on the training results, the initial screening output of the health discriminant factor is calculated using an extreme learning machine. The expression for the initial screening output is as follows: （10） in, The input to the extreme learning machine represents the first... A sample vector, Indicates the first The weights of each hidden node from the hidden layer to the output layer. Indicates the first Activation function for each sample vector; The initial screening output is processed for anomaly detection and sample labeling based on a preset anomaly detection expression to obtain initial screening candidate samples. The anomaly detection expression is shown below: （11） in, This indicates the preset abnormal threshold.

7. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers according to claim 6, characterized in that, The step of training the health discrimination factors using an extreme learning machine, batch-filtering sampling channels corresponding to health factors exceeding a preset threshold based on the training results, and marking abnormal sampling channels to obtain preliminary candidate samples, further includes: The Moore–Penrose generalized inverse matrix is ​​optimized by introducing a regularization parameter. The regularization expression for the Moore–Penrose generalized inverse matrix is ​​as follows: （12） in, This indicates the introduced regularization parameter. This represents the total number of sample vectors in the sample set composed of the offline exported data from the sampling channel.

8. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers according to claim 1, characterized in that, The step of performing fuzzy weighted scoring on each of the initial screening candidate samples, matching the score values ​​with predefined risk levels, and then performing graded quantification on the initial screening candidate samples based on the matching results includes: The fuzzy membership degree of each of the initial screening candidate samples is calculated, and the functional expression of the fuzzy membership degree is as follows: （13） in, Indicates the first The first preliminary screening candidate sample Diagnostic value corresponding to each feature In the Membership degree of risk level , They represent the first The boundary values ​​of the scoring intervals for each feature dimension at different risk levels; The fuzzy weighted score for each of the initial candidate samples is calculated, and the fuzzy weighted score expression is as follows: （14） in, Indicates the first Fuzzy weighted scores of initial screening candidate samples Indicates the first The first preliminary screening candidate sample Normalized weights of each feature, Indicates a predefined risk level Fixed grade coefficient, These represent predefined risk levels; By using the predefined risk level and fuzzy weighted score mapping relationship, the preliminary screening candidate samples are mapped to the corresponding risk level for graded quantification based on the fuzzy weighted score calculation results.

9. The offline diagnostic method for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers according to claim 8, characterized in that, The step of performing fuzzy weighted scoring on each of the initial screening candidate samples, matching the score values ​​with predefined risk levels, and performing graded quantification on the initial screening candidate samples based on the matching results further includes: The fuzzy weighted score is normalized, and the function expression for fuzzy weighted score normalization is shown below: （15） in, This represents the sum of fuzzy responses across all risk levels.

10. An offline diagnostic system for nonlinear characteristic compression, initial screening, and graded evaluation of sampling channels of substation voltage transformers, characterized in that, The system includes: The data preprocessing module is used to acquire multi-source offline data from the sampling channels of voltage transformers in substations, extract multi-source features from the multi-source offline data to form sample vectors, and preprocess the sample vectors to generate a unified input health discrimination factor. The data filtering module is used to train the health discrimination factors using an extreme learning machine, and to batch filter the sampling channels corresponding to health discrimination factors that exceed a preset threshold based on the training results and mark abnormal sampling channels to obtain preliminary candidate samples. The risk grading module is used to perform fuzzy weighted scoring on each of the initial screening candidate samples, match the score value with a predefined risk level, and perform grading and quantification on the initial screening candidate samples based on the matching result. The risk diagnosis module is used to rank the abnormal sampling channels corresponding to each preliminary screening candidate sample according to the risk level quantification result, and to make maintenance decisions based on the ranking result.

Images