Power equipment data anomaly detection method and system based on LSTM-COF

Abstract

The invention discloses an LSTM-COF-based power equipment data anomaly detection method and system, and relates to the technical field of power equipment state monitoring, and the method comprises the following steps: collecting historical data, and constructing a three-dimensional data matrix; predicting equipment parameters at the extreme temperature through an LSTM model; compressing the features, quantifying the covariance deviation degree between the features in combination with a correlation abnormal factor algorithm, and detecting abnormal points; the system integrates a data collection module, a data preprocessing module, a data prediction module, a detection model generation module and a visualization module. According to the method, the multi-dimensional historical operation data of the power equipment is collected, the equipment parameters in the extreme temperature environment are predicted by using the LSTM model, the principal component analysis dimensionality reduction and correlation abnormal factor algorithms are combined, insulation degradation type and electrical connection type faults can be dynamically identified, the data distribution change is adapted through the incremental learning mechanism, and the fault diagnosis accuracy is improved. The problems of low high-dimensional data processing efficiency and poor anomaly detection adaptability due to manual experience dependence in a traditional method are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring, and in particular to a method and system for detecting anomaly in power equipment data based on LSTM-COF. Background Art

[0002] Power equipment plays an important role in the national power grid. Traditional power equipment fault detection mainly relies on manual experience or basic mathematical models, which have some defects: empirical judgment is easily affected by subjective factors, the rules lack universality, and it is difficult to adapt to complex and changeable equipment operating conditions; the characteristic data of power equipment is high-dimensional (such as load, insulation resistance, polarization index, etc.), and traditional mathematical models are difficult to model efficiently, resulting in computational redundancy and insufficient accuracy; existing density-based methods such as the LOF algorithm can identify local outliers, but they tend to ignore the correlation between features in high-dimensional data, resulting in an increased misjudgment rate, especially when the dynamic data features are complex in extreme temperature environments.

[0003] In recent years, some studies have attempted to improve traditional methods by using neural networks for data prediction. However, these methods are not combined with effective dimensionality reduction methods. High-dimensional data is directly input into the model, which easily introduces noise and affects the prediction stability. In addition, the detection algorithm is single and only relies on a single anomaly detection model. The impact of inter-feature correlation on anomaly judgment is not considered, resulting in insufficient detection efficiency and robustness.

[0004] Based on this, a method and system for detecting anomaly in power equipment data based on LSTM-COF are now provided, which can eliminate the drawbacks of the existing technical solutions. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for detecting anomaly in power equipment data based on LSTM-COF, so as to solve the problems in traditional methods such as reliance on manual experience, low efficiency in high-dimensional data processing, and poor adaptability of anomaly detection.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for detecting anomaly in power equipment data based on LSTM-COF specifically includes the following steps: S1. Collect historical operating data of power equipment at room temperature as a basis. The historical operating data includes nine parameters: load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, furfural content, core grounding current, core insulation resistance, and winding DC resistance. Generate time series data at 20-minute intervals, eliminate dimensional differences through standardization, and uniformly intercept 200 time nodes to construct a three-dimensional data matrix. S2. Input the data preprocessed in step S1 into the LSTM model for time series modeling. The input is the standardized normal temperature data, and the output is the predicted values ​​of the parameters under high and low temperature environments. The parameters of the output predicted values ​​include the core grounding current and the winding resistance. S3. Perform secondary normalization on the prediction data and perform principal component analysis dimensionality reduction on the high-dimensional prediction data output by the model to compress the 9-dimensional features into 2-dimensional features, retaining the parameter combinations that are strongly correlated with temperature-sensitive faults; S4. Based on the data after the dimensionality reduction operation in step S3, a correlation anomaly factor algorithm is introduced to quantify the covariance deviation between the features after dimensionality reduction, and anomaly detection is performed on the predicted data after dimensionality reduction. Samples with significant local density deviation are determined as outliers; S5. Output the test results and perform a visual display operation. The test results include a prediction curve graph, a COF value distribution graph, and a two-dimensional feature scatter plot. Power system maintenance personnel analyze the test results and make decisions.

[0007] Preferably, the step S1 specifically includes: S11. Collect historical operating data of the transformer, using load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, furfural content, core grounding current, core insulation resistance, and winding DC resistance as detection features. The historical operating data of the transformer retains timestamp information. S12, normalizing the transformer historical operation data in step S11 to the interval [0, 1] by using a normalization formula to eliminate interference caused by different dimensions on model training; S13. Perform a time alignment operation, uniformly intercept 200 time nodes, and construct a three-dimensional data matrix using a three-dimensional data matrix formula.

[0008] Preferably, the LSTM model in step S2 includes a multi-layer LSTM neural network with more than two hidden layers, wherein the hidden layer is a network layer composed of multiple LSTM units, and any of the LSTM units includes a forget gate, an input gate, a cell state update and an output gate. The LSTM model adopts a sliding window structure, and the input data of the LSTM model is the multidimensional state data of the preprocessed power equipment, and the output data of the LSTM model is the predicted future state data.

[0009] Preferably, in step S2, the LSTM model optimizes the training model parameters through a loss function, and performs back propagation and parameter update through the adaptive gradient optimization algorithm Adam, and the loss function adopts the mean square error between the predicted value and the true value.

[0010] Preferably, the step S3 specifically includes: S31. Focusing and eliminating redundancy on key features of the power equipment, and performing secondary normalization on the power equipment prediction data under extreme temperature environments output by the model to eliminate the impact of temperature dimension differences on the features. The extreme temperatures are below -26°C and above 62°C. S32. Analyze the linear correlation between features through the covariance matrix, use principal component analysis to extract and screen the principal components that are strongly associated with equipment failures, and perform principal component analysis dimensionality reduction on the high-dimensional prediction data output by the model; S33. Filter the principal components based on the cumulative contribution rate threshold, where the threshold is ≥50%, compress the 9-dimensional original features into 2-dimensional feature data, and set the 2-dimensional feature data after dimensionality reduction as PC1 and PC2. The 2-dimensional feature data is classified and stored according to the device number, timestamp, and ambient temperature label, retaining parameter combinations that are strongly correlated with temperature-sensitive faults; Among them, PC1 is the feature related to the degradation of insulation material performance, including dielectric loss and furfural content, and PC2 is the feature related to the stability of electrical connection, including winding DC resistance and partial discharge.

[0011] Preferably, the step S4 specifically includes: S41. Based on the data after the dimensionality reduction operation in step S3, the local reachability density of any sample point is calculated, and the correlation deviation between the power equipment features is quantified using the correlation anomaly factor algorithm; S42. Using the bagging ensemble strategy, we generate several sub-datasets from the original dimensionality-reduced data through the bootstrap sampling method. We independently train the COF base learner on each sub-dataset to obtain an anomaly scoring model. We assign weights based on the performance of the base learner on the validation set, and perform weighted fusion to obtain the final anomaly score. S43. Based on historical data statistics, set an abnormality determination threshold, compare the abnormality determination threshold with the above abnormality score, determine whether the corresponding sample point is an abnormal point, and classify the abnormal point according to the physical meaning of the principal component; S44. Design an incremental learning mechanism that adapts to the dynamic changes in power equipment data. Regularly input newly collected equipment data into the model, use mini-batch gradient descent to update the base learner parameters, and dynamically adjust the threshold according to the distribution of recent detection results. The calculation formula of the local reachable density is: ; The calculation formula of the correlation deviation is: ; in, is the sample point, is a sample point in the neighborhood, For sample points The average local reachability density of the neighborhood is is a sample point in the neighborhood The average local reachability density of the neighborhood is For sample points The distance to the kth nearest neighbor, is the k-distance neighborhood, represents the number of sample points in the k-distance neighborhood, Includes all distances not exceeding The sample points, For sample points The correlation deviation of , For sample points and sample points in the neighborhood The reachable distance, For sample points and sample points in the neighborhood actual distance.

[0012] Preferably, the abnormality determination threshold in step S43 is set to If the final anomaly score is greater than the anomaly determination threshold, the sample point is determined to be an anomaly point. The classification condition of the anomaly point is: When satisfied If PC1 is abnormal, it is determined to be an insulation degradation fault; When satisfied If PC2 is abnormal, it is determined to be an electrical connection fault; When satisfied When both PC1 and PC2 are abnormal, it is determined to be a compound abnormal fault.

[0013] Preferably, the incremental learning mechanism in step S44 is used to adapt to the data distribution changes caused by the grid connection of new energy sources, and the formula used is: ,in, are the new model parameters, are the old model parameters, is the learning rate, , is the loss function, Enter data for batches.

[0014] A power equipment data anomaly detection system based on LSTM-COF is applied to a power equipment data anomaly detection method based on LSTM-COF, including: The data collection module executes step S1 to collect historical data of the power equipment under normal temperature operation and collect multi-source heterogeneous data of the power equipment in real time; The data preprocessing module executes step S1 to preprocess the various data collected by the data collection module, perform standardization, time alignment and noise repair operations, and construct a three-dimensional data matrix; The data prediction module executes step S2 to convert the processed data into a time series format and construct an anomaly detection dataset to predict the parameter sequence and dynamic operating status of the power equipment under extreme temperatures based on the LSTM model; The detection model generation module executes steps S3 and S4 to extract core features from power equipment parameters through principal component analysis, reduce data dimensions and eliminate redundant information, combine the correlation anomaly factor algorithm, quantify the degree of collaborative anomaly between features, and train the integrated detection model; The visualization module executes step S5 to convert the detection results into an intuitive graphical interface.

[0015] Preferably, it also includes a memory and a processor, the memory is used to store a program for implementing the LSTM-COF-based power equipment data anomaly detection method, and the processor is used to execute the program for implementing the LSTM-COF-based power equipment data anomaly detection method to implement the steps of the LSTM-COF-based power equipment data anomaly detection method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: In this LSTM-COF-based power equipment data anomaly detection method, the LSTM model is used to capture the time series dynamic characteristics of power equipment parameters, predict the operating status under extreme temperatures, and provide a reliable data basis for anomaly detection. Principal component analysis is used to compress 9-dimensional features into 2-dimensional features, retaining parameter combinations that are strongly correlated with temperature-sensitive faults, significantly reducing computational redundancy and improving processing efficiency. The correlation anomaly factor algorithm is introduced to quantify the covariance deviation of the features after dimensionality reduction, and the bagging integration strategy is combined to improve the robustness of the model and accurately identify anomalies with local density deviation. A dynamic threshold mechanism and an incremental learning mechanism are provided to adapt to the data distribution changes caused by the grid connection of new energy, facilitate the optimization of detection performance, and reduce manual intervention. A power equipment data anomaly detection system is provided to facilitate the generation of prediction curves, COF distribution diagrams, and two-dimensional scatter plots through a visualization module, intuitively displaying anomalies and their classification, assisting maintenance personnel in making quick decisions, reducing manual inspection costs, and improving fault location efficiency under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the present invention.

[0018] Figure 2 Schematic diagram of the process of step S1 of the present invention.

[0019] Figure 3 Schematic diagram of the process of step S3 of the present invention.

[0020] Figure 4Schematic diagram of the process of step S4 of the present invention.

[0021] Figure 5 Schematic diagram of the system structure of the present invention.

[0022] Figure 6 Schematic diagram of the interaction between the data prediction module and the detection model generation module of the present invention.

[0023] Notes on the accompanying drawings: data collection module 10, data preprocessing module 20, data prediction module 30, detection model generation module 40, visualization module 50, memory 60, processor 70. DETAILED DESCRIPTION

[0024] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0025] In this embodiment, if Figures 1-6 As shown, a method for detecting anomaly in power equipment data based on LSTM-COF specifically includes the following steps: S1. Collect historical operating data of power equipment at room temperature as a basis. The historical operating data includes nine parameters: load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, furfural content, core grounding current, core insulation resistance, and winding DC resistance. Generate time series data at 20-minute intervals, eliminate dimensional differences through standardization, and uniformly intercept 200 time nodes to construct a three-dimensional data matrix. Specifically, the historical operation data of the transformer is collected, and the load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, aldehyde content, core grounding current, core insulation resistance, and winding DC resistance are used as detection features. The historical operation data of the transformer are all retained with timestamp information. The historical operation data of the power equipment (such as transformers and circuit breakers) under normal temperature environment are collected by the data collection module 10 as the basis. The 9 types of key parameters generate time series data at intervals of 20 minutes, and the ambient temperature and humidity data are recorded synchronously. The surface temperature distribution map of the equipment is obtained by the infrared thermal imager. The data storage adopts a time series database, which supports real-time streaming access and batch import. The data preprocessing module 20 is used according to the normalization formula Normalized to the interval [0,1], is a characteristic value in the original data, is the minimum value of this feature, is the maximum value of this feature, The normalized value is in the range of [0,1], eliminating the dimension difference of the time series data of power equipment, uniformly intercepting 200 time nodes, ensuring the consistency of data length, and constructing a three-dimensional data matrix (number of devices × time × features), solve the problem of inconsistent sensor sampling frequencies, use cubic spline interpolation for missing data, detect and eliminate outliers using the isolation forest algorithm, unify data formats, and improve processing efficiency; S2. Input the data preprocessed in step S1 into the LSTM model for time series modeling. The input is the standardized normal temperature data, and the output is the predicted values ​​of the parameters under high and low temperature environments. The parameters of the output predicted values ​​include the core grounding current and the winding resistance. Specifically, the LSTM model contains a multi-layer LSTM neural network with more than two hidden layers. The hidden layer is a network layer composed of multiple LSTM units. Any LSTM unit contains a forget gate, an input gate, a cell state update, and an output gate. The LSTM model uses a sliding window structure and only needs to process local time series data to reduce memory usage. The sliding window size is set to , then the training data set becomes: , input data All are pre-processed multi-dimensional state data of power equipment, where each window As the input data of the LSTM model, the output data is the predicted future state data ; The forget gate is used to control the retention ratio of historical information, and its calculation formula is: ,in is the output of the forget gate, is the weight matrix of the forget gate, is the bias, is the hidden state at the previous moment, is the current input data, is the sigmoid activation function; The input gate is used to determine the storage of the current input information and consists of two parts: 、 ,in 、 are the weight matrices of the input gate and candidate cell states, 、 is the bias term, is the new candidate cell state, is the hyperbolic tangent function, is the input gate output, is the candidate cell state; The cell state update is used to combine the forget gate and the input gate to update the cell state. The calculation formula is: ,in is the cell state at the current moment, is the cell state at the previous moment; The output gate is used to control the output information at the current moment and consists of two parts: 、 ,in is the weight matrix of the output gate, is the bias term, is the output gate output, is the hidden state at the current moment; Final output of the LSTM model Calculated by the fully connected layer: ,in and are all fully connected layer parameters, The predicted state of the power equipment at the next moment; The LSTM model captures the temporal correlation between temperature and parameters, simulating this dynamic response process and providing high-reliability data support for anomaly detection. Based on preprocessed data, this process constructs a long-short-term memory network model to simulate the dynamic behavior of power equipment under extreme temperatures. The model has 32 neurons in the hidden layer, uses a Sigmoid activation function, and an Adam optimizer. It inputs standardized normal-temperature data and outputs sequences of key parameters under high and low-temperature environments, such as core grounding current and winding DC resistance. The LSTM model uses the synergistic effect of forget gates and input gates to capture the nonlinear effects of sudden temperature changes on insulation material aging, such as the tendency of high temperatures to accelerate dielectric loss degradation of insulating oil. During training, the data is divided into training, validation, and test sets, and model parameters are optimized through continuous iteration. During model training, the first 40 sets of equipment data serve as the training set, and the last 10 sets serve as the validation set. The training process can be summarized into the following steps: first, a sliding window is constructed on the multi-dimensional time series data of the power equipment to form input samples with a time series structure. Then, a neural network model containing multiple LSTM layers is constructed. The long-term dependency relationship in the time series is modeled through nonlinear transformations of the forget gate, input gate, and output gate. The mean square error is used as the loss function, and the training model parameters are optimized through the loss function. The formula is: ,in is the number of training samples, For the The true value of the sample, For the The LSTM prediction value of samples is calculated, where L is the mean square error, and back propagation and parameter update are performed through the adaptive gradient optimization algorithm Adam. The formula is: , where t is the number of iterations, is the current gradient, and is the first-order moment estimate of the gradient, and is the gradient second moment estimate, 、 is the momentum hyperparameter (set to , ), is the learning rate, are the updated model parameters, are the model parameters before updating, The denominator is kept to a minimum to prevent it from being zero, thereby minimizing the prediction error. Ultimately, an LSTM model that can accurately predict the status of power equipment is obtained, providing high-quality prediction data support for subsequent dimensionality reduction and anomaly detection. S3. Perform secondary normalization on the prediction data and perform principal component analysis dimensionality reduction on the high-dimensional prediction data output by the model to compress the 9-dimensional features into 2-dimensional features, retaining the parameter combinations that are strongly correlated with temperature-sensitive faults; Specifically, the key features of the power equipment are focused and redundantly eliminated. The prediction data of the power equipment under the extreme temperature environment output by the model is secondary standardized to eliminate the influence of the temperature dimension difference on the features. The extreme temperatures are below -26℃ and above 62℃. The mean of each characteristic parameter is calculated. and variance Among them, the corresponding 9 types of characteristics such as load and partial discharge are calculated according to the formula Standardize the data to ensure that each feature follows a distribution with a mean of 0 and a variance of 1, where For the Samples, The standardized value of the feature, The original data Samples, The original value of the feature, For the The mean of the features, For the The variance of the feature; Through the covariance matrix formula Analyze the linear correlation between features, Indicates the summation of the elements of a 9×9 real matrix set, whose elements For the variables and The covariance element corresponding to the variables reflects the degree of linear correlation between the two variables. is the number of samples, For the In the sample The standardized value of the variable, For the In the sample The standardized value of the variable, As the degree of freedom correction term, principal component analysis (PCA) is used to extract and screen the principal components that are strongly associated with equipment failures, and solve the characteristic equation of the covariance matrix , get the eigenvalue and the corresponding orthogonal unit eigenvectors is the characteristic value; Perform principal component analysis and dimensionality reduction on the high-dimensional prediction data output by the model, and select the principal components according to the cumulative contribution rate threshold. The calculation formula for the cumulative contribution rate is: , The number of principal components is used to compress the 9-dimensional original features into 2-dimensional feature data. The 2-dimensional feature data after dimensionality reduction is set as PC1 and PC2. The 2-dimensional feature data is classified and stored according to the device number, timestamp, and ambient temperature label, and the parameter combination that is strongly correlated with the temperature-sensitive fault is retained. In this method, strong correlation refers to a parameter combination that is highly correlated with the temperature-sensitive fault and can significantly reflect the impact of temperature changes on the device status. When the temperature changes, the change in the parameter combination can more accurately reflect whether the device has a temperature-sensitive fault. The parameter combination includes features related to the degradation of the insulation material performance and features related to the stability of the electrical connection, specifically including the dielectric loss furfural content. , winding DC resistance and partial discharge, dielectric loss and furfural content are closely related to the performance of the insulation material. When the temperature rises, the insulation material will accelerate aging, resulting in increased dielectric loss and changes in furfural content. Changes in the two parameters of dielectric loss and furfural content can directly reflect the degree of degradation of the insulation material and are strongly correlated with temperature-sensitive insulation degradation faults. Temperature changes will affect the contact resistance of the electrical connection, causing changes in the winding DC resistance. Abnormal temperature may also trigger partial discharge. Changes in the two parameters of winding DC resistance and partial discharge can reflect the fault condition of the electrical connection and are strongly correlated with temperature-sensitive electrical connection faults. Assuming that the cumulative contribution rate threshold is ≥50%, the first two principal components (PC1 and PC2) are selected, and the expression is: , ,in is the eigenvector component, is the original feature, , ,PC1 is a feature related to the degradation of insulation material performance, including dielectric loss and furfural content, and PC2 is a feature related to the stability of electrical connections, including winding DC resistance and partial discharge. The two-dimensional data after dimensionality reduction can reduce computational redundancy and directly associate the actual fault type through physical meaning mapping. For example, an abnormal increase in PC1 indicates the risk of insulating oil aging, and an abnormal fluctuation in PC2 indicates poor winding contact; S4. Based on the data after the dimensionality reduction operation in step S3, a correlation anomaly factor algorithm is introduced to quantify the covariance deviation between the features after dimensionality reduction, and anomaly detection is performed on the predicted data after dimensionality reduction. Samples with significant local density deviation are determined as outliers; Specifically, assuming that any sample point is sample point o, based on the above dimensionality reduction data, the formula is: Calculate the local reachability density of the sample point o to characterize the compactness of the data distribution in its neighborhood, where is the sample point, is a sample point in the neighborhood, For sample points The average local reachability density of the neighborhood is is a sample point in the neighborhood The average local reachability density of the neighborhood is For sample points The distance to the kth nearest neighbor, is the k-distance neighborhood, represents the number of sample points in the k-distance neighborhood, Includes all distances not exceeding The sample points, For sample points The correlation deviation of , For sample points and sample points in the neighborhood The reachable distance, For sample points and sample points in the neighborhood The actual distance is calculated by the correlation anomaly factor (COF) algorithm Quantify the correlation deviation between power equipment characteristics; The bagging ensemble strategy is used. The bagging ensemble strategy is a method to improve model performance by building multiple base learners and combining prediction results. It is used to improve the robustness and accuracy of the model. Specifically, it includes: resampling to generate sub-datasets, generating 10 sub-datasets from the original dimensionality reduction data through the bootstrap sampling method, and independently training the COF base learner on each sub-dataset to obtain 10 groups of anomaly scoring models. , assign weights based on the performance of the base learner on the validation set , weighted fusion to obtain the final anomaly score , where the weight The accuracy of the base learner Dynamic calculation: separate the sample set with real abnormal labels from the historical data (such as 100 sets of transformer data with known faults) to build a validation set, and each base learner calculates the accuracy on the validation set ,in is the number of abnormal samples correctly identified, is the number of correctly identified normal samples, is the number of abnormal samples that are misidentified, is the number of normal samples that are misidentified, For the The weight of a base learner is: , For the The base learner pairs of sample points The correlation anomaly factor score, Indicates the representative The accuracy calculated by the base learner on the validation set is calculated based on historical data statistics, and the abnormality judgment threshold is set. Set as , compare the anomaly judgment threshold with the above anomaly score to determine whether the corresponding sample point is an anomaly point. If the final anomaly score is greater than the anomaly judgment threshold, the sample point is determined to be an anomaly point. The anomaly point is classified according to the physical meaning of the principal component. The classification conditions of the anomaly point are: When satisfied If PC1 is abnormal, it is determined to be an insulation degradation fault; When satisfied If PC2 is abnormal, it is determined to be an electrical connection fault; When satisfied If both PC1 and PC2 are abnormal, it is determined to be a compound abnormal fault; Ten COF-based learners were trained using a bagging strategy to dynamically weight differences between power equipment batches (e.g., different transformer models). This improved the robustness of detecting concurrent faults such as core multi-point grounding and winding overheating. The physical meaning of the principal components was combined to further classify the fault type. For example, a high COF value accompanied by an abnormal PC1 value indicates core multi-point grounding, while an abnormal PC2 value indicates winding overheating. An incremental learning mechanism is designed to adapt to the dynamic changes in power equipment data. The incremental learning mechanism is used to adapt to the changes in data distribution caused by the grid connection of new energy sources. The formula used is: ,in, are the new model parameters, are the old model parameters, is the learning rate, , is the loss function, To input data in batches, the newly collected device data is regularly input into the model, and the base learner parameters are updated using the mini-batch gradient descent method. The threshold is dynamically adjusted according to the distribution of recent detection results. The formula is: ,in is the actual abnormality rate, is the expected abnormality rate, is the adjustment coefficient, is the adjusted threshold, is the threshold before adjustment; S5. Output the test results and perform a visual display operation. The test results include a prediction curve graph, a COF value distribution graph, and a two-dimensional feature scatter plot. Power system maintenance personnel analyze the test results and make decisions; Specifically, based on the two-dimensional principal component data after dimensionality reduction, dynamic visualization detection results are generated in high and low temperature environments. This can intuitively distinguish between the normal and abnormal states of power equipment. The data distribution probability density is calculated through the kernel density estimation algorithm, and the envelope boundary is drawn. Normal data is concentrated in the blue envelope, and abnormal points are located outside the red envelope. The abnormal equipment numbers are marked. Combined with the physical characteristics of the equipment, the shape of the envelope surface dynamically reflects the impact of temperature on faults such as insulation material aging and winding resistance imbalance. The details of abnormal parameter deviations are displayed in real time, which helps auxiliary power system maintenance personnel quickly locate the source of the fault. Step S5 further includes diagnostic report generation and output: Based on the mapping of COF test results and the physical meaning of the principal components, a structured diagnostic report is automatically generated. The report content covers the abnormal equipment number, fault type classification, key parameter deviation values ​​and maintenance recommendations. The report supports PDF and SCADA system interface protocols and is transmitted to the operation and maintenance terminal through encryption to ensure data security. The test results can be visualized and multi-dimensionally diagnosed, significantly reducing manual inspection costs and improving fault location efficiency under complex working conditions. It is particularly suitable for batch monitoring and intelligent operation and maintenance of equipment such as transformers and circuit breakers in new energy grid-connected scenarios; Among them Figure 5 As shown, a power equipment data anomaly detection system based on LSTM-COF is applied to a power equipment data anomaly detection method based on LSTM-COF, including: The data collection module 10 is used to collect historical data of power equipment under normal temperature operation, and to collect multi-source heterogeneous data in real time during power equipment operation, including physical parameters such as load, partial discharge, insulating oil dielectric loss, and core grounding current of key equipment such as transformers and circuit breakers. This avoids the subjectivity and reliance on experience of manual inspections, and outputs raw data with timestamps for use by the data preprocessing module 20. The data preprocessing module 20 is used to preprocess the various data collected by the data collection module 10, perform standardization, time alignment, and noise repair operations, and construct a three-dimensional data matrix. The data is scaled to the interval [0, 1] using a normalization formula, 200 time nodes are uniformly intercepted, missing values ​​are filled, and outliers are detected and eliminated using the isolation forest algorithm. This solves the problem of inconsistent sampling frequencies of multi-source data and eliminates the need for manual intervention. The output is a standardized three-dimensional matrix; The data prediction module 30 is used to convert the processed data into a time series format and construct an anomaly detection dataset. Based on the LSTM model, the module predicts the parameter sequence and dynamic operating status of power equipment at extreme temperatures. For example, the model inputs the nonlinear trend of insulation aging at room temperature and outputs the predicted values ​​of parameters at extreme temperatures, such as core grounding current and winding resistance. The module uses a mean square error loss function and an Adam optimizer during training to ensure prediction accuracy, provide a reliable data foundation for anomaly detection, and automatically learn the time series dynamic characteristics of historical data without the need for manual setting of rules or thresholds. The detection model generation module 40 is used to extract core features from power equipment parameters through principal component analysis, reduce data dimensions and eliminate redundant information, combine the correlation anomaly factor algorithm to quantify the degree of collaborative anomaly between features, train an integrated detection model, and implement anomaly detection through PCA dimensionality reduction and the COF algorithm. PCA compresses 9-dimensional features into 2-dimensional features, retaining principal components with a cumulative contribution rate of ≥50%. The COF algorithm automatically quantifies the covariance deviation between features, determines anomalies, and outputs classification results, reducing manual intervention. The visualization module 50 is used to convert the test results into an intuitive graphical interface, generate a data envelope distribution map under high and low temperature environments, mark the abnormal equipment number and the deviation trend of key parameters, and display the equipment health status through interactive charts to facilitate decision-making by operation and maintenance personnel; Among them Figure 5 As shown, the system further includes a memory 60 and a processor 70. The memory 60 is used to store a program for implementing the method for detecting anomaly data of power equipment based on LSTM-COF, and the processor 70 is used to execute the program for implementing the method for detecting anomaly data of power equipment based on LSTM-COF, so as to implement the steps of the method for detecting anomaly data of power equipment based on LSTM-COF. The processor 70 pushes data to a visual display interface, supports historical data query, and implements subsequent analysis operations. Specifically, the system also includes collection equipment, background management equipment and user terminals. The above modules are all arranged inside the background management equipment. The background management equipment has a built-in database, media interface, storage and actuator. The background management equipment is equipped with a server with an encryption chip, which can be a desktop computer, laptop computer, server and other electronic devices. It can be an independent proxy server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as big data and artificial intelligence platforms and other basic cloud computing services. This server is an electronic device used to provide background services.

[0026] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting anomaly in power equipment data based on LSTM-COF, characterized in that: The specific steps include: S1. Collect historical operating data of power equipment at room temperature as a basis. The historical operating data includes nine parameters: load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, furfural content, core grounding current, core insulation resistance, and winding DC resistance. Generate time series data at 20-minute intervals, eliminate dimensional differences through standardization, and uniformly intercept 200 time nodes to construct a three-dimensional data matrix. S2. Input the data preprocessed in step S1 into the LSTM model for time series modeling. The input is the standardized normal temperature data, and the output is the predicted values ​​of the parameters under high and low temperature environments. The parameters of the output predicted values ​​include the core grounding current and the winding resistance. S3. Perform secondary normalization on the prediction data and perform principal component analysis dimensionality reduction on the high-dimensional prediction data output by the model to compress the 9-dimensional features into 2-dimensional features, retaining the parameter combinations that are strongly correlated with temperature-sensitive faults; S4. Based on the data after the dimensionality reduction operation in step S3, a correlation anomaly factor algorithm is introduced to quantify the covariance deviation between the features after dimensionality reduction, and anomaly detection is performed on the predicted data after dimensionality reduction. Samples with significant local density deviation are determined as outliers; S5. Output the test results and perform a visual display operation. The test results include a prediction curve graph, a COF value distribution graph, and a two-dimensional feature scatter plot. Power system maintenance personnel analyze the test results and make decisions.

2. The method for detecting anomaly of power equipment data based on LSTM-COF according to claim 1, characterized in that: The step S1 specifically includes: S11. Collect historical operating data of the transformer, using load, partial discharge, insulating oil dielectric loss, polarization index, volume resistivity, furfural content, core grounding current, core insulation resistance, and winding DC resistance as detection features. The historical operating data of the transformer retains timestamp information. S12, normalizing the transformer historical operation data in step S11 to the interval [0, 1] by using a normalization formula to eliminate interference caused by different dimensions on model training; S13. Perform a time alignment operation, uniformly intercept 200 time nodes, and construct a three-dimensional data matrix using a three-dimensional data matrix formula.

3. The method for detecting anomaly of power equipment data based on LSTM-COF according to claim 1, characterized in that: The LSTM model in step S2 includes a multi-layer LSTM neural network with more than two hidden layers, wherein the hidden layer is a network layer composed of multiple LSTM units, and any of the LSTM units includes a forget gate, an input gate, a cell state update, and an output gate. The LSTM model adopts a sliding window structure, and the input data of the LSTM model is the preprocessed multidimensional state data of the power equipment, and the output data of the LSTM model is the predicted future state data.

4. The method for detecting anomaly in power equipment data based on LSTM-COF according to claim 3 is characterized in that: In step S2, the LSTM model optimizes the training model parameters through the loss function, and performs back propagation and parameter update through the adaptive gradient optimization algorithm Adam, wherein the loss function adopts the mean square error between the predicted value and the true value.

5. The method for detecting anomaly of power equipment data based on LSTM-COF according to claim 1, characterized in that: The step S3 specifically includes: S31. Focusing and eliminating redundancy on key features of the power equipment, and performing secondary normalization on the power equipment prediction data under extreme temperature environments output by the model to eliminate the impact of temperature dimension differences on the features. The extreme temperatures are below -26°C and above 62°C. S32. Analyze the linear correlation between features through the covariance matrix, use principal component analysis to extract and screen the principal components that are strongly associated with equipment failures, and perform principal component analysis dimensionality reduction on the high-dimensional prediction data output by the model; S33. Filter the principal components based on the cumulative contribution rate threshold, where the threshold is ≥50%, compress the 9-dimensional original features into 2-dimensional feature data, and set the 2-dimensional feature data after dimensionality reduction as PC1 and PC2. The 2-dimensional feature data is classified and stored according to the device number, timestamp, and ambient temperature label, retaining parameter combinations that are strongly correlated with temperature-sensitive faults; Among them, PC1 is the feature related to the degradation of insulation material performance, including dielectric loss and furfural content, and PC2 is the feature related to the stability of electrical connection, including winding DC resistance and partial discharge.

6. The method for detecting anomaly in power equipment data based on LSTM-COF according to claim 1, characterized in that: The step S4 specifically includes: S41. Based on the data after the dimensionality reduction operation in step S3, the local reachability density of any sample point is calculated, and the correlation deviation between the power equipment features is quantified using the correlation anomaly factor algorithm; S42. Using the bagging ensemble strategy, we generate several sub-datasets from the original dimensionality-reduced data through the bootstrap sampling method. We independently train the COF base learner on each sub-dataset to obtain an anomaly scoring model. We assign weights based on the performance of the base learner on the validation set, and perform weighted fusion to obtain the final anomaly score. S43. Based on historical data statistics, set an abnormality determination threshold, compare the abnormality determination threshold with the above abnormality score, determine whether the corresponding sample point is an abnormal point, and classify the abnormal point according to the physical meaning of the principal component; S44. Design an incremental learning mechanism that adapts to the dynamic changes in power equipment data. Regularly input newly collected equipment data into the model, use mini-batch gradient descent to update the base learner parameters, and dynamically adjust the threshold according to the distribution of recent detection results. The calculation formula of the local reachable density is: ; The calculation formula of the correlation deviation is: ; in, is the sample point, is a sample point in the neighborhood, For sample points The average local reachability density of the neighborhood is is a sample point in the neighborhood The average local reachability density of the neighborhood is For sample points The distance to the kth nearest neighbor, is the k-distance neighborhood, represents the number of sample points in the k-distance neighborhood, Includes all distances not exceeding The sample points, For sample points The correlation deviation of , For sample points and sample points in the neighborhood The reachable distance, For sample points and sample points in the neighborhood actual distance.

7. The method for detecting anomaly in power equipment data based on LSTM-COF according to claim 6, characterized in that: The abnormality determination threshold in step S43 is set to If the final anomaly score is greater than the anomaly determination threshold, the sample point is determined to be an anomaly point. The classification condition of the anomaly point is: When satisfied If PC1 is abnormal, it is determined to be an insulation degradation fault; When satisfied If PC2 is abnormal, it is determined to be an electrical connection fault; When satisfied When both PC1 and PC2 are abnormal, it is determined to be a compound abnormal fault.

8. The method for detecting anomaly in power equipment data based on LSTM-COF according to claim 6, characterized in that: The incremental learning mechanism in step S44 is used to adapt to the data distribution changes caused by the grid connection of new energy sources, and the formula used is: ,in, are the new model parameters, are the old model parameters, is the learning rate, , is the loss function, Enter data for batches.

9. A power equipment data anomaly detection system based on LSTM-COF, applied to a power equipment data anomaly detection method based on LSTM-COF according to any one of claims 1 to 8, characterized in that: include: The data collection module (10) executes step S1 to collect historical data of the power equipment under normal temperature operation and collect multi-source heterogeneous data of the power equipment in real time; The data preprocessing module (20) executes step S1 to preprocess the various data collected by the data collection module (10), perform standardization, time alignment and noise repair operations, and construct a three-dimensional data matrix; The data prediction module (30) executes step S2 to convert the processed data into a time series format and construct an anomaly detection data set to predict the parameter sequence and dynamic operating state of the power equipment under extreme temperature based on the LSTM model; A detection model generation module (40) executes steps S3-S4 to extract core features from power equipment parameters through principal component analysis, reduce data dimensions and eliminate redundant information, combine with correlation anomaly factor algorithm, quantify the degree of collaborative anomaly between features, and train an integrated detection model; The visualization module (50) executes step S5 to convert the detection results into an intuitive graphical interface.

10. The power equipment data anomaly detection system based on LSTM-COF according to claim 9 is characterized in that: The invention also includes a memory (60) and a processor (70), wherein the memory (60) is used to store a program for implementing the method for detecting anomaly of power equipment data based on LSTM-COF, and the processor (70) is used to execute the program for implementing the method for detecting anomaly of power equipment data based on LSTM-COF, so as to implement the steps of the method for detecting anomaly of power equipment data based on LSTM-COF.

Images