A transformer state monitoring method, device and electronic equipment

Abstract

This application discloses a transformer condition monitoring method, device, and electronic equipment, relating to the field of power equipment. It involves sequentially performing data normalization, outlier detection, and missing value processing on the acquired raw transformer data. The correlation between various features in the preprocessed data is analyzed, and the preprocessed data is divided according to the analysis results to obtain feature data characterizing transformer condition changes. A prediction model is invoked to extract and predict features from the feature data, obtaining feature prediction data for a preset future time period. Key feature data is extracted from the feature prediction data using a feature selection algorithm, and a fault identification model is invoked to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing an artificial neural network model or support vector machine based on a gravity search algorithm. By performing multiple data processing and filtering processes from data preprocessing to model prediction, the reliability and accuracy of subsequent model predictions are ensured.

Description

Technical Field

[0001] This application relates to the field of power equipment technology, and in particular to a transformer condition monitoring method, device and electronic equipment. Background Technology

[0002] As a key electrical equipment in the entire power grid system, power transformers play a vital role in power generation, transmission, transformation, and distribution. Their operating condition directly affects the stability of the power grid and production safety. A failure in a large power transformer can lead to major accidents such as large-scale power outages. Therefore, monitoring the operating conditions of large power transformers is crucial. Because transformers are located at the center of the power grid and operate in a complex environment, they are frequently subjected to various adverse operating conditions, making them more prone to failure. Therefore, how to achieve more accurate and reliable monitoring of transformer conditions has become an urgent problem to be solved. Summary of the Invention

[0003] In view of the above problems, this application provides a transformer condition monitoring method, device, and electronic equipment to improve the reliability and accuracy of prediction. The specific solution is as follows:

[0004] The first aspect of this application provides a transformer condition monitoring method, including:

[0005] The acquired raw transformer data is subjected to data normalization, outlier detection, and missing value processing in sequence to obtain preprocessed data of the raw transformer data.

[0006] The correlation between the features in the preprocessed data is analyzed, and the preprocessed data is divided according to the analysis results to obtain feature data characterizing the changes in the transformer state;

[0007] The prediction model is invoked to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset time period in the future.

[0008] Key feature data is extracted from the feature prediction data according to the feature selection algorithm, and the fault identification model is called to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing the artificial neural network model or support vector machine based on the gravity search algorithm.

[0009] In one possible implementation, the process of sequentially performing data normalization, outlier detection, and missing value processing on the acquired raw transformer data to obtain preprocessed data of the raw transformer data includes:

[0010] The original transformer data is normalized based on the standard deviation normalization algorithm;

[0011] The isolated forest algorithm is used to identify outliers in the original transformer data and remove the outliers.

[0012] Missing data is filled in using the completion strategy corresponding to the missing data pattern.

[0013] In one possible implementation, the outlier identification of the original transformer data based on the isolated forest algorithm includes:

[0014] Determine the path length of the data point from the root node to the isolated leaf node and the average path length;

[0015] Based on the path length and the average path length, an outlier score for the data point is determined.

[0016] If the outlier score is greater than the outlier score threshold, then the value of the data point is determined to be an outlier.

[0017] In one possible implementation, the missing value completion process based on the completion strategy corresponding to the missing data pattern includes:

[0018] When the missing pattern is completely random, a corresponding type imputation strategy is determined based on the data distribution characteristics, and missing value completion is performed based on the type imputation strategy.

[0019] When the missing pattern is random missing, missing value completion is performed based on the least squares regression algorithm or the support vector regression algorithm.

[0020] When the missing pattern is non-random missing, the missing values ​​are imputed based on the observed data to obtain multiple imputed datasets, and the multiple imputed datasets are merged to obtain the missing value completion result.

[0021] In one possible implementation, the step of analyzing the correlation between features in the preprocessed data and dividing the preprocessed data according to the analysis results to obtain feature data characterizing the transformer state changes includes:

[0022] The correlation coefficient algorithm is used to determine the correlation values ​​that characterize the degree of correlation between the parameters in the preprocessed data, and the preprocessed data is divided based on the correlation values.

[0023] In one possible implementation, the invocation of the prediction model to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset future time period includes:

[0024] Based on the convolutional network in the prediction model, local and global features are extracted from the feature data to obtain a high-order spatial feature sequence.

[0025] The high-order spatial feature sequence is predicted based on the long short-term memory network in the prediction model to obtain the feature prediction data.

[0026] In one possible implementation, the step of extracting key feature data from the feature prediction data according to the feature selection algorithm and calling the fault identification model to perform fault identification processing on the key feature data includes:

[0027] Based on the feature selection algorithm, a preset number of feature data with a correlation greater than a preset value are selected from the feature prediction data, and the fault identification model is called to perform fault identification processing on the preset number of feature data.

[0028] In one possible implementation, the optimization process of the fault identification model includes:

[0029] The fitness function in the gravity search algorithm is used to determine the quality score that characterizes the model performance, and the parameters of the model are adjusted based on the quality score.

[0030] A second aspect of this application provides a transformer condition monitoring device, comprising:

[0031] The data preprocessing module is used to perform data normalization, outlier detection, and missing value processing on the acquired raw transformer data in sequence to obtain the preprocessed data of the raw transformer data.

[0032] The correlation analysis module is used to analyze the correlation between the features in the preprocessed data, and divide the preprocessed data according to the analysis results to obtain feature data characterizing the changes in the transformer state.

[0033] The data prediction module is used to call a prediction model to perform feature extraction and prediction processing on the feature data, and obtain feature prediction data within a preset time period in the future; and,

[0034] The fault prediction module is used to extract key feature data from the feature prediction data according to the feature selection algorithm, and call the fault identification model to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing an artificial neural network model or support vector machine based on the gravity search algorithm.

[0035] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the transformer condition monitoring method described in the first aspect or any implementation thereof.

[0036] A fourth aspect of this application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0037] The memory is used to store computer programs;

[0038] The processor is used to execute the computer program so that the electronic device can implement the transformer condition monitoring method of the first aspect or any implementation thereof.

[0039] The fifth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the transformer condition monitoring method described in the first aspect or any implementation thereof.

[0040] By employing the aforementioned technical solution, the transformer condition monitoring method provided in this application preserves the distribution characteristics of the original data and ensures that each feature dimension has equal importance weight in subsequent modeling and analysis by sequentially performing data normalization, outlier detection, and missing value processing on the acquired raw transformer data. The correlation between features in the preprocessed data is analyzed, and the preprocessed data is divided according to the analysis results to obtain feature data characterizing transformer condition changes. A prediction model is invoked to extract and predict features from the feature data, obtaining feature prediction data for a preset future time period. Key feature data is extracted from the feature prediction data using a feature selection algorithm, and a fault identification model is invoked to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing an artificial neural network model or support vector machine based on a gravity search algorithm. Through multiple data processing and filtering processes from data preprocessing to model prediction, the reliability and accuracy of subsequent model predictions are guaranteed. Attached Figure Description

[0041] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0042] Figure 1 A flowchart of a transformer condition monitoring method provided in this application;

[0043] Figure 2 A flowchart for data normalization and outlier handling is provided in this application;

[0044] Figure 3 The result diagram of the hydrogen gas before abnormal data processing provided in this application;

[0045] Figure 4 A graph showing the results of processing the abnormal hydrogen data provided in this application;

[0046] Figure 5 The resulting graph shows the methane data provided in this application before anomaly processing;

[0047] Figure 6 The resulting graph shows the methane data provided in this application after anomaly data processing;

[0048] Figure 7 A graph showing the results of data error compensation for the hydrogen data provided in this application;

[0049] Figure 8 A graph showing the results of data error compensation for the methane data provided in this application;

[0050] Figure 9 Spearman correlation coefficient analysis plot of the DGA data provided in this application;

[0051] Figure 10 The long short-term memory neural network structure diagram provided in this application;

[0052] Figure 11 The overall flowchart of the hybrid feature selection and gravity search algorithm provided in this application;

[0053] Figure 12 A flowchart of the model optimization based on the gravity search algorithm provided for this application;

[0054] Figure 13 This application provides a structural diagram of a transformer condition monitoring device.

[0055] Figure 14 This is a structural diagram of an electronic device provided in this application. Detailed Implementation

[0056] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0057] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0058] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0059] The rapid increase in electricity demand not only places higher demands on transformer performance but also presents unprecedented challenges to their reliability and safety. Extreme environmental changes, such as high temperatures, humidity, and smog, as well as aging and corrosion of the equipment itself due to long-term operation, can all lead to reduced transformer efficiency and increased failure rates. Many power transformer accidents are caused by a lack of monitoring methods, allowing minor defects in power transformers to gradually escalate into major safety incidents.

[0060] Currently, using online monitoring data, experimental data, and empirical data to assess the health status of transformers is a common method. This assessment can clearly identify transformers requiring immediate repair or replacement. Furthermore, health status assessment plays a crucial role in developing transformer maintenance and operation strategies, guiding which transformers require enhanced monitoring, which should be scheduled for outage maintenance, and which have serious fault risks requiring immediate outage maintenance. Condition-based maintenance strategies also include using time-series analysis of online monitoring data to predict future transformer health trends. This allows for early detection of potential fault signs, making maintenance plans more rational and timely. This proactive maintenance strategy not only improves transformer reliability but also ensures the stable operation of the power system. Although a large amount of transformer sensing data exists, real data is noisy and incomplete. For example, data from oil gas detection, vibration, and temperature monitoring in power transformers increases in scale over time, and the amount of abnormal data increases continuously. At the same time, different data standards from heterogeneous data sources further exacerbate the difficulty of data cleaning. Data cleaning of multi-source heterogeneous data usually occurs before and after data fusion, and data fusion is optional and proactive; fixed cleaning procedures cannot flexibly handle this process.

[0061] To address the aforementioned problems, this application provides a transformer condition monitoring method. The transformer condition monitoring method of this application embodiment will be described in detail below with reference to the accompanying drawings.

[0062] Reference Figure 1 , Figure 1 This is a flowchart illustrating a transformer condition monitoring method provided in an embodiment of this application, as shown below. Figure 1 As shown in the figure, the transformer condition monitoring method provided in this application embodiment may include steps S101 to S104, which are described in detail below.

[0063] Step S101: Perform data normalization, outlier detection, and missing value processing on the obtained raw transformer data in sequence to obtain preprocessed raw transformer data.

[0064] Specifically, the Z-score method (standard deviation normalization algorithm) can be used to normalize the multi-dimensional sensing data of hydrogen, methane, and other mixtures in the DGA (Dissolved Gas Analysis) data of the collected transformer data. Then, the isolated forest algorithm is used for collaborative cleaning of outliers. Finally, the processed gas mixture data is completed using appropriate completion strategies according to the missing data pattern.

[0065] Step S102: Analyze the correlation between the features in the preprocessed data, and divide the preprocessed data according to the analysis results to obtain feature data characterizing the changes in transformer state.

[0066] After obtaining the preprocessed data, the correlation between multiple parameters can be analyzed. In terms of analysis methods, Spearman's correlation coefficient can be used to calculate the correlation, and the data can be further divided based on the correlation results to obtain data with influence relationships.

[0067] Step S103: Call the prediction model to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset time period in the future.

[0068] Furthermore, a multi-feature temporal prediction model based on a two-dimensional convolutional neural network and a variant of a long short-term memory network can be used for model prediction and classification to obtain the corresponding prediction data.

[0069] Step S104: Extract key feature data from the feature prediction data according to the feature selection algorithm, and call the fault identification model to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing the artificial neural network model or support vector machine based on the gravity search algorithm.

[0070] Specifically, significant features can be extracted from the aforementioned prediction data using feature selection methods. Furthermore, the fault identification model can be optimized using gravity search algorithms, such as fault identification models built on SVM (Support Vector Machine) and ANN (Artificial Neural Network).

[0071] This transformer condition monitoring method normalizes the data, then uses the Isolation Forest algorithm for outlier detection, further constructs a corresponding completion strategy, verifies data consistency, and finally utilizes hybrid feature selection and gravity search algorithms to reduce redundant input and optimize the cleaning of multi-dimensional sensing information parameters. This approach not only preserves the distribution characteristics of the original data but also ensures that each feature dimension has equal importance weight in subsequent modeling and analysis. The Isolation Forest algorithm can naturally handle large-scale datasets and has good scalability. The hybrid feature selection and gravity search algorithms are beneficial for optimizing key parameters of the artificial intelligence model, improving fault classification accuracy, achieving multi-dimensional sensing data cleaning, and enabling value judgment and screening of the data before and after cleaning, thereby making the fault identification model's identification results more accurate and reliable.

[0072] In some embodiments, to improve the effectiveness of data preprocessing, the above-mentioned data normalization, outlier detection, and missing value processing are performed sequentially on the acquired raw transformer data to obtain preprocessed raw transformer data, which may specifically include:

[0073] The original transformer data is normalized using the standard deviation normalization algorithm.

[0074] The isolated forest algorithm is used to identify and remove outliers from the original transformer data.

[0075] Missing data is filled in using the completion strategy corresponding to the missing data pattern.

[0076] Specifically, the original data is transformed into a standard normal distribution with a mean of 0 and a standard deviation of 1 through linear transformation. The mathematical expression of the Z-score standardization method is:

[0077] (1);

[0078] Where x represents the original data point, m is the arithmetic mean of the dataset, s is the standard deviation, and x' is the standardized data value.

[0079] The Z-score standardization method transforms the data distribution into a standard scale through two steps: centering (subtracting the mean) and scaling (dividing by the standard deviation). This preserves the original data's distribution characteristics and ensures that each feature dimension has equal importance and weight in subsequent modeling and analysis. It provides a more accurate data foundation for handling outliers in the original data based on the results, further improving the effectiveness of the entire multi-dimensional perceptual information parameter cleaning method.

[0080] For specific procedures on handling outliers, please refer to [link / reference]. Figure 2 As shown, initialize the parameters of the isolated forest model, including key hyperparameters such as the number of subtrees (n_estimators) and the maximum depth (max_depth), and build the initial anomaly detection model.

[0081] The algorithm randomly selects feature dimensions and generates split points within their value range to achieve a bipartite partition of the data space. This recursive process continues until one of the following termination conditions is met: the preset maximum tree depth is reached or the subset size is reduced to a single instance.

[0082] Calculating the path length from the root node to the isolated leaf node allows for the quantification of the degree of data anomalies across multiple dimensions. The formula for calculating the path length is:

[0083] (2);

[0084] Among them, h i (x j Let x be the data point in the i-th tree. j The path length is given, and the total number of trees is T.

[0085] Furthermore, outlier scores are calculated; data points with shorter path lengths are typically considered outliers. Outlier score s(x) j The formula for calculating ) is:

[0086] (3);

[0087] Where h(x) j ) is the data point x j The average path length is given by c(n), which is the average path length under normal circumstances.

[0088] Specifically, c(n) can be calculated based on the number of sample points n, and the calculation formula is as follows: Formula (4)

[0089] (4);

[0090] The outlier score ranges from [0, 1]. The closer the score is to 1, the more likely the data point is to be an outlier.

[0091] Furthermore, based on actual multi-dimensional perception information parameters, a threshold Sthreshold is set to distinguish between normal points and outliers. When the outlier score s(x) of a data point is... j If the value is greater than the threshold, the point is considered an outlier.

[0092] Figures 3 to 6 The data shows Z-score normalization for hydrogen and methane gas data, and a comparison of scatter plots before and after processing with the Isolation Forest algorithm. Visualizing the data intuitively demonstrates the effectiveness of outlier detection and removal. Figure 3 The figure shown is a graph illustrating the hydrogen data of the transformer before abnormal data processing in one embodiment of this application. Figure 4 The figure shown is a graph illustrating the result of abnormal data processing of the transformer's hydrogen data in one embodiment of this application. Figure 5 The image shown is a result of abnormal data processing for the transformer methane data in one embodiment of this application. Figure 6 The image shown is a result of abnormal data processing on the transformer's methane data in one embodiment of this application. It can be seen that a significant data cleaning effect has been achieved.

[0093] When performing missing value completion processing, missing data can be completed based on the completion strategy corresponding to the missing data pattern. In this case, missing pattern identification can be performed on the gas mixture data after processing in step S101. The missing patterns include completely random missing, random missing and non-random missing.

[0094] Specifically, let the original data matrix be:

[0095] (5);

[0096] Where n represents the number of samples and 𝑝 represents the number of feature variables. Further, the corresponding missing indicator matrix is ​​constructed:

[0097] (6);

[0098] Where, x ij When missing, m ij =1; otherwise, m ij It is 0.

[0099] By analyzing the distribution characteristics of the missing indicator matrix 𝑀, and combining the missing rate of each variable, the distribution of missing locations, the co-occurrence relationship of missing items, and the correlation between missing items and observed data, the missing patterns are determined.

[0100] Furthermore, using the missing indicator variable 𝑀 and the observed data X obs Unobserved data X misBased on the dependencies between them, missing patterns satisfy the following classification criteria:

[0101] when

[0102] (7);

[0103] When the condition is met, it is determined to be completely random and missing.

[0104] when

[0105] (8);

[0106] When the time is right, it is determined to be a random missing item;

[0107] (9);

[0108] When the time is right, it is determined to be a non-random missing number;

[0109] Specifically, after determining the missing patterns, data completion is performed using imputation methods based on central tendency measures, regression-based imputation methods, and multiple imputation methods for different missing patterns.

[0110] Specifically, we can first count the missing rate, missing location distribution, and missing co-occurrence of each variable, and then determine the missing pattern by combining the correlation between the missing data and the observed data.

[0111] Furthermore, if the occurrence of missing data is not statistically significantly correlated with any of the observed variables, it is determined to be completely random missing data; if the occurrence of missing data is statistically correlated with some of the observed variables, it is determined to be random missing data; if the occurrence of missing data cannot be fully explained by the observed variables, and based on the data distribution characteristics and actual data collection conditions, it is determined to be more likely to be related to the unobserved values ​​themselves, it is determined to be non-random missing data.

[0112] Specifically, after determining the missing patterns, data completion is performed using imputation methods based on central tendency measures, regression-based imputation methods, and multiple imputation methods for different missing patterns.

[0113] When the missing pattern is completely random, the corresponding type imputation strategy is determined based on the data distribution characteristics, and the missing value is filled in based on the type imputation strategy.

[0114] When the missing value pattern is random, missing value completion is performed based on the least squares regression algorithm or the support vector regression algorithm.

[0115] When the missing value pattern is non-random, the missing values ​​are imputed based on the observed data to obtain multiple imputed datasets. These multiple imputed datasets are then merged to obtain the missing value completion result.

[0116] Specifically, Missing Completely at Random (MCAR) is an idealized missing data mechanism. Under the MCAR mechanism, the missing patterns are not statistically correlated with any variable in the dataset, and the complete subset of data can be regarded as an unbiased sample of the population.

[0117] To address the MCAR missing data mechanism, an imputation method based on a measure of central tendency is chosen. Specifically, the choice of imputation strategy should strictly adhere to the data distribution characteristics: for continuous variables following a normal distribution, mean imputation is used; when the data exhibits significant skewness or outliers, median imputation is used; for discrete categorical variables, mode imputation is selected. Under the MCAR assumption, the unbiasedness and effectiveness of the imputation process can be ensured.

[0118] Missing at Random (MAR) is a conditionally independent missing data mechanism. Under the MAR mechanism, the missing pattern is statistically correlated with some observed variables. By controlling for the correlated variables, the complete subset of data can represent the overall distribution.

[0119] For the MAR missing mechanism, regression-based imputation methods can be divided into two categories based on the dependency structure between features: linear and nonlinear.

[0120] For the linear correlation characteristics of MAR, the regression coefficients are estimated by minimizing the sum of squared residuals using least squares regression (OLS).

[0121] When complex nonlinear relationships exist between features, the Support Vector Regression (SVR) method is selected, and the kernel trick is introduced to map the data to the Reproducing Kernel Hilbert Space (RKHS). The optimization objective is:

[0122] (10);

[0123] Where, f(x) i () represents a nonlinear mapping, C is the penalty parameter, and x is the insensitive region. By selecting an appropriate kernel function, SVR can effectively capture nonlinear patterns in high-dimensional feature spaces and exhibits good generalization ability when processing high-dimensional small sample data.

[0124] Therefore, the SVR method is chosen over OLS, offering the following advantages: 1) it enhances robustness to noise through an insensitive loss function; 2) the introduction of kernel methods overcomes the limitations of linear models; and 3) it reduces overfitting risk based on the principle of minimizing structural risk. Thus, in MAR missing scenarios with complex nonlinear relationships, SVR can provide more accurate and robust imputation results.

[0125] For the Missing Not at Random (MNAR) pattern, the probability of missing data is related to the unobserved data values ​​themselves.

[0126] Multiple imputation methods can be used. Multiple imputation datasets are generated, and the specific process is as follows: first, the missing values ​​are imputed using the observed data to generate multiple imputation datasets; then, each imputation dataset is analyzed; finally, the analysis results of each dataset are combined to obtain a comprehensive and robust estimate.

[0127] Reference Figure 7 and Figure 8 As shown, the graph illustrates the data compensation error results after applying corresponding data compensation strategies based on different missing data types for hydrogen and methane gas data. By visualizing the data, the graph intuitively displays the error results between the data after algorithm compensation and the original data.

[0128] in Figure 7 The diagram shows the result of data compensation error for the hydrogen data of the transformer in one embodiment of this application, as shown. Figure 7 As shown, the errors between the three sets of hydrogen-compensated data and the original data are 0.146, 0.733, and 0.135, respectively. Figure 8 The diagram shows the result of data compensation error for the transformer's methane data in one embodiment of this application, as shown. Figure 8 As shown, the errors between the three sets of methane-compensated data and the original data are 0.046, 0.078, and 0.039, respectively. This indicates that the data effectively compensates for missing values.

[0129] In other embodiments, the correlation between features in the preprocessed data is analyzed, and the preprocessed data is divided according to the analysis results to obtain feature data characterizing transformer state changes, including:

[0130] The correlation coefficient algorithm is used to determine the correlation values ​​that characterize the degree of correlation between parameters in the preprocessed data, and the preprocessed data is divided based on the correlation values.

[0131] Specifically, the Spearman correlation coefficient is used to assess the statistical correlation between gas characteristic variables under different fault types. This coefficient is suitable for nonlinear and non-normally distributed data analysis and is relatively robust to outliers.

[0132] When determining association based on the Spearman correlation coefficient and the corresponding significance test p-value, the following can be specifically included:

[0133] When two sets of gas characteristic variables show a significant strong correlation (the absolute value of the correlation coefficient is close to 1 and the P value is < 0.05), it is sufficient to retain one set of characteristic variables in the subsequent fault prediction analysis.

[0134] When the correlation between two sets of gas characteristic variables is weak or not significant (the absolute value of the correlation coefficient is close to 0 or the P value is ≥0.05), it is determined that the two sets of characteristic variables will be retained for transformer fault diagnosis.

[0135] in:

[0136] The Spearman correlation coefficient ranges from [-1, +1]: -1 indicates a perfect negative correlation, 0 indicates no linear correlation, and +1 indicates a perfect positive correlation.

[0137] The p-value, as a statistical significance indicator, is used to determine the reliability of the correlation coefficient; a p-value less than the preset significance level (e.g., 0.05) indicates that the correlation is statistically significant. Figure 9 The image shows a Spearman correlation coefficient analysis plot of the transformer's DGA data in one embodiment of this application. The correlation between the data points is clearly shown, facilitating data segmentation.

[0138] When calling a prediction model to extract and predict features from feature data to obtain predicted feature data for a preset future time period, the specific steps may include:

[0139] The convolutional network in the prediction model is used to extract local and global features from the feature data to obtain a high-order spatial feature sequence.

[0140] The high-order spatial feature sequences are predicted using the long short-term memory network in the prediction model to obtain feature prediction data.

[0141] Specifically, a multi-feature temporal prediction model based on two-dimensional convolutional neural networks and long short-term memory networks is constructed and applied for feature learning and prediction. Figure 10 This is a diagram of the long short-term memory neural network structure proposed in this application.

[0142] The time-series data of transformer sensing parameters are organized into a three-dimensional tensor, where the dimensions correspond to the time step, feature dimension, and number of samples, respectively, to meet the input requirements of a two-dimensional convolutional neural network.

[0143] Furthermore, the above three-dimensional tensor is input into a two-dimensional convolutional neural network, and the following operations are performed:

[0144] Convolution operation: The convolution kernel slides on a two-dimensional plane consisting of the time step and the feature dimension to extract local feature combination patterns and generate feature maps.

[0145] Nonlinear activation: Applying the ReLU activation function to the convolution output introduces a nonlinear transformation to learn complex relationships.

[0146] Pooling operation: Perform pooling operation on the activated feature map to downsample, reduce computational complexity and retain key information.

[0147] Feature abstraction: By stacking multiple layers of convolution, activation and pooling operations, feature representations from basic local features to complex global spatial features are gradually extracted and abstracted.

[0148] Furthermore, the high-order spatial feature sequence output by the aforementioned two-dimensional convolutional neural network is input into an LSTM (Long Short-Term Memory) network for temporal dependency modeling. The LSTM network regulates the information flow through its internal gating mechanisms (including forget gate, input gate, and output gate), learns long-term and short-term temporal dependencies, and outputs the final temporal feature representation.

[0149] The forget gate, input gate, and output gate are as follows:

[0150] (11);

[0151] in: , , These are the state calculation results for the forget gate, input gate, and output gate, respectively. For activation function, , , These are the weight matrices for the forget gate, input gate, and output gate, respectively. , , These are the bias terms for the forget gate, input gate, and output gate, respectively. The final output of the LSTM is determined by the output gates and the cell states, as shown in the following equation:

[0152] (12);

[0153] In the formula: t c W represents the cell state input at time t. c Let b be the input cell state weight matrix. c This is the input cell state bias term, where ⨀ indicates element-wise multiplication.

[0154] When training and predicting the aforementioned Long Short-Term Memory (LSTM) neural network, time-series segments of historical normal operation data are extracted for each type or state of fault to be predicted. Each segment is divided into an input sequence and a corresponding subsequent prediction target sequence. The combined model of the aforementioned two-dimensional convolutional neural network and LSTM is trained using the constructed dataset. During training, the Mean Squared Error Loss (MSE) function is used as the optimization objective, and the Adam optimizer is used for parameter optimization. The trained model is used to process new time-series data, where the output of the LSTM layer is a multi-dimensional time-series feature vector representing the evolution of the device state.

[0155] The process involves extracting key feature data from feature prediction data using a feature selection algorithm, and then applying the fault identification model to perform fault identification processing on the key feature data, including:

[0156] Based on the feature selection algorithm, a preset number of feature data with a correlation greater than a preset value are selected from the feature prediction data, and the fault identification model is called to perform fault identification processing on the preset number of feature data.

[0157] Specifically, two feature selection methods, stepwise regression and robust regression, can be used to determine the three most significant features from the DGA input data. Figure 11 This is a flowchart illustrating the overall process of fault prediction using the hybrid feature selection-artificial intelligence-gravity search algorithm proposed in this application. The calculated probabilities or p-values ​​for each gas are obtained using these two methods. These p-values ​​are used to determine the three most critical gases by selecting the three smallest p-values. A regression method is used to determine the optimal quantity and combination of critical input gas components without an iterative process. The statistical characteristics of the data are analyzed using the p-values ​​of partial f-statistics, the formula for which the partial f-statistic is calculated is:

[0158] (13);

[0159] Where SSj and SSk represent the sum of squared residuals of model j and model k, respectively, tj and tk represent the number of parameters in the corresponding models, dfk represents the residual degrees of freedom of model k, and Fdfj−dfk, dfk is a partial F statistic.

[0160] Robust regression is used to limit the influence of outliers to achieve consistent results. It is used to analyze the relationship between one or more inputs and outputs of DGA data. m-estimation is the most common method for calculating p-values ​​in robust regression; the m-estimator transformation can convert a linear model into the following form:

[0161] (14);

[0162] Where n is the number of samples, ei is the residual of the i-th sample, yi is the output observation of the i-th sample, and xi is the input feature vector of the i-th sample. Let xi be the transpose of xi, b be the regression coefficient vector, and p(⋅) be the loss function in robust regression.

[0163] Furthermore, a quality score characterizing the model's performance can be determined using the fitness function in the gravity search algorithm, and the model's parameters can be adjusted based on the quality score.

[0164] Specifically, refer to Figure 12As shown, the agent's quality, reflecting its performance, is calculated using the fitness function in GSA. The fitness function is set as the percentage of correctly identified transformer fault types. The solution to the problem is given by the location of each driver. An efficient driver corresponds to a heavier quality. The model parameters are optimized by updating the rate and adjusting the location accordingly.

[0165] The above describes a transformer condition monitoring method provided by the embodiments of this application. The following describes the apparatus for performing the above transformer condition monitoring method.

[0166] Please see Figure 13 , Figure 13 This is a schematic diagram of the structure of a transformer condition monitoring device provided in an embodiment of this application. Figure 13 As shown, the transformer condition monitoring device includes:

[0167] The data preprocessing module 1301 is used to perform data normalization, outlier detection and missing value processing on the acquired raw transformer data in sequence to obtain preprocessed data of the raw transformer data.

[0168] The correlation analysis module 1302 is used to analyze the correlation between various features in the preprocessed data and divide the preprocessed data according to the analysis results to obtain feature data characterizing the changes in transformer state.

[0169] Data prediction module 1303 is used to call the prediction model to perform feature extraction and prediction processing on feature data, and obtain feature prediction data within a preset time period in the future; and,

[0170] The fault prediction module 1304 is used to extract key feature data from the feature prediction data according to the feature selection algorithm, and call the fault identification model to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing the artificial neural network model or support vector machine based on the gravity search algorithm.

[0171] In one possible implementation, the data preprocessing module 1301 performs data normalization, outlier detection, and missing value processing sequentially on the acquired raw transformer data to obtain preprocessed data of the raw transformer data, including:

[0172] The original transformer data is normalized using the standard deviation standardization algorithm;

[0173] The isolated forest algorithm is used to identify and remove outliers from the original transformer data.

[0174] Missing data is filled in using the completion strategy corresponding to the missing data pattern.

[0175] In one possible implementation, the data preprocessing module 1301 performs outlier identification on the raw transformer data based on the isolated forest algorithm, including:

[0176] Determine the path length of the data point from the root node to the isolated leaf node and the average path length;

[0177] Outlier scores for data points are determined based on path length and the average path length.

[0178] If the outlier score is greater than the outlier score threshold, then the data point is determined to be an outlier.

[0179] In one possible implementation, the data preprocessing module 1301 performs missing value completion processing on the missing data based on the completion strategy corresponding to the missing data pattern, including:

[0180] When the missing pattern is completely random, the corresponding type imputation strategy is determined according to the data distribution characteristics, and the missing value is filled in based on the type imputation strategy.

[0181] When the missing value pattern is random missing, missing value completion is performed based on the least squares regression algorithm or the support vector regression algorithm.

[0182] When the missing value pattern is non-random, the missing values ​​are imputed based on the observed data to obtain multiple imputed datasets. These multiple imputed datasets are then merged to obtain the missing value completion result.

[0183] In one possible implementation, the correlation analysis module 1302 analyzes the correlation between features in the preprocessed data and divides the preprocessed data according to the analysis results to obtain feature data characterizing the transformer state changes, including:

[0184] The correlation coefficient algorithm is used to determine the correlation values ​​that characterize the degree of correlation between parameters in the preprocessed data, and the preprocessed data is divided based on the correlation values.

[0185] In one possible implementation, the data prediction module 1303 calls a prediction model to perform feature extraction and prediction processing on the feature data to obtain the feature prediction data for a preset future time period, including:

[0186] The convolutional network in the prediction model is used to extract local and global features from the feature data to obtain a high-order spatial feature sequence.

[0187] The high-order spatial feature sequences are predicted using the long short-term memory network in the prediction model to obtain feature prediction data.

[0188] In one possible implementation, the fault prediction module 1304 extracts key feature data from the feature prediction data according to the feature selection algorithm, and calls the fault identification model to perform fault identification processing on the key feature data, including:

[0189] Based on the feature selection algorithm, a preset number of feature data with a correlation greater than a preset value are selected from the feature prediction data, and the fault identification model is called to perform fault identification processing on the preset number of feature data.

[0190] In one possible implementation, the optimization process of the fault identification model in the fault prediction module 1304 includes:

[0191] The fitness function in the gravity search algorithm is used to determine the quality score that characterizes the model performance, and the model parameters are adjusted based on the quality score.

[0192] This application also provides an electronic device in its embodiments. (See reference...) Figure 14 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, desktop computers, etc. Figure 14 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0193] like Figure 14 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 1401, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1402 or a program loaded from a storage device 1408 into a random access memory (RAM) 1403. When the electronic device is powered on, the RAM 1403 also stores various programs and data required for the operation of the electronic device. The processing unit 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An input / output (I / O) interface 1405 is also connected to the bus 1404.

[0194] Typically, the following devices can be connected to I / O interface 1405: input devices 1406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 1407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1408 including, for example, memory cards, hard drives, etc.; and communication devices 1409. Communication device 1409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 14Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0195] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the transformer condition monitoring methods provided in this application.

[0196] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the transformer condition monitoring methods provided in this application.

[0197] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0198] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0199] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0200] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A transformer condition monitoring method, characterized in that, include: The acquired raw transformer data is subjected to data normalization, outlier detection, and missing value processing in sequence to obtain preprocessed data of the raw transformer data. The correlation between the features in the preprocessed data is analyzed, and the preprocessed data is divided according to the analysis results to obtain feature data characterizing the changes in the transformer state; The prediction model is invoked to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset time period in the future. Key feature data is extracted from the feature prediction data according to the feature selection algorithm, and the fault identification model is called to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing the artificial neural network model or support vector machine based on the gravity search algorithm.

2. The transformer condition monitoring method according to claim 1, characterized in that, The process of performing data normalization, outlier detection, and missing value processing on the acquired raw transformer data sequentially yields preprocessed data of the raw transformer data, including: The original transformer data is normalized based on the standard deviation normalization algorithm; The isolated forest algorithm is used to identify outliers in the original transformer data and remove the outliers. Missing data is filled in using the completion strategy corresponding to the missing data pattern.

3. The transformer condition monitoring method according to claim 2, characterized in that, The outlier identification of the original transformer data based on the isolated forest algorithm includes: Determine the path length of the data point from the root node to the isolated leaf node and the average path length; Based on the path length and the average path length, an outlier score for the data point is determined. If the outlier score is greater than the outlier score threshold, then the value of the data point is determined to be an outlier.

4. The transformer condition monitoring method according to claim 2, characterized in that, The missing data completion strategy based on the missing data pattern performs missing value completion processing, including: When the missing pattern is completely random, a corresponding type imputation strategy is determined based on the data distribution characteristics, and missing value completion is performed based on the type imputation strategy. When the missing pattern is random missing, missing value completion is performed based on the least squares regression algorithm or the support vector regression algorithm. When the missing pattern is non-random missing, the missing values ​​are imputed based on the observed data to obtain multiple imputed datasets, and the multiple imputed datasets are merged to obtain the missing value completion result.

5. The transformer condition monitoring method according to any one of claims 1 to 4, characterized in that, The process involves analyzing the correlations among the features in the preprocessed data and dividing the preprocessed data according to the analysis results to obtain feature data characterizing the transformer's state changes, including: The correlation coefficient algorithm is used to determine the correlation values ​​that characterize the degree of correlation between the parameters in the preprocessed data, and the preprocessed data is divided based on the correlation values.

6. The transformer condition monitoring method according to any one of claims 1 to 4, characterized in that, The invocation of the prediction model to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset time period in the future includes: Based on the convolutional network in the prediction model, local and global features are extracted from the feature data to obtain a high-order spatial feature sequence. The high-order spatial feature sequence is predicted based on the long short-term memory network in the prediction model to obtain the feature prediction data.

7. The transformer condition monitoring method according to any one of claims 1 to 4, characterized in that, The step of extracting key feature data from the feature prediction data according to the feature selection algorithm and calling the fault identification model to perform fault identification processing on the key feature data includes: Based on the feature selection algorithm, a preset number of feature data with a correlation greater than a preset value are selected from the feature prediction data, and the fault identification model is called to perform fault identification processing on the preset number of feature data.

8. The transformer condition monitoring method according to any one of claims 1 to 4, characterized in that, The optimization process of the fault identification model includes: The fitness function in the gravity search algorithm is used to determine the quality score that characterizes the model performance, and the parameters of the model are adjusted based on the quality score.

9. A transformer condition monitoring device, characterized in that, include: The data preprocessing module is used to perform data normalization, outlier detection, and missing value processing on the acquired raw transformer data in sequence to obtain the preprocessed data of the raw transformer data. The correlation analysis module is used to analyze the correlation between the features in the preprocessed data, and divide the preprocessed data according to the analysis results to obtain feature data characterizing the changes in the transformer state. The data prediction module is used to call the prediction model to perform feature extraction and prediction processing on the feature data to obtain feature prediction data within a preset time period in the future. as well as, The fault prediction module is used to extract key feature data from the feature prediction data according to the feature selection algorithm, and call the fault identification model to perform fault identification processing on the key feature data. The fault identification model is obtained by optimizing an artificial neural network model or support vector machine based on the gravity search algorithm.

10. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the transformer condition monitoring method as described in any one of claims 1 to 8.

Images