Health state evaluation method and system for fiber optic mutual inductor based on feature fusion

Abstract

The present application relates to the field of power equipment operation and maintenance and state monitoring technology, in particular to a fiber-optic mutual inductor health state evaluation method and system based on feature fusion. The present application collects sampling signals of multiple fiber-optic current transformers and pre-processes them; constructs a historical data matrix for principal component analysis, establishes a principal component space and calculates a squared prediction error control limit value; constructs a simulated fault data set and projects it to the principal component space to calculate a squared prediction error time series; performs wavelet packet decomposition on the squared prediction error time series to extract a normalized energy feature vector and trains a one-dimensional convolutional neural network model; projects real-time sampling signals to the principal component space to determine whether they are out of limit; and if they are out of limit, outputs a health state grade through the trained model. The present application realizes accurate grading and real-time online evaluation of the health state of fiber-optic mutual inductors.

Description

Technical Field

[0001] This invention relates to the field of power equipment operation and maintenance and condition monitoring technology, specifically to a method and system for assessing the health status of fiber optic instrument transformers based on feature fusion. Background Technology

[0002] Fiber optic current transformers are widely used in high-voltage power systems due to their superior insulation performance and fast response capability. However, they will gradually deteriorate due to environmental factors and their own aging during long-term operation, resulting in a decrease in measurement accuracy.

[0003] Traditional health status assessments primarily rely on periodic offline calibration, which suffers from drawbacks such as long cycles, high costs, and the inability to monitor in real time. Existing online methods are mostly based on single characteristics for classification, making it difficult to comprehensively reflect the multi-dimensional degradation characteristics of equipment, and their ability to process nonlinear and non-stationary signals is insufficient, resulting in low assessment accuracy. More importantly, when one of multiple instrument transformers at the same measurement point begins to deteriorate, relying solely on the measurement value of a single device is insufficient to determine whether the error is caused by grid fluctuations or equipment degradation. However, by analyzing group measurement data, the correlation between instrument transformers can be effectively utilized to more sensitively detect abnormal equipment. Summary of the Invention

[0004] This invention provides a method and system for assessing the health status of fiber optic instrument transformers based on feature fusion, which solves the problems of poor real-time performance, single feature, and weak nonlinear processing capability of traditional methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a method for assessing the health status of fiber optic instrument transformers based on feature fusion, comprising:

[0007] 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;

[0008] S200: Construct a historical data matrix of the multiple fiber optic current transformers under normal operating conditions and perform principal component analysis, establish a principal component space and calculate the squared prediction error control limit.

[0009] S300: Construct a simulated fault dataset by superimposing error signals of different amplitudes into the signal of one of the fiber optic current transformers to construct corresponding fault data and corresponding health status level labels, and project the simulated fault dataset onto the principal component space to calculate the squared prediction error time series.

[0010] 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.

[0011] S500: Project the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series, and determine whether it exceeds the squared prediction error control limit;

[0012] S600: If the control limit is exceeded, multi-layer wavelet packet decomposition is performed on the real-time squared prediction error time series and the normalized energy feature vector is extracted. The normalized energy feature vector is then input into the trained one-dimensional convolutional neural network model to output the health status level. If the limit is not exceeded, the system is judged to be in a normal state.

[0013] As a preferred embodiment of the present invention, the step of establishing the principal space includes:

[0014] The historical data matrix is ​​standardized.

[0015] Calculate the covariance matrix of the standardized historical data matrix and perform eigenvalue decomposition. Sort the eigenvalues ​​from largest to smallest to obtain the eigenvectors.

[0016] The number of principal components is determined based on the cumulative variance contribution rate, and the corresponding eigenvectors are selected to form the principal component space.

[0017] As a preferred embodiment of the present invention, the step of constructing the simulated fault dataset includes:

[0018] Based on the historical data matrix, the signal of one of the fiber optic current transformers is selected;

[0019] Error signals of different amplitudes are superimposed on the selected fiber optic current transformer signals to construct fault data corresponding to four health status levels, while the signals of the other fiber optic current transformers remain unchanged.

[0020] Each fault data point is labeled with a corresponding health status level label based on the range of the superimposed error.

[0021] As a preferred embodiment of the present invention, the health status level label includes four levels:

[0022] The error range for Level 1 is 0% to 0.05%;

[0023] The error range for Level 2 is 0.05% to 0.2%.

[0024] The error range for Level 3 is 0.2% to 0.5%.

[0025] Level 4 corresponds to an error range of 0.5% or more.

[0026] As a preferred embodiment of the present invention, the step of performing multi-level wavelet packet decomposition on the squared prediction error time series includes:

[0027] Perform the squared prediction error time series Layer wavelet packet decomposition, yielding Sub-band reconstructed signal;

[0028] 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.

[0029] 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.

[0030] As a preferred embodiment of the present invention, the step of projecting the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series includes:

[0031] The real-time sampled signal is divided into segments with a fixed number of data points, and each segment serves as an evaluation window;

[0032] Project the data within each evaluation window onto the principal component space and calculate the squared prediction error statistic corresponding to that window.

[0033] The squared prediction error statistics for each evaluation window are arranged in chronological order to form a real-time squared prediction error time series.

[0034] As a preferred embodiment of the present invention, the input of the one-dimensional convolutional neural network model is the normalized energy feature vector, and the output layer uses the Softmax activation function to output the probability distribution of four health status levels, and selects the level corresponding to the highest probability as the health status assessment result.

[0035] This invention also proposes a fiber optic instrument transformer health status assessment system based on feature fusion, comprising:

[0036] The signal preprocessing module is used to collect sampling signals from multiple fiber optic current transformers at the same measurement point, and to perform noise reduction and synchronization calibration on the sampling signals.

[0037] The principal component modeling module is used to construct the historical data matrix of the multiple fiber optic current transformers under normal operating conditions and perform principal component analysis, establish the principal component space and calculate the squared prediction error control limit.

[0038] The data construction module is used to construct a simulated fault dataset. By superimposing error signals of different amplitudes into the signal of one of the fiber optic current transformers, the corresponding fault data and corresponding health status level labels are constructed. The simulated fault dataset is then projected onto the principal component space to calculate the squared prediction error time series.

[0039] The model training module is used to 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.

[0040] The real-time monitoring module is used to project the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series and determine whether it exceeds the squared prediction error control limit.

[0041] The status assessment module is used to perform multi-layer wavelet packet decomposition on the real-time squared prediction error time series and extract the normalized energy feature vector if the control limit is exceeded, and input it into the trained one-dimensional convolutional neural network model to output the health status level; if the limit is not exceeded, it is determined to be a normal state.

[0042] The beneficial effects of this invention are:

[0043] 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.

[0044] 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.

[0045] 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

[0046] 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:

[0047] Figure 1 This is a flowchart illustrating the fiber optic instrument health status assessment method based on feature fusion of the present invention.

[0048] 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;

[0049] 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

[0050] 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.

[0051] Example 1: As Figure 1 As shown, the fiber optic instrument transformer health status assessment method based on feature fusion includes:

[0052] 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;

[0053] Specifically, this step is used to collect the secondary output sampling signals of multiple fiber optic current transformers at the same measurement point, and to perform noise reduction and synchronization calibration processing to provide a high-quality data foundation for subsequent group characteristic analysis.

[0054] Multiple fiber optic current transformers at the same measurement point are selected as monitoring objects. Preferably, the number of transformers is... The secondary output signals of multiple instrument transformers are synchronously acquired with high precision and high frequency to obtain a sampling signal matrix.

[0055] Digital filtering technology is used to reduce noise in the sampled signals, filtering out high-frequency noise interference and retaining effective measurement signal components. Since there may be slight deviations in the sampling times of each instrument transformer, time alignment calibration of the sampled signals is required to ensure that the sampling data from each instrument transformer at the same time can accurately correspond, thus laying the foundation for subsequent group correlation analysis.

[0056] After preprocessing, the dimension is obtained as follows: The sampled signal matrix, where The number of sampling points. This represents the number of mutual inductors; each row in the matrix represents the same time period. The output values ​​of the current transformers are displayed, with each column representing the time series data of one current transformer.

[0057] S200: Construct a historical data matrix of the multiple fiber optic current transformers under normal operating conditions and perform principal component analysis, establish a principal component space and calculate the squared prediction error control limit.

[0058] Furthermore, the step of establishing the principal space includes:

[0059] The historical data matrix is ​​standardized.

[0060] Calculate the covariance matrix of the standardized historical data matrix and perform eigenvalue decomposition. Sort the eigenvalues ​​from largest to smallest to obtain the eigenvectors.

[0061] The number of principal components is determined based on the cumulative variance contribution rate, and the corresponding eigenvectors are selected to form the principal component space.

[0062] Specifically, this step uses historical data from multiple fiber optic current transformers under normal operating conditions to establish a principal component space for the group correlation pattern through principal component analysis (PCA) and determine the squared prediction error (SPE) control limit, providing a benchmark model for subsequent anomaly detection.

[0063] First, historical synchronous sampling data from multiple current transformers under normal operating conditions are collected to construct a historical data matrix. Its dimension is ,in This represents the number of historical sampling points. This refers to the number of current transformers. Preferably, To ensure the stability of statistical properties.

[0064] Historical data matrix Standardization is performed by subtracting the mean from each column of data and dividing by the standard deviation to obtain a standardized historical data matrix. Standardization eliminates the influence of dimensional differences in measurements from different transformers, bringing all variables to the same scale.

[0065] Calculate the covariance matrix of the standardized historical data matrix The covariance matrix is ​​then decomposed into eigenvalues, and the eigenvectors are obtained by sorting the eigenvalues ​​from largest to smallest. The eigenvalues ​​reflect the magnitude of the variance along the corresponding principal component direction, and the eigenvectors represent the direction of the principal component.

[0066] The number of principal components is determined based on the cumulative variance contribution rate. Specifically, before calculation The cumulative variance contribution rate of each feature value is used to stop selection when the cumulative variance contribution rate reaches a preset threshold. Preferably, this threshold is set to be no less than 85% to ensure that the principal component space can capture the main change patterns of the population data, while avoiding increased computational complexity caused by too many principal components. The corresponding top... The eigenvectors constitute the projection matrix of the principal space. .

[0067] Based on the established principal component space, calculate the historical data matrix. Squared prediction error for each sample For the first Sample The original data is projected onto the principal component space to obtain principal components. The data is then reconstructed using the principal components and the projection matrix to obtain the reconstructed values. The difference between the original data and the reconstructed values ​​is the projection residual. The SPE statistic is defined as the sum of squares of the components of the projection residual vector, representing the degree to which the sample deviates from the principal component space.

[0068] SPE control limits were determined using statistical methods. Preferably, it is assumed that the SPE statistic follows a chi-square distribution, and the control limits are determined based on the 99% confidence interval. This control limit serves as a threshold benchmark for determining whether real-time data deviates from the normal population pattern.

[0069] S300: Construct a simulated fault dataset by superimposing error signals of different amplitudes into the signal of one of the fiber optic current transformers to construct corresponding fault data and corresponding health status level labels, and project the simulated fault dataset onto the principal component space to calculate the squared prediction error time series.

[0070] Furthermore, the step of constructing the simulated fault dataset includes:

[0071] Based on the historical data matrix, the signal of one of the fiber optic current transformers is selected;

[0072] Error signals of different amplitudes are superimposed on the selected fiber optic current transformer signals to construct fault data corresponding to four health status levels, while the signals of the other fiber optic current transformers remain unchanged.

[0073] Each fault data point is labeled with a corresponding health status level label based on the range of the superimposed error.

[0074] Furthermore, the health status level label includes four levels:

[0075] The error range for Level 1 is 0% to 0.05%;

[0076] The error range for Level 2 is 0.05% to 0.2%.

[0077] The error range for Level 3 is 0.2% to 0.5%.

[0078] Level 4 corresponds to an error range of 0.5% or more.

[0079] Specifically, the historical data matrix collected in step S200 Based on this, the signal of one fiber optic current transformer is selected as the fault simulation object, while the signals of the other transformers remain unchanged. Error signals of different amplitudes are superimposed on the selected fiber optic current transformer signal to construct fault data corresponding to four health status levels.

[0080] The correspondence between health status levels and error ranges is defined as follows: Level 1 corresponds to a normal state, with an error range of 0 to 0.05%. Level 2 corresponds to a slightly abnormal or minor deterioration state, with an error range of 0.05% to 0.2%. Level 3 corresponds to a higher risk or moderate deterioration state, with an error range of 0.2% to 0.5%. Level 4 corresponds to a severe failure state, with an error range of over 0.5%.

[0081] Based on the aforementioned error range, an error of corresponding amplitude is superimposed on the selected transformer signal to generate simulated fault data of different levels. Each fault data point is labeled with a corresponding health status level based on the amplitude range of the superimposed error, forming a simulated fault dataset. and its corresponding tag set.

[0082] Simulated fault dataset Each sample is projected onto the principal component space established in step S200, and its projection residual relative to the principal component space is calculated. Then, the squared prediction error statistic (SPE) is calculated. The SPE statistics of each sample are arranged in chronological order to obtain the time series of the squared prediction error. .

[0083] To facilitate subsequent phased evaluations, 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.

[0084] 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.

[0085] 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.

[0086] Furthermore, the step of performing multi-level wavelet packet decomposition on the squared prediction error time series includes:

[0087] Perform the squared prediction error time series Layer wavelet packet decomposition, yielding Sub-band reconstructed signal;

[0088] 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.

[0089] 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.

[0090] 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.

[0091] Within each evaluation window Sequence fragments Layer wavelet packet decomposition, yielding 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.

[0092] Calculate the first Layer Energy value of reconstructed signal in each sub-band The energy value is defined as the sum of the squares of the values ​​at each sampling point of the reconstructed signal in that sub-band, i.e. It is equal to the sum of the squares of all sampled values ​​in the reconstructed signal of that sub-band. The energy value reflects the energy distribution intensity of the signal within that frequency band.

[0093] The energy values ​​of each sub-band are combined to form an energy vector, which contains... Each element represents an energy value for a sub-frequency band. The energy vector is then normalized so that the sum of its elements equals 1, resulting in a normalized energy feature vector. Normalization eliminates the influence of signal amplitude variations, making the characteristics more stable.

[0094] Repeat the wavelet packet decomposition and energy feature extraction process described above for all evaluation windows in the simulated fault dataset to form a complete training set feature library. Each normalized energy feature vector corresponds to the health status level label of its respective evaluation window.

[0095] A one-dimensional convolutional neural network model is constructed, comprising at least two one-dimensional convolutional layers for feature pattern extraction, pooling layers for dimensionality reduction, fully connected layers for nonlinear mapping, and an output layer employing the Softmax activation function to output the probability distribution of four health state levels. This one-dimensional convolutional neural network can automatically learn local patterns and nonlinear mapping relationships in feature vectors.

[0096] A one-dimensional convolutional neural network model is trained using a labeled training set feature library. During training, the network parameters are continuously adjusted using the backpropagation algorithm to minimize the cross-entropy loss between the model's output level probability distribution and the true labels. After training, the model can accurately predict the corresponding health status level based on the input normalized energy feature vector.

[0097] This step performs wavelet packet analysis on the time series of squared prediction errors to extract time-frequency domain energy features, and uses labeled feature vectors to train a one-dimensional convolutional neural network model to establish a nonlinear mapping relationship between features and health status levels.

[0098] S500: Project the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series, and determine whether it exceeds the squared prediction error control limit;

[0099] Further, the step of projecting the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series includes:

[0100] The real-time sampled signal is divided into segments with a fixed number of data points, and each segment serves as an evaluation window;

[0101] Project the data within each evaluation window onto the principal component space and calculate the squared prediction error statistic corresponding to that window.

[0102] The squared prediction error statistics for each evaluation window are arranged in chronological order to form a real-time squared prediction error time series.

[0103] 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.

[0104] 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.

[0105] 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.

[0106] 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.

[0107] 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. .

[0108] 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 an abnormality warning is triggered, it indicates that an abnormality exists and further health status level assessment needs to be performed in step S600.

[0109] S600: If the control limit is exceeded, multi-layer wavelet packet decomposition is performed on the real-time squared prediction error time series and the normalized energy feature vector is extracted. The normalized energy feature vector is then input into the trained one-dimensional convolutional neural network model to output the health status level. If the limit is not exceeded, the system is judged to be in a normal state.

[0110] The input to the one-dimensional convolutional neural network model is the normalized energy feature vector. The output layer uses the Softmax activation function to output the probability distribution of four health status levels, and selects the level corresponding to the highest probability as the health status assessment result.

[0111] Specifically, this step performs a refined health status level assessment on the assessment window that is determined to be abnormal in step S500, and outputs the specific health status level using the trained one-dimensional convolutional neural network model.

[0112] If in step S500 the real-time squared prediction error time series is determined... All values ​​did not exceed the control limits. If the current assessment window is determined to be in a normal state, i.e., level 1, then no further deep learning classification processing is required.

[0113] If in step S500 the real-time squared prediction error time series is determined... Exceeding control limits Then it is necessary to consider the evaluation window corresponding to Further analysis of the sequence fragments is needed. Within the current evaluation window... The sequence fragments undergo the same wavelet packet analysis processing as in step S400.

[0114] Specifically, regarding Sequence fragments Layer wavelet packet decomposition, preferably ,get Reconstruct the signal for each sub-band. Calculate the energy value of the reconstructed signal for each sub-band. The energy value is the sum of the squares of the values ​​at each sampling point of the reconstructed signal for that sub-band. Construct an energy vector from the energy values ​​of each sub-band and normalize it so that the sum of all elements equals 1, obtaining the normalized energy feature vector. .

[0115] Normalized energy eigenvectors As input, it is fed into the one-dimensional convolutional neural network model trained in step S400. This model extracts feature patterns through convolutional layers and performs non-linear mapping through fully connected layers. The model's output layer uses the Softmax activation function, outputting a probability distribution of four health state levels, where each level corresponds to a probability value, and the sum of the four probability values ​​equals 1.

[0116] The level corresponding to the highest probability is selected as the health status assessment result for the current assessment window. This assessment result characterizes the health status of the fiber optic current transformer during the current time period, with levels 1 to 4 representing normal status, minor anomaly, high risk, and serious fault, respectively.

[0117] The health status assessment results are output to the monitoring system or human-machine interface for maintenance personnel to refer to and make decisions. These assessment results can reflect the equipment's degradation trend in real time, providing a basis for preventative maintenance. Due to the simple structure of the one-dimensional convolutional neural network, the single-sample inference time can be controlled within 10ms, meeting the requirements for online real-time assessment of fiber optic current transformers.

[0118] Example 2: This example uses the health status monitoring of the fiber optic current transformer at a busbar measurement point in a 220kV substation as an example to demonstrate the application effect of the technical solution of the present invention in actual engineering.

[0119] The substation busbar measurement point is equipped with three fiber optic current transformers for redundant measurement. The fiber optic current transformer health status assessment system based on feature fusion, as described in this invention, is used for online monitoring. The structure of this system is as follows: Figure 2 As shown, the system 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 follows: Figure 3 As shown.

[0120] The signal preprocessing module acquires the secondary output signals of three current transformers under normal operating conditions, with a sampling point count of [number missing]. The dimension is greater than 2000, and after noise reduction and synchronization calibration, it is obtained as follows: Historical data matrix.

[0121] The principal component modeling module standardizes the historical data matrix and performs principal component analysis. The number of principal components is determined based on a cumulative variance contribution rate of at least 85%, and the principal component space is established and calculated. Control Limits .

[0122] The data construction module builds a simulated fault dataset based on normal historical data. By superimposing error signals of different amplitudes, training data covering four health status levels is generated. Level 1 corresponds to an error range of 0 to 0.05%, Level 2 corresponds to 0.05% to 0.2%, Level 3 corresponds to 0.2% to 0.5%, and Level 4 corresponds to above 0.5%.

[0123] The model training module performs a 3-layer wavelet packet decomposition on the SPE sequence to obtain a reconstructed signal in 8 sub-bands and calculates the normalized energy feature vector. A one-dimensional convolutional neural network model is then trained using the labeled feature vector.

[0124] After running for a period of time, the real-time monitoring module will process the real-time sampled signal into segments of 100 data points and calculate... Statistic. A statistic for a certain evaluation window. The sequence contains values ​​that exceed the control limits. .

[0125] The state assessment module performs wavelet packet decomposition on the SPE sequence of the abnormal window and extracts feature vectors, which are then input into a trained one-dimensional convolutional neural network model. The model outputs a probability distribution of four levels, and the level corresponding to the highest probability is selected as the assessment result. The system determines the current health status to be level 2, with an error range of 0.05% to 0.2%.

[0126] The maintenance personnel arranged offline verification based on the assessment results, and the results showed that the actual error of the current transformer was about 0.1%, which was completely consistent with the system assessment.

[0127] Traditional offline verification methods are time-consuming and require power outages. This invention achieves online real-time monitoring, enabling immediate detection and accurate classification of anomalies. Traditional online methods struggle to distinguish between grid fluctuations and equipment degradation. This invention extracts population correlation features through principal component analysis, and the SPE statistic effectively separates normal fluctuations from abnormal degradation. Traditional methods only provide simple alarms; this invention combines wavelet packet analysis and deep learning to achieve a precise four-level classification. A single system assessment takes less than 10ms, meeting real-time monitoring requirements and effectively ensuring the accuracy and reliability of power system measurements.

[0128] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention 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 substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing the health status of fiber optic instrument transformers based on feature fusion, characterized in that, include: 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; S200: Construct a historical data matrix of the multiple fiber optic current transformers under normal operating conditions and perform principal component analysis, establish a principal component space and calculate the squared prediction error control limit. S300: Construct a simulated fault dataset by superimposing error signals of different amplitudes into the signal of one of the fiber optic current transformers to construct corresponding fault data and corresponding health status level labels, and project the simulated fault dataset onto the principal component space to calculate the squared prediction error time series. 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. S5 00: Project the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series, and determine whether it exceeds the squared prediction error control limit; S600: If the control limit is exceeded, multi-layer wavelet packet decomposition is performed on the real-time squared prediction error time series and the normalized energy feature vector is extracted. The normalized energy feature vector is then input into the trained one-dimensional convolutional neural network model to output the health status level. If it does not exceed the limit, it is considered to be in a normal state; The steps for performing multi-level wavelet packet decomposition on the squared prediction error time series include: Perform the squared prediction error time series Layer wavelet packet decomposition, yielding 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 into an energy vector, and the energy vector is normalized to obtain a normalized energy feature vector. The step of projecting the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series includes: The real-time sampled signal is divided into segments with a fixed number of data points, and each segment serves as an evaluation window; Project the data within each evaluation window onto the principal component space and calculate the squared prediction error statistic corresponding to that 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.

2. The method for assessing the health status of fiber optic transformers based on feature fusion according to claim 1, characterized in that, The steps for establishing the principal space include: The historical data matrix is ​​standardized. Calculate the covariance matrix of the standardized historical data matrix and perform eigenvalue decomposition. Sort the eigenvalues ​​from largest to smallest to obtain the eigenvectors. The number of principal components is determined based on the cumulative variance contribution rate, and the corresponding eigenvectors are selected to form the principal component space.

3. The method for assessing the health status of fiber optic transformers based on feature fusion according to claim 1, characterized in that, The steps for constructing the simulated fault dataset include: Based on the historical data matrix, the signal of one of the fiber optic current transformers is selected; Error signals of different amplitudes are superimposed on the selected fiber optic current transformer signals to construct fault data corresponding to four health status levels, while the signals of the other fiber optic current transformers remain unchanged. Each fault data point is labeled with a corresponding health status level label based on the range of the superimposed error.

4. The method for assessing the health status of fiber optic transformers based on feature fusion according to claim 1, characterized in that, The health status level label includes four levels: The error range for Level 1 is 0% to 0.05%; The error range for Level 2 is 0.05% to 0.2%. The error range for Level 3 is 0.2% to 0.5%. Level 4 corresponds to an error range of 0.5% or more.

5. The method for assessing the health status of fiber optic transformers based on feature fusion according to claim 1, characterized in that, The input to the one-dimensional convolutional neural network model is the normalized energy feature vector. The output layer uses the Softmax activation function to output the probability distribution of four health status levels, and selects the level corresponding to the highest probability as the health status assessment result.

6. A fiber optic instrument transformer health status assessment system based on feature fusion, characterized in that, include: The signal preprocessing module is used to collect sampling signals from multiple fiber optic current transformers at the same measurement point, and to perform noise reduction and synchronization calibration on the sampling signals. The principal component modeling module is used to construct the historical data matrix of the multiple fiber optic current transformers under normal operating conditions and perform principal component analysis, establish the principal component space and calculate the squared prediction error control limit. The data construction module is used to construct a simulated fault dataset. By superimposing error signals of different amplitudes into the signal of one of the fiber optic current transformers, the corresponding fault data and corresponding health status level labels are constructed. The simulated fault dataset is then projected onto the principal component space to calculate the squared prediction error time series. The model training module is used to 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. The real-time monitoring module is used to project the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series and determine whether it exceeds the squared prediction error control limit. The status assessment module is used to perform multi-layer wavelet packet decomposition on the real-time squared prediction error time series and extract the normalized energy feature vector if the control limit is exceeded, and input the normalized energy feature vector into the trained one-dimensional convolutional neural network model to output the health status level. If it does not exceed the limit, it is considered to be in a normal state; The model training module performs multi-level wavelet packet decomposition on the squared prediction error time series, including: Perform the squared prediction error time series Layer wavelet packet decomposition, yielding 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 into an energy vector, and the energy vector is normalized to obtain a normalized energy feature vector. The real-time monitoring module projects the real-time sampled signal onto the principal component space to calculate the real-time squared prediction error time series, including: The real-time sampled signal is divided into segments with a fixed number of data points, and each segment serves as an evaluation window; Project the data within each evaluation window onto the principal component space and calculate the squared prediction error statistic corresponding to that 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.

Images