Feature fusion-based optical fiber transformer health state evaluation method and system

Abstract

The invention relates to the technical field of power equipment operation and maintenance and state monitoring, in particular to an optical fiber transformer health state assessment method and system based on feature fusion. According to the invention, sampling signals of a plurality of optical fiber current transformers are collected and preprocessed; constructing a historical data matrix for principal component analysis, establishing a principal component space and calculating a square prediction error control limit value; constructing a simulated fault data set and projecting the simulated fault data set to a principal component space to calculate a square prediction error time sequence; performing wavelet packet decomposition on the square prediction error time sequence to extract a normalized energy feature vector, and training a one-dimensional convolutional neural network model; projecting the real-time sampling signal to a principal component space to judge whether the real-time sampling signal exceeds the limit or not; if yes, the health state level is output through the trained model. According to the invention, accurate grading and real-time online evaluation of the health state of the optical fiber transformer are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power equipment operation and maintenance and state monitoring, in particular to a fiber-optic mutual inductor health state evaluation method and system based on feature fusion. BACKGROUND

[0002] The fiber-optic current transformer is widely used in high-voltage power systems due to its superior insulation performance and fast response capability, but it will gradually deteriorate due to environmental factors and its own aging during long-term operation, resulting in reduced measurement accuracy.

[0003] Traditional health state evaluation mainly relies on periodic offline calibration, which has the defects of long cycle, high cost and inability to monitor in real time. The existing online methods are mostly based on single feature grading, which is difficult to fully reflect the multi-dimensional deterioration characteristics of the equipment, and the processing capability for non-linear and non-stationary signals is insufficient, resulting in low evaluation accuracy. More importantly, when a certain transformer in the same measurement point starts to deteriorate, it is difficult to determine whether the error is caused by power grid fluctuation or equipment deterioration by relying only on single device measurement value, while the analysis of group measurement data can effectively utilize the correlation between transformers and more sensitively find abnormal equipment. SUMMARY

[0004] The present application provides a fiber-optic mutual inductor health state evaluation method and system based on feature fusion, which solves the problems of poor real-time performance, single feature and weak non-linear processing capability of traditional methods.

[0005] To achieve the above purpose, the present application provides the following technical solutions: The present application provides a fiber-optic mutual inductor health state evaluation method based on feature fusion, comprising: S100: Collecting sampling signals of multiple fiber-optic current transformers under the same measurement point, and performing noise reduction and synchronization calibration on the sampling signals; S200: Constructing a historical data matrix of the multiple fiber-optic current transformers under normal operating conditions and performing principal component analysis, establishing a principal component space and calculating a squared prediction error control limit value; S300: Constructing a simulated fault data set by superimposing error signals of different amplitudes in the signal of one of the fiber-optic current transformers, constructing corresponding fault data and corresponding health state grade labels, and projecting the simulated fault data set to the principal component space to calculate a squared prediction error time series; S400: Performing multi-layer wavelet packet decomposition on the squared prediction error time series, calculating the energy values of the reconstructed signals of each sub-band and normalizing to construct a normalized energy feature vector, and training a one-dimensional convolutional neural network model using the normalized energy feature vector and the corresponding health state grade labels; S500: Projecting the real-time sampling signal to the principal component space to calculate a real-time squared prediction error time series, and determining whether the squared prediction error control limit is exceeded; S600: If the control limit is exceeded, performing multi-layer wavelet packet decomposition on the real-time squared prediction error time series and extracting a normalized energy feature vector, and inputting the one-dimensional convolutional neural network model trained to output a health state grade; if not, determining that the state is normal.

[0006] As a preferred technical solution of the present application, the step of establishing the principal component space comprises: standardizing the historical data matrix; calculating the covariance matrix of the standardized historical data matrix and performing eigenvalue decomposition to obtain the eigenvectors in descending order of eigenvalues; determining the number of principal components according to the cumulative variance contribution rate, and selecting the corresponding eigenvectors to form the principal component space.

[0007] As a preferred technical solution of the present application, the step of constructing the simulated fault data set comprises: selecting the signal of one of the fiber-optic current transformers from the historical data matrix; superimposing error signals of different amplitudes in the selected fiber-optic current transformer signal to construct fault data corresponding to four health state grades, and keeping the signals of the remaining fiber-optic current transformers unchanged; According to the amplitude range of the superimposed error, each fault data is labeled with a corresponding health state grade label.

[0008] As a preferred technical solution of the present application, the health state grade label comprises four grades: Grade one corresponds to an error range of 0% to 0.05%; Grade two corresponds to an error range of 0.05% to 0.2%; Grade three corresponds to an error range of 0.2% to 0.5%; Grade four corresponds to an error range of 0.5% or more.

[0009] As a preferred technical solution of the present application, the step of performing multi-layer wavelet packet decomposition on the squared prediction error time series comprises: performing multi-layer wavelet packet decomposition on the squared prediction error time series to obtain sub-band reconstruction signals; calculating the energy value of the layer sub-band reconstruction signal, the energy value being the sum of the squares of the values of each sampling point of the sub-band reconstruction signal; The energy values of the sub-frequency bands are combined into an energy vector, and a normalized energy feature vector is obtained by normalizing the energy vector.

[0010] As a preferred technical solution of the present application, the projecting the real-time sampling signal into the principal component space to calculate the real-time squared prediction error time sequence comprises: The real-time sampling signal is segmented according to a fixed number of data points, and each segment is taken as an evaluation window; The data in each evaluation window is projected into the principal component space, and the squared prediction error statistic corresponding to the window is calculated; The squared prediction error statistics of each evaluation window are combined into a real-time squared prediction error time sequence in time sequence.

[0011] As a preferred technical solution of the present application, the input of the one-dimensional convolutional neural network model is the normalized energy feature vector, the output layer adopts a Softmax activation function to output the probability distribution of four health state grades, and the grade corresponding to the maximum probability is selected as the health state evaluation result.

[0012] The present application also proposes a fiber mutual inductor health state evaluation system based on feature fusion, comprising: A signal preprocessing module is used to collect sampling signals of multiple fiber current transformers at the same measurement point, and to perform noise reduction and synchronous calibration on the sampling signals; A principal component modeling module is used to construct a historical data matrix of the multiple fiber current transformers in a normal operating state and perform principal component analysis, establish a principal component space and calculate a squared prediction error control limit value; A data construction module is used to construct a simulated fault data set by superimposing error signals of different amplitudes in the signal of one of the fiber current transformers, construct corresponding fault data and corresponding health state grade labels, and project the simulated fault data set into the principal component space to calculate a squared prediction error time sequence; A model training module is used to perform multi-layer wavelet packet decomposition on the squared prediction error time sequence, calculate the energy values of the reconstructed signals of each sub-frequency band, normalize the energy values to construct a normalized energy feature vector, and train a one-dimensional convolutional neural network model using the normalized energy feature vector and the corresponding health state grade labels; A real-time monitoring module is used to project a real-time sampling signal into the principal component space to calculate a real-time squared prediction error time sequence, and determine whether the control limit value is exceeded; A state evaluation module is used to perform multi-layer wavelet packet decomposition on the real-time squared prediction error time sequence and extract a normalized energy feature vector if the control limit value is exceeded, input the one-dimensional convolutional neural network model to output a health state grade; and determine that the state is normal if the control limit value is not exceeded.

[0013] The beneficial effects of this invention are: 1. By establishing a group correlation pattern of multiple instrument transformers through principal component analysis and extracting the squared prediction error statistic as a degradation index, this invention effectively separates normal power grid fluctuations from abnormal equipment degradation. Compared with traditional single-unit monitoring methods, this invention significantly improves the sensitivity and accuracy of fault detection by utilizing group redundancy information.

[0014] 2. By integrating principal component analysis, wavelet packet analysis, and convolutional neural networks, an integrated evaluation process of "group feature separation → time-frequency feature extraction → deep learning-based classification" is constructed. This approach can extract multi-dimensional time-frequency energy features from nonlinear, non-stationary squared prediction error signals, breaking through the limitations of traditional methods that rely on single feature evaluation, and achieving accurate and automated classification of four levels corresponding to the error range.

[0015] 3. A one-dimensional convolutional neural network is used for health status classification. The model structure is simple, and the inference time for a single sample can be controlled within 10ms, meeting the requirements for online real-time evaluation of fiber optic transformers. Compared with traditional offline verification methods, this invention achieves continuous monitoring around the clock and immediate anomaly detection, effectively ensuring the accuracy and reliability of power system measurements. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the fiber optic instrument health status assessment method based on feature fusion of the present invention. Figure 2 This is a schematic diagram of the structure of the fiber optic instrument transformer health status assessment system based on feature fusion according to the present invention; Figure 3 This is a schematic diagram of the overall evaluation process of the fiber optic transformer health status assessment system based on feature fusion according to the present invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] Example 1: As Figure 1 As shown, the fiber optic instrument transformer health status assessment method based on feature fusion includes: S100: Collect sampling signals from multiple fiber optic current transformers at the same measurement point, and perform noise reduction and synchronization calibration on the sampling signals; Specifically, this step is used to collect the secondary output sampling signals of multiple optical fiber current transformers at the same measurement point, and to perform noise reduction and synchronous calibration processing, thereby providing a high-quality data basis for subsequent group characteristic analysis.

[0019] Multiple optical fiber current transformers at the same measurement point are selected as the monitoring objects, and preferably, the number of transformers The secondary output signals of multiple transformers are collected with high precision and high frequency synchronization, and a sampling signal matrix is obtained.

[0020] Digital filtering technology is used to reduce noise in the sampling signal, filter out high-frequency noise interference, and retain effective measurement signal components. Since there may be slight deviations in the sampling time of each transformer, the sampling signal needs to be time-aligned and calibrated to ensure that the sampling data of each transformer at the same time can be accurately corresponded, thereby laying a foundation for subsequent group correlation analysis.

[0021] After preprocessing, a sampling signal matrix with dimensions of is obtained, where is the number of sampling points, is the number of transformers, and each row in the matrix represents the output value of transformers at the same time, and each column represents the time series data of one transformer.

[0022] S200: Construct a historical data matrix of the multiple optical fiber current transformers in a normal operating state and perform principal component analysis to establish a principal component space and calculate a squared prediction error control limit value;

[0023] Further, the step of establishing the principal component space comprises: standardizing the historical data matrix; calculating the covariance matrix of the standardized historical data matrix and performing eigenvalue decomposition to obtain eigenvectors in descending order of eigenvalues; determine the number of principal components according to the cumulative variance contribution rate, and select the corresponding eigenvectors to form the principal component space.

[0024] Specifically, this step is based on the historical data of multiple optical fiber current transformers in a normal operating state, and establishes a principal component space of the group correlation mode through principal component analysis (PCA), and determines a squared prediction error (SPE) control limit value, thereby providing a benchmark model for subsequent anomaly detection.

[0025] First, the historical synchronous sampling data of multiple transformers in a normal operating state is collected to construct a historical data matrix with dimensions of , where is the number of historical sampling points, is the number of transformers. Preferably, to ensure the stability of statistical characteristics.

[0026] The historical data matrix is standardized, that is, each column of data is subtracted by its mean value and divided by the standard deviation, to obtain a standardized historical data matrix. The standardization process can eliminate the influence of the dimension difference of different mutual inductor measurement values, so that each variable is on the same scale.

[0027] The covariance matrix of the standardized historical data matrix is calculated , and the covariance matrix is subjected to eigenvalue decomposition to obtain eigenvectors in descending order of eigenvalues. The eigenvalues reflect the variance in the corresponding principal component direction, and the eigenvectors represent the direction of the principal component.

[0028] The number of principal components is determined according to the cumulative variance contribution rate . Specifically, the cumulative variance contribution rate of the first eigenvalues is calculated, and the selection is stopped when the cumulative variance contribution rate reaches a preset threshold. Preferably, the threshold is set to not less than 85% to ensure that the principal component space can capture the main change pattern of the group data, while avoiding the increase of computational complexity caused by too many principal components. The first eigenvectors are selected to form a projection matrix of the principal component space .

[0029] Based on the established principal component space, the squared prediction error of each sample in the historical data matrix is calculated . For the th sample , the sample is projected into the principal component space to obtain principal component components, and the difference between the original data and the reconstructed value is calculated as the projection residual. The SPE statistic is defined as the sum of the squares of the components of the projection residual vector, representing the degree of deviation of the sample from the principal component space.

[0030] The SPE control limit is determined by statistical method . Preferably, assuming that the SPE statistic follows a chi-square distribution, the control limit is determined according to a 99% confidence interval. The control limit serves as a threshold reference for judging whether the real-time data deviates from the normal group pattern.

[0031] S300: Constructing a simulated fault data set, superimposing error signals of different amplitudes in the signal of one of the optical fiber current transformers to construct corresponding fault data and corresponding health status grade labels, projecting the simulated fault data set into the principal component space to calculate a squared prediction error time series; Further, the step of constructing a simulated fault data set comprises: On the basis of the historical data matrix, the signal of one of the fiber-optic current transformers is selected as the fault simulation object, and the signals of the remaining transformers remain unchanged. Different amplitude error signals are superimposed in the selected fiber-optic current transformer signal to construct fault data corresponding to four health state levels. On the basis of the historical data matrix, the signal of one of the fiber-optic current transformers is selected as the fault simulation object, and the signals of the remaining transformers remain unchanged. Different amplitude error signals are superimposed in the selected fiber-optic current transformer signal to construct fault data corresponding to four health state levels. According to the amplitude range of the superimposed error, the corresponding health state level label is marked for each fault data.

[0032] Further, the health state level label includes four levels: Level one corresponds to an error range of 0% to 0.05%; Level two corresponds to an error range of 0.05% to 0.2%; Level three corresponds to an error range of 0.2% to 0.5%; Level four corresponds to an error range of 0.5% or more.

[0033] Specifically, on the basis of the historical data matrix collected in step S200 , the signal of one of the fiber-optic current transformers is selected as the fault simulation object, and the signals of the remaining transformers remain unchanged. Different amplitude error signals are superimposed in the selected fiber-optic current transformer signal to construct fault data corresponding to four health state levels.

[0034] The corresponding relationship between the health state level and the error range is defined as follows. Level 1 corresponds to a normal state, and the error range is 0 to 0.05%. Level 2 corresponds to a light abnormality or slight degradation state, and the error range is 0.05% to 0.2%. Level 3 corresponds to a higher risk or moderate degradation state, and the error range is 0.2% to 0.5%. Level 4 corresponds to a serious fault state, and the error range is 0.5% or more.

[0035] According to the above error range, the corresponding amplitude error is superimposed on the selected transformer signal to generate simulated fault data of different levels. According to the amplitude range of the superimposed error, the corresponding health state level label is marked for each fault data to form a simulated fault data set and its corresponding label set.

[0036] Each sample in the simulated fault data set is projected into the principal component space established in step S200, and its projection residual relative to the principal component space is calculated, and then the squared prediction error statistic is calculated. The SPE statistics of each sample are arranged in time sequence to obtain a squared prediction error time sequence .

[0037] For ease of subsequent phased evaluation, the SPE statistics of each sample in the simulated fault data set The sequence is segmented according to a fixed number of data points. Preferably, each segment contains 100 data points, serving as an evaluation window. Each evaluation window corresponds to a set of... Sequence fragments, and inherit their corresponding health status level labels.

[0038] This step involves artificially adding error signals of varying magnitudes to normal historical data to construct a simulated fault dataset for training a deep learning hierarchical model. This method can generate training data covering different health state levels even in the absence of real fault samples.

[0039] S400: Perform multi-layer wavelet packet decomposition on the squared prediction error time series, calculate the energy value of the reconstructed signal of each sub-band and normalize it to construct a normalized energy feature vector, and use the normalized energy feature vector and the corresponding health status level label to train a one-dimensional convolutional neural network model. Furthermore, the step of performing multi-level wavelet packet decomposition on the squared prediction error time series includes: Perform the squared prediction error time series Layer wavelet packet decomposition yields Sub-band reconstructed signal; Calculate the first Layer The energy value of the reconstructed signal of each sub-band is the sum of the squares of the values ​​of each sampling point of the reconstructed signal of that sub-band. The energy values ​​of each sub-band are combined to form an energy vector, and the energy vector is normalized to obtain a normalized energy feature vector.

[0040] Specifically, the time series of squared prediction errors obtained in step S300 Perform multi-level wavelet packet decomposition. Wavelet packet analysis can simultaneously perform fine decomposition of the low-frequency and high-frequency components of a signal, and is suitable for handling nonlinear and non-stationary signals. Statistical signals can capture subtle local changes in the early stages of a fault.

[0041] Within each evaluation window Sequence fragments Layer wavelet packet decomposition yields The sub-band reconstructed signal. In this embodiment, the number of decomposition layers is preferably set. ,get Each sub-band reconstructed signal corresponds to a component of the original signal within a different frequency range. This layer selection ensures frequency resolution while avoiding excessive feature dimensionality due to over-decomposition.

[0042] Calculate the first Layer Energy value of each sub-band reconstructed signal The energy value is defined as the sum of squares of each sample point value of the sub-band reconstructed signal, that is The energy value is equal to the sum of squares of all sample point values in the sub-band reconstructed signal. The energy value reflects the energy distribution intensity of the signal in the frequency band.

[0043] The energy values of each sub-band are combined into an energy vector, which contains elements, each element corresponding to the energy value of a sub-band. The energy vector is normalized so that the sum of each element is equal to 1, obtaining the normalized energy feature vector The normalization process eliminates the influence of signal amplitude changes, making the feature more stable.

[0044] Repeat the wavelet packet decomposition and energy feature extraction process for all evaluation windows in the simulated fault data set to form a complete training set feature library. Each normalized energy feature vector corresponds to the health state level label of its own evaluation window.

[0045] A one-dimensional convolutional neural network model is constructed, which includes at least two one-dimensional convolutional layers for feature extraction, pooling layers for dimension reduction, fully connected layers for nonlinear mapping, and an output layer using Softmax activation function to output the probability distribution of four health state levels. One-dimensional convolutional neural network can automatically learn local patterns and nonlinear mapping relationships in feature vectors.

[0046] The labeled training set feature library is used to train the one-dimensional convolutional neural network model. The training process continuously adjusts the network parameters through the backpropagation algorithm to minimize the cross-entropy loss between the model output level probability distribution and the true label. After training is completed, the model can accurately predict the corresponding health state level according to the input normalized energy feature vector.

[0047] This step performs wavelet packet analysis on the squared prediction error time series, extracts time-frequency energy features, and trains a one-dimensional convolutional neural network model using labeled feature vectors to establish a nonlinear mapping relationship between features and health state levels.

[0048] S500: Projecting the real-time sampling signal into the principal component space to calculate the real-time squared prediction error time series, and determining whether it exceeds the squared prediction error control limit value; Further, the projecting the real-time sampling signal into the principal component space to calculate the real-time squared prediction error time series comprises: Segmenting the real-time sampling signal by a fixed number of data points, each segment serving as an evaluation window; Projecting the data in each evaluation window into the principal component space to calculate the squared prediction error statistic corresponding to the window; The squared prediction error statistics for each evaluation window are arranged in chronological order to form a real-time squared prediction error time series.

[0049] Specifically, this step involves online monitoring of the real-time collected data to be detected, calculating the real-time squared prediction error statistic using the established principal component space model, and comparing it with the control limit to determine whether there are any abnormalities, thus providing triggering conditions for subsequent refined classification of health status.

[0050] The real-time acquired data to be detected, data2, undergoes the same noise reduction and synchronization calibration preprocessing as in step S100 to obtain the preprocessed real-time sampling signal matrix, the dimension of which is ,in This represents the number of sampling points in real time. This refers to the number of mutual inductors.

[0051] The real-time sampled signal is segmented into segments with a fixed number of data points, each segment serving as an evaluation window. Preferably, each evaluation window contains 100 data points, consistent with the segmentation method of the training data in step S300. This segmented processing method enables phased and continuous evaluation of the device's health status.

[0052] The data within each evaluation window are standardized using the mean and standard deviation parameters calculated in step S200 when establishing the principal component space, ensuring consistency in standardization. The standardized data are then projected onto the principal component space established in step S200, using a projection matrix. Perform projection and reconstruction, and calculate the projection residual.

[0053] Calculate the squared prediction error statistic for each sample within the evaluation window. For the nth sample within the evaluation window... One sample, its The statistic is equal to the sum of the squares of the components of the projected residual vector of the sample. The squared prediction error statistics of all samples within each evaluation window are arranged in chronological order to form a real-time squared prediction error time series. .

[0054] Real-time squared prediction error time series Compared with the control limit determined in step S200 Perform a comparison. Make a judgment. Does the sequence contain values ​​exceeding In this situation. If The values ​​in the sequence did not exceed the control limits. This indicates that the population measurement pattern within the current evaluation window has not deviated from the normal pattern and is judged to be in a normal state. A certain evaluation window's... The sequence contains values ​​that exceed the control limits. If the abnormal early warning is triggered, it indicates that there is an anomaly, and further health state level evaluation needs to be performed in step S600.

[0055] S600: If the control limit is exceeded, the real-time square prediction error time series is subjected to multi-layer wavelet packet decomposition and a normalized energy feature vector is extracted, and a trained one-dimensional convolutional neural network model is input to output the health state level; if it is not exceeded, it is determined to be in a normal state.

[0056] The input of the one-dimensional convolutional neural network model is the normalized energy feature vector, the output layer uses a Softmax activation function to output the probability distribution of four health state levels, and the level corresponding to the maximum probability is selected as the health state evaluation result.

[0057] Specifically, this step performs refined health state level evaluation for the evaluation window determined to be abnormal in step S500, and uses a trained one-dimensional convolutional neural network model to output a specific health state level.

[0058] If all values in the real-time square prediction error time series in step S500 do not exceed the control limit , the health state corresponding to the current evaluation window is directly determined to be normal, i.e., level 1, and no subsequent deep learning classification processing is required.

[0059] If the real-time square prediction error time series exceeds the control limit in step S500, further analysis of the sequence segment corresponding to the evaluation window is required. The same wavelet packet analysis processing as in step S400 is performed on the sequence segment in the current evaluation window.

[0060] Specifically, the sequence segment is subjected to layer wavelet packet decomposition, preferably , to obtain sub-band reconstruction signals. The energy values of the sub-band reconstruction signals are calculated, and the energy value is the sum of the squares of the sampling point values of the sub-band reconstruction signals. The energy values of the sub-band are combined to form an energy vector, which is normalized so that the sum of the elements is equal to 1, to obtain a normalized energy feature vector .

[0061] The normalized energy feature vector As input, the one-dimensional convolutional neural network model trained in step S400 is input. The model extracts feature patterns through the convolutional layer and performs nonlinear mapping through the fully connected layer. The output layer of the model adopts the Softmax activation function, and outputs the probability distribution of four health state levels, wherein each level corresponds to a probability value, and the sum of the four probability values is equal to 1.

[0062] The level corresponding to the maximum probability is selected as the health state evaluation result of the current evaluation window. The evaluation result represents the health condition of the fiber-optic current transformer in the current period, and the levels 1 to 4 correspond to the normal state, the slight abnormality, the higher risk and the serious failure, respectively.

[0063] The health state evaluation result is output to the monitoring system or the human-computer interaction interface for reference by the operation and maintenance personnel. The evaluation result can reflect the degradation trend of the device in real time, and provide a basis for preventive maintenance. Due to the simple structure of the one-dimensional convolutional neural network, the single sample inference time can be controlled within 10 ms, which meets the needs of online real-time evaluation of the fiber-optic current transformer.

[0064] Embodiment two: This embodiment takes the health state monitoring of the fiber-optic current transformer of a bus measurement point of a certain 220kV substation as an example to show the application effect of the technical solution of the present application in actual engineering.

[0065] The bus measurement point of the substation is configured with three fiber-optic current transformers for redundant measurement. The health state evaluation system of the fiber-optic transformer based on feature fusion of the present application is used for online monitoring, the structure of the system is as shown in Figure 2 , which includes a signal preprocessing module, a principal component modeling module, a data construction module, a model training module, a real-time monitoring module and a state evaluation module. The overall evaluation process of the system is as shown in Figure 3 .

[0066] The signal preprocessing module collects the secondary output signals of the three transformers in the normal operating state, the sampling point number is greater than 2000, and the historical data matrix with a dimension of is obtained after noise reduction and synchronous calibration.

[0067] The principal component modeling module performs standardization processing on the historical data matrix and performs principal component analysis. The number of principal components is determined according to the standard that the cumulative variance contribution rate is not less than 85%, the principal component space is established, and the control limit value is calculated.

[0068] The data construction module constructs a simulated fault data set on the basis of normal historical data. By superimposing error signals of different amplitudes, training data covering four health state levels are generated. Level 1 corresponds to an error range of 0-0.05%, level 2 corresponds to 0.05%-0.2%, level 3 corresponds to 0.2%-0.5%, and level 4 corresponds to more than 0.5%.

[0069] The model training module performs 3-layer wavelet packet decomposition on the SPE sequence to obtain 8 sub-band reconstruction signals and calculate a normalized energy feature vector. A one-dimensional convolutional neural network model is trained using the labeled feature vector.

[0070] After running for a period of time, the real-time monitoring module processes the real-time sampling signal in segments of 100 data points, calculates statistical quantities. In the sequence of a certain evaluation window, there are values exceeding the control limit .

[0071] The state evaluation module performs wavelet packet decomposition on the SPE sequence of the abnormal window and extracts a feature vector, which is input into the trained one-dimensional convolutional neural network model. The model outputs a probability distribution of four levels, and the level corresponding to the maximum probability is selected as the evaluation result. The system determines that the current health state is level 2, corresponding to an error range of 0.05%-0.2%.

[0072] According to the evaluation result, the operation and maintenance personnel arrange offline verification, and the result shows that the actual error of the mutual inductor is about 0.1%, which is completely consistent with the system evaluation.

[0073] The traditional offline verification method has a long cycle and requires power-off operation, while the present application realizes online real-time monitoring, immediately discovers abnormalities and accurately classifies them. The traditional online method cannot distinguish between power grid fluctuations and device degradation, while the present application extracts group correlation features through principal component analysis and effectively separates normal fluctuations from abnormal degradation. The traditional method can only provide simple alarms, while the present application combines wavelet packet analysis and deep learning to achieve four-level accurate classification. The system takes less than 10ms for a single evaluation, meeting the real-time monitoring requirements and effectively ensuring the accuracy and reliability of power system measurement.

[0074] Finally, it should be noted that the above-described preferred embodiments of the present application are not intended to limit the present application, and although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements of some technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A health state evaluation method for an optical fiber mutual inductor based on feature fusion, characterized in that, The method comprises the following steps: S100: Collecting sampling signals of multiple fiber-optic current transformers at the same measuring point, and performing noise reduction and synchronization calibration on the sampling signals; S200: Constructing a historical data matrix of the multiple fiber-optic current transformers in a normal operating state and performing principal component analysis, establishing a principal component space and calculating a squared prediction error control limit value; S300: Constructing a simulated fault data set by superimposing error signals of different amplitudes in the signal of one of the fiber-optic current transformers, constructing corresponding fault data and corresponding health status grade labels, projecting the simulated fault data set into the principal component space to calculate a squared prediction error time sequence; S400: Performing multi-layer wavelet packet decomposition on the squared prediction error time sequence, calculating the energy values of the reconstructed signals of each sub-band, normalizing to construct a normalized energy feature vector, and training a one-dimensional convolutional neural network model using the normalized energy feature vector and the corresponding health status grade labels; S5 S00: Projecting real-time sampling signals into the principal component space to calculate a real-time squared prediction error time sequence, and determining whether the squared prediction error control limit value is exceeded; S600: If the control limit value is exceeded, performing multi-layer wavelet packet decomposition on the real-time squared prediction error time sequence and extracting a normalized energy feature vector, inputting the trained one-dimensional convolutional neural network model to output a health status grade; if not, determining that the state is normal.

2. The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The step of establishing the principal component space comprises: standardizing the historical data matrix; calculating the covariance matrix of the standardized historical data matrix and performing eigenvalue decomposition to obtain eigenvectors in descending order of eigenvalues; determining the number of principal components according to the cumulative variance contribution rate, and selecting the corresponding eigenvectors to form the principal component space. 3.The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The step of constructing the simulated fault data set comprises: selecting the signal of one of the fiber-optic current transformers based on the historical data matrix; superimposing error signals of different amplitudes in the selected fiber-optic current transformer signal to construct fault data corresponding to four health status grades, and keeping the signals of the remaining fiber-optic current transformers unchanged; annotating the corresponding health status grade labels for each fault data according to the amplitude range of the superimposed error.

4. The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The health status grade labels comprise four grades: Grade one corresponds to an error range of 0% to 0.05%; Grade two corresponds to an error range of 0.05% to 0.2%; Grade three corresponds to an error range of 0.2% to 0.5%; Grade four corresponds to an error range of 0.5% or more.

5. The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The step of performing multi-layer wavelet packet decomposition on the squared prediction error time sequence comprises: performing a wavelet packet decomposition on the square prediction error time series a layer wavelet packet decomposition, to obtain a sub-band reconstruction signal; Calculate the first Layer The energy value of the reconstructed signal of each sub-band is the sum of the squares of the values ​​of each sampling point of the reconstructed signal of that sub-band. grouping the energy values of each sub-band into an energy vector, and normalizing the energy vector to obtain a normalized energy feature vector.

6. The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The step of projecting the real-time sampling signals into the principal component space to calculate a real-time squared prediction error time sequence comprises: segmenting the real-time sampling signals into segments according to a fixed number of data points, and taking each segment as an evaluation window; projecting the data in each evaluation window into the principal component space to calculate the squared prediction error statistics corresponding to the window; arranging the squared prediction error statistics of each evaluation window in time sequence to form a real-time squared prediction error time sequence.

7. The health state evaluation method of the optical fiber mutual inductor based on feature fusion according to claim 1, characterized in that, The one-dimensional convolutional neural network model takes the normalized energy feature vector as input, and adopts a Softmax activation function in an output layer to output a probability distribution of four health state grades, and selects a grade corresponding to a maximum probability as a health state evaluation result.

8. The health state evaluation system of the optical fiber mutual inductor based on feature fusion, characterized in that, The method comprises the following steps: A signal preprocessing module is configured to collect sampling signals of multiple fiber-optic current transformers at a same measuring point, and to perform noise reduction and synchronization calibration on the sampling signals; A principal component modeling module is configured to construct a historical data matrix of the multiple fiber-optic current transformers in a normal operating state, and to perform principal component analysis, establish a principal component space, and calculate a squared prediction error control limit value; A data construction module is configured to construct a simulated fault data set by superimposing error signals of different amplitudes in a signal of one of the fiber-optic current transformers, construct corresponding fault data and corresponding health state grade labels, and project the simulated fault data set to the principal component space to calculate a squared prediction error time sequence; A model training module is configured to perform multi-layer wavelet packet decomposition on the squared prediction error time sequence, calculate energy values of reconstructed signals of each sub-band, normalize the energy values to construct a normalized energy feature vector, and train a one-dimensional convolutional neural network model using the normalized energy feature vector and the corresponding health state grade labels; A real-time monitoring module is configured to project real-time sampling signals to the principal component space to calculate a real-time squared prediction error time sequence, and determine whether the real-time squared prediction error time sequence exceeds the squared prediction error control limit value; A state evaluation module is configured to, if the real-time squared prediction error time sequence exceeds the control limit value, perform multi-layer wavelet packet decomposition on the real-time squared prediction error time sequence, extract a normalized energy feature vector, input the normalized energy feature vector into the trained one-dimensional convolutional neural network model, and output a health state grade; and if the real-time squared prediction error time sequence does not exceed the control limit value, determine that the state is normal.

Images