A transformer insulation life prediction method and system based on partial discharge signals

Abstract

The application discloses a transformer insulation life prediction method based on partial discharge signals, and aims to solve the problem of insufficient precision caused by the fact that existing prediction methods do not fully consider environmental factors and noise influence. The method comprises the following steps: sampling transformer monitoring data (partial discharge signals, temperature, load, humidity), obtaining operation sequences through resampling and time alignment; extracting partial discharge statistical values and environmental change values as initial features through a sliding window; calculating the time sequence correlation values of the two and mapping to obtain correction coefficients, and meanwhile, calculating noise weights based on noise levels; weighting and correcting the initial features by using the correction coefficients and the noise weights; inputting the corrected features into an aging model to obtain aging sequences, and obtaining aging features through trend calculation; obtaining stable features through residual calculation and smoothing correction; finally, inputting the stable features into a life model, and outputting prediction results according to the convergence of life distribution variance. The application improves the accuracy and reliability of transformer insulation life prediction.

Description

Technical Field

[0001] This invention relates to the field of transformer life prediction, and more specifically, to a method and system for predicting transformer insulation life based on partial discharge signals. Background Technology

[0002] Transformers are crucial power conversion devices in power systems, and their reliable operation directly impacts grid stability. Transformers have a higher probability of failure towards the end of their insulation life. For economic reasons, maintainers need to understand the actual expected insulation life of transformers to avoid replacing equipment prematurely or too late; they can also anticipate transformer failure risks and prepare for contingencies in advance; and assessing related parts and similar equipment can guide targeted and precise maintenance.

[0003] In related technologies, insulation life prediction is often achieved through threshold determination, empirical formula fitting, or single time series trend analysis, mainly focusing on lifetime estimation based on changes in partial discharge amplitude or aging indicators. However, under actual operating conditions, partial discharge signals are simultaneously affected by temperature changes, load fluctuations, humidity interference, and acquisition noise, resulting in significant dynamic fluctuations in partial discharge characteristics over time. If analysis is based solely on a single discharge characteristic or static aging model, it is difficult to distinguish between short-term fluctuations caused by environmental disturbances and long-term degradation caused by insulation aging.

[0004] The relevant technologies have the technical problem of making it difficult to accurately predict the remaining life of transformer insulation based on partial discharge signals under environmental disturbances and changes in operating conditions. Summary of the Invention

[0005] This invention overcomes the shortcomings of related technologies and provides a method and system for predicting the insulation life of transformers based on partial discharge signals. It can accurately predict the remaining insulation life of transformers based on partial discharge signals under environmental disturbances and changes in operating conditions.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for predicting transformer insulation life based on partial discharge signals includes the following steps: The monitoring data of the sampled transformer is preprocessed to obtain the operating sequence. The preprocessing includes resampling and time alignment. The running sequence is segmented using a sliding window and initial features are extracted, including partial discharge statistics and environmental change values. The time-series correlation between partial discharge statistics and environmental change values ​​is calculated, and a weighted mapping is performed based on the time-series correlation values ​​to obtain the correction coefficients; The noise level is calculated based on the running sequence, and a confidence mapping is performed according to the noise level to obtain the noise weight; The initial features are weighted and corrected according to the correction coefficient and the noise weight to obtain the corrected features; The correction features are input into a preset aging model to obtain an aging sequence, and the trend is calculated based on the aging sequence to obtain aging features. The residuals are calculated based on the aging characteristics to obtain a residual sequence, and the aging characteristics are smoothed and corrected based on the residual sequence to obtain stable characteristics; The stable features are input into a preset lifetime model to obtain a lifetime distribution. The lifetime distribution is then compared with a preset convergence condition to obtain a distribution state. Based on the distribution state, the stable features are calibrated and predicted to obtain a prediction result.

[0007] Preferably, the process of sampling transformer monitoring data and preprocessing it to obtain the operating sequence includes: Acquire the transformer's partial discharge signal, temperature data, load data, and humidity data, along with the timestamps generated at the corresponding acquisition times; The partial discharge signal, temperature data, load data, humidity data, and timestamp are written into the sampling record to obtain the original monitoring sequence; Using the sampling time axis of the partial discharge signal as a reference, the original monitoring sequence is interpolated and resampled to obtain an environmental resampled sequence; The environmental resampling sequence and the partial discharge signal are time-aligned according to the corresponding timestamps to obtain the running sequence.

[0008] Preferably, the process of performing sliding window segmentation on the running sequence and extracting initial features includes: The running sequence is segmented into sliding windows according to a preset window length and a preset sliding step size to obtain the window sequence; Statistical calculations are performed on the partial discharge signals in the window sequence to obtain the mean intensity of partial discharge signals, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency. The mean value of the partial discharge signal intensity, the standard deviation of the partial discharge signal intensity, and the discharge pulse frequency are determined as the statistical value of partial discharge. The rate of change of the window sequence is calculated to obtain the rate of temperature change, the load fluctuation amplitude, and the rate of humidity change. The rate of temperature change, the magnitude of load fluctuation, and the rate of humidity change are defined as the environmental change values. The initial features are obtained by combining the partial discharge statistics and environmental change values ​​corresponding to the same window.

[0009] Preferably, the process of calculating the time-series correlation between partial discharge statistics and environmental change values, and then performing weighted mapping based on the time-series correlation values ​​to obtain the correction coefficients includes: The initial features are arranged in a time sequence to obtain a feature time sequence that includes partial discharge statistics and environmental change values. By performing hysteresis cross-correlation analysis on the characteristic time series, the optimal matching offset between the partial discharge statistics and the environmental change values ​​can be obtained. The feature time sequence is aligned according to the optimal matching offset to obtain the feature matching sequence; The correlation between the partial discharge statistics and the environmental change values ​​in the feature matching sequence is calculated using the Pearson correlation coefficient algorithm to obtain the window correlation value corresponding to each window. The time-series correlation values ​​are obtained by arranging the window-related values ​​in chronological order. The time-series correlation value is input into a preset weight mapping model to obtain the correction coefficient.

[0010] Preferably, the process of calculating the noise level based on the running sequence and performing a confidence mapping based on the noise level to obtain the noise weight includes: Wavelet decomposition is performed on the partial discharge signal in the running sequence to obtain high-frequency components; The amplitude dispersion of the high-frequency component is calculated to obtain the amplitude standard deviation of the high-frequency component, and the amplitude standard deviation is determined as the noise level; The noise level is matched with a preset noise range to determine the noise level corresponding to the noise level. Based on the noise level, the confidence value corresponding to the noise level is extracted from the preset confidence mapping table, and the confidence value is determined as the noise weight.

[0011] Preferably, the process of weighting and correcting the initial features according to the correction coefficient and the noise weight to obtain the corrected features includes: Extract the partial discharge statistics and environmental change values ​​from the initial features; The mean value of partial discharge signal intensity, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency in the partial discharge statistics are corrected and weighted using the correction coefficients to obtain the corrected partial discharge statistics. The noise correction statistics are obtained by weighting the corrected partial discharge statistics with the noise weights to improve their reliability. The noise correction statistics and the environmental change values ​​are vector-concatenated according to a preset feature order to obtain the correction features.

[0012] Preferably, the process of inputting the correction features into a preset aging model to obtain an aging sequence, and performing trend calculation based on the aging sequence to obtain aging features includes: The correction features are input into a preset aging model to obtain the aging index values ​​corresponding to each time window. The aging index values ​​are arranged in chronological order to obtain the aging sequence; Regression fitting is performed on the aging sequence to obtain a trend fitting curve; The slope value is calculated based on the trend fitting curve to obtain the aging trend slope; The aging rate is obtained by calculating the difference in aging index values ​​corresponding to adjacent time windows in the aging sequence. The aging trend slope and the aging rate are defined as the aging characteristics.

[0013] Preferably, the step of calculating residuals based on the aging characteristics to obtain a residual sequence, and then smoothing and correcting the aging characteristics based on the residual sequence to obtain stable characteristics includes: The aging trend slope and aging rate in the aging characteristics are fitted and calculated to obtain fitted characteristic values; The difference between the fitted feature value and the actual feature value in the aging feature is calculated to obtain the residual value; The residual values ​​are arranged in chronological order to obtain the residual sequence; The residual sequence is compared with a preset residual threshold to determine the deviation feature segment; The deviation feature segments are smoothed and corrected by local weighted regression to obtain stable features.

[0014] Preferably, the process of inputting the stable features into a preset lifetime model to obtain a lifetime distribution, comparing the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrating and predicting the stable features based on the distribution state to obtain a prediction result includes: The stable characteristics are input into a preset lifetime model to calculate the dynamic changes in the insulation aging index. The lifetime distribution is obtained by performing distribution fitting based on the dynamic change values ​​and preset historical aging comparison data; The variance of the lifetime distribution is calculated by performing variance calculation on the lifetime distribution. The distribution state is obtained by comparing the variance of the lifetime distribution with a preset convergence threshold. When the variance of the lifetime distribution, which characterizes the distribution state, is higher than the convergence threshold, the stability feature is calibrated to obtain the calibration feature. The calibration features are re-input into the lifetime model to obtain the calibration lifetime distribution, and the mean lifetime corresponding to the calibration lifetime distribution is determined as the prediction result. When the variance of the lifetime distribution, as characterized by the distribution state, is lower than or equal to the convergence threshold, the mean lifetime corresponding to the lifetime distribution is determined as the prediction result.

[0015] A transformer insulation life prediction system based on partial discharge signals includes: The sampling preprocessing module is used to sample the monitoring data of the transformer and preprocess it to obtain the operating sequence. The preprocessing includes resampling and time alignment. The feature extraction module is used to perform sliding window segmentation on the running sequence and extract initial features, including partial discharge statistics and environmental change values. The correlation analysis module is used to calculate the time-series correlation value between the partial discharge statistics and the environmental change value based on the initial characteristics, and to perform weight mapping based on the time-series correlation value to obtain the correction coefficient; The noise assessment module is used to calculate the noise level based on the running sequence and perform a confidence mapping based on the noise level to obtain the noise weight; The feature correction module is used to perform weighted correction on the initial features according to the correction coefficient and the noise weight to obtain the corrected features; The aging analysis module is used to input the correction features into a preset aging model to obtain an aging sequence, and to perform trend calculations based on the aging sequence to obtain aging features. A smoothing correction module is used to perform residual calculation based on the aging characteristics to obtain a residual sequence, and to perform smoothing correction on the aging characteristics based on the residual sequence to obtain stable characteristics; The lifetime prediction module is used to input the stable features into a preset lifetime model to obtain a lifetime distribution, compare the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrate and predict the stable features based on the distribution state to obtain a prediction result.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention can effectively integrate multi-source monitoring data (partial discharge signal, temperature, load, humidity) through sampling, and extract partial discharge statistics and environmental change values ​​in segments through sliding window, which significantly improves the accuracy and comprehensiveness of feature characterization and lays a solid data foundation for subsequent lifetime prediction. (2) Through time-series correlation analysis and noise assessment mechanism, this invention can fully consider the dynamic correlation between partial discharge and environmental factors, and perform credibility weighting according to noise level, effectively reducing the impact of noise interference on prediction results, significantly improving the robustness of prediction model, and solving the problem that the existing technology is unable to distinguish between short-term fluctuations caused by environmental disturbances and long-term degradation caused by insulation aging. (3) This invention constructs a complete prediction closed loop through feature correction, aging analysis, smoothing correction and convergence judgment mechanism, which can realize accurate prediction of transformer insulation life, provide scientific basis for power equipment operation and maintenance decision-making, and has important engineering application value. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the sampling preprocessing process in this invention; Figure 2 This is a flowchart illustrating the feature extraction process in this invention; Figure 3 This is a flowchart illustrating the relevant analysis process in this invention; Figure 4 This is a flowchart illustrating the noise assessment process in this invention; Figure 5 This is a flowchart illustrating the feature correction process in this invention; Figure 6 This is a schematic diagram of the aging analysis process in this invention; Figure 7 This is a flowchart illustrating the smoothing correction process in this invention; Figure 8 This is a flowchart illustrating the lifetime prediction process in this invention; Figure 9 This is a schematic diagram of the system in this invention. Detailed Implementation

[0018] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] Example 1: A method for predicting transformer insulation life based on partial discharge signals includes the following steps: The monitoring data of the sampled transformer is preprocessed to obtain the operating sequence. The preprocessing includes resampling and time alignment. The running sequence is segmented using a sliding window and initial features are extracted, including partial discharge statistics and environmental change values. The time-series correlation between partial discharge statistics and environmental change values ​​is calculated, and a weighted mapping is performed based on the time-series correlation values ​​to obtain the correction coefficients; The noise level is calculated based on the running sequence, and a confidence mapping is performed according to the noise level to obtain the noise weight; The initial features are weighted and corrected according to the correction coefficient and the noise weight to obtain the corrected features; The correction features are input into a preset aging model to obtain an aging sequence, and the trend is calculated based on the aging sequence to obtain aging features. The residuals are calculated based on the aging characteristics to obtain a residual sequence, and the aging characteristics are smoothed and corrected based on the residual sequence to obtain stable characteristics; The stable features are input into a preset lifetime model to obtain a lifetime distribution. The lifetime distribution is then compared with a preset convergence condition to obtain a distribution state. Based on the distribution state, the stable features are calibrated and predicted to obtain a prediction result.

[0021] Among them, such as Figure 1 As shown, the process of sampling transformer monitoring data and preprocessing it to obtain the operating sequence includes: Acquire the transformer's partial discharge signal, temperature data, load data, and humidity data, along with the timestamps generated at the corresponding acquisition times; The partial discharge signal, temperature data, load data, humidity data, and timestamp are written into the sampling record to obtain the original monitoring sequence; Using the sampling time axis of the partial discharge signal as a reference, the original monitoring sequence is interpolated and resampled to obtain an environmental resampled sequence; The environmental resampling sequence and the partial discharge signal are time-aligned according to the corresponding timestamps to obtain the running sequence.

[0022] It is worth noting that the partial discharge signal is output by the partial discharge sensor, the temperature data is output by the temperature sensor, the load data is output by the load monitoring unit, and the humidity data is output by the humidity sensor. In this embodiment, a unified time base is used to record the acquisition time for the partial discharge signal, temperature data, load data, and humidity data. This unified time base is written into the partial discharge sensor, temperature sensor, load monitoring unit, and humidity sensor by the acquisition controller during the installation and commissioning phase. Specifically, in this embodiment, the acquisition controller outputs a standard time. Each sensor receives the standard time, corrects its local timing value, and records the sampling time according to the corrected local timing value during subsequent sampling. The partial discharge sensor synchronously records the corresponding sampling time when outputting the partial discharge signal; the temperature sensor synchronously records the corresponding sampling time when outputting the temperature data; the load monitoring unit synchronously records the corresponding sampling time when outputting the load data; and the humidity sensor synchronously records the corresponding sampling time when outputting the humidity data. Each sampling time is written into the corresponding data record to form the corresponding timestamp.

[0023] It should be noted that the sampling record is a structured record unit established based on individual sample values. Each sampling record includes a data type field, a sample value field, and a corresponding timestamp field. The data type field records the category of the current sample value, which includes partial discharge signal, temperature data, load data, or humidity data. The sample value field records the numerical value of the current sample value, and the corresponding timestamp field records the acquisition time corresponding to the current sample value. In this embodiment, all sampling records are written into a unified record table in ascending order of their corresponding timestamps. The set of sampling records after writing constitutes the original monitoring sequence.

[0024] In one implementation, all corresponding timestamps of the partial discharge signal in the original monitoring sequence are extracted, and the corresponding timestamps are determined as target time nodes. Temperature data subsequences, load data subsequences, and humidity data subsequences are extracted from the original monitoring sequence. For any target time node, two adjacent temperature sampling points located before and after the target time node are found in the temperature data subsequence, and the sampling value, corresponding timestamp, sampling value, and corresponding timestamp of the preceding and following temperature sampling points are read. According to the time ratio between the preceding and following temperature sampling points, the sampling values ​​of the preceding and following temperature sampling points are linearly interpolated to obtain the temperature interpolation result corresponding to the target time node. Using the same interpolation processing method as the temperature data, the load data subsequence and humidity data subsequence are interpolated and resampled to obtain the load interpolation result and humidity interpolation result corresponding to the target time node. In this embodiment, the temperature interpolation result, load interpolation result, and humidity interpolation result corresponding to all target time nodes are arranged in chronological order to obtain the environmental resampling sequence.

[0025] It should be noted that in this embodiment, each target time node corresponding to the partial discharge signal is traversed, and the temperature interpolation result, load interpolation result, and humidity interpolation result that are the same as the current target time node are read from the environmental resampling sequence. In this embodiment, the sampled value of the partial discharge signal, the temperature interpolation result, the load interpolation result, the humidity interpolation result, and the corresponding timestamp corresponding to the current target time node are written into the same alignment record. In this embodiment, all alignment records are arranged in ascending order of the corresponding timestamps to obtain the running sequence.

[0026] like Figure 2 As shown, the process of performing sliding window segmentation and extracting initial features for the running sequence includes: The running sequence is segmented into sliding windows according to a preset window length and a preset sliding step size to obtain the window sequence; Statistical calculations are performed on the partial discharge signals in the window sequence to obtain the mean intensity of partial discharge signals, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency. The mean value of the partial discharge signal intensity, the standard deviation of the partial discharge signal intensity, and the discharge pulse frequency are determined as the statistical value of partial discharge. The rate of change of the window sequence is calculated to obtain the rate of temperature change, the load fluctuation amplitude, and the rate of humidity change. The rate of temperature change, the magnitude of load fluctuation, and the rate of humidity change are defined as the environmental change values. The initial features are obtained by combining the partial discharge statistics and environmental change values ​​corresponding to the same window.

[0027] This embodiment uses an offline calibration method to determine the preset window length and preset sliding step size. The absolute value of the difference between adjacent sampling points in the load data of the running sequence is calculated to obtain a load differential sequence. The 25th percentile of the load differential sequence is determined as the stable differential threshold, and the 75th percentile is determined as the variable differential threshold. In the running sequence, data intervals with continuous load differential values ​​not greater than the stable differential threshold are determined as stable intervals, and data segments that pass through an interval with a load differential value greater than the variable differential threshold and enter another stable interval are determined as continuous data segments. For each continuous data segment, the changes in partial discharge signal intensity, discharge pulse frequency, temperature data, load data, and humidity data are statistically analyzed. For any continuous data segment, this embodiment reads the median of the corresponding statistic within the stable interval before the change as the initial stable value, and reads the median of the corresponding statistic within the stable interval after the change as the final stable value; calculates the difference between the final stable value and the initial stable value to obtain the final stable change; starting from the change point of the continuous data segment, finds the sampling time when the change of the corresponding statistic first reaches 80% of the final stable change, and determines the duration between this sampling time and the change point as the feature response duration; divides the feature response duration by the sampling period of the running sequence to obtain the number of candidate window points; takes the median of the number of candidate window points corresponding to all continuous data segments and rounds it to obtain the preset window length.

[0028] In one implementation, the aforementioned continuous data segments are reused when determining the preset sliding step size. For any continuous data segment, starting from the change point of the continuous data segment, the sampling time when the changes in partial discharge signal intensity, discharge pulse frequency, temperature data, load data, and humidity data first reach 50% of the corresponding final stable change is found, and the duration between this sampling time and the change point is determined as the half-response duration; the half-response duration is divided by the sampling period of the running sequence to obtain the number of candidate sliding step size points; the median of the candidate sliding step size points corresponding to all continuous data segments is taken and rounded to obtain the preset sliding step size; when the preset sliding step size is greater than the preset window length, in this embodiment, the preset window length is determined as the updated preset sliding step size.

[0029] For example, if the number of candidate window points obtained in the offline calibration stage is 8k (k represents 1000), 9k, 9k, 10k, and 11k respectively, then in this embodiment, the median of 9k is determined as the preset window length; if the number of candidate sliding step length points obtained in the offline calibration stage is 4k, 4.5k, 4.5K, 5k, and 5.5k respectively, then in this embodiment, the median of 4.5K is determined as the preset sliding step length.

[0030] It is worth noting that, starting from the first running record in the running sequence, a number of running records equal to the preset window length are read consecutively to obtain the first window data; then, the window's starting position is moved according to the preset sliding step size, and another number of running records equal to the preset window length are read consecutively to obtain the next window data; this window movement and running record reading operation is repeated until the number of remaining running records is less than the preset window length. In this embodiment, all window data are arranged in ascending order of window start time to obtain the window sequence.

[0031] It should be noted that in this embodiment, the background signal of the partial discharge sensor collected during the offline calibration stage is first read, and the mean amplitude and standard deviation of the background signal are calculated. The mean amplitude of the background signal is added to three times the standard deviation of the background signal to obtain the discharge pulse identification threshold. For any window data in the window sequence, the sampled value of the partial discharge signal in the window data is read, and the process of the sampled value of the partial discharge signal crossing the threshold relative to the discharge pulse identification threshold is detected. When the sampled value of the partial discharge signal changes from below the discharge pulse identification threshold to above the discharge pulse identification threshold, a discharge pulse start point is recorded. The maximum amplitude of the sampled value of the partial discharge signal between two adjacent discharge pulse start points is extracted to obtain the peak amplitude of a single discharge pulse event. The mean of the peak amplitudes of all discharge pulse events in the current window is calculated to obtain the mean of the partial discharge signal intensity. The standard deviation of the peak amplitudes of all discharge pulse events in the current window is calculated to obtain the standard deviation of the partial discharge signal intensity. The number of discharge pulse events in the current window is counted, and the number of discharge pulse events is divided by the duration of the current window to obtain the discharge pulse frequency. In this embodiment, the mean value of the partial discharge signal intensity, the standard deviation of the partial discharge signal intensity, and the discharge pulse frequency are determined as the statistical values ​​of partial discharge.

[0032] In one implementation, this embodiment reads the temperature data corresponding to the start time and the end time of the current window, divides the difference between the two by the duration of the current window to obtain the temperature change rate; reads the maximum and minimum values ​​of all load data within the current window, and determines the load fluctuation amplitude by the difference between the maximum and the minimum values; reads the humidity data corresponding to the start time and the end time of the current window, divides the difference between the two by the duration of the current window to obtain the humidity change rate. This embodiment determines the temperature change rate, the load fluctuation amplitude, and the humidity change rate as the environmental change value.

[0033] In one implementation, this embodiment determines the partial discharge statistics and the environmental change values ​​corresponding to the same window as the initial features.

[0034] like Figure 3 As shown, the process of calculating the time-series correlation between partial discharge statistics and environmental change values, and then performing weighted mapping based on the time-series correlation values ​​to obtain the correction coefficients includes: The initial features are arranged in a time sequence to obtain a feature time sequence that includes partial discharge statistics and environmental change values. By performing hysteresis cross-correlation analysis on the characteristic time series, the optimal matching offset between the partial discharge statistics and the environmental change values ​​can be obtained. Align the feature time sequence according to the optimal matching offset to obtain the feature matching sequence; The correlation between the partial discharge statistics and the environmental change values ​​in the feature matching sequence is calculated using the Pearson correlation coefficient algorithm to obtain the window correlation value corresponding to each window. The time-series correlation values ​​are obtained by arranging the window-related values ​​in chronological order. The time-series correlation value is input into a preset weight mapping model to obtain the correction coefficient.

[0035] It is worth noting that in this embodiment, the initial features corresponding to each window are read and arranged in ascending order of window start time to obtain the feature time series. Specifically, a feature time series is generated based on the initial features, and multiple historical data segments in which environmental change values ​​change continuously are extracted from the feature time series. The window position with the largest change in environmental change value and the window position with the largest change in partial discharge statistics value are read from each historical data segment. The number of windows between the two window positions is determined as the offset sample value. The maximum absolute value among all offset sample values ​​is read, and the number of integer windows between the negative maximum absolute value and the positive maximum absolute value is determined as the candidate offset set.

[0036] It is worth noting that in this embodiment, the environmental change value is shifted according to the current candidate offset, and the feature records that are still in the same window position as the partial discharge statistics value after shifting are retained to obtain the candidate alignment sequence corresponding to the current candidate offset. The partial discharge statistics value and environmental change value in the candidate alignment sequence are standardized respectively. The standardization process includes subtracting the mean of each component in the candidate alignment sequence and then dividing by the standard deviation of the component in the candidate alignment sequence. The normalized mean of partial discharge signal intensity, standard deviation of partial discharge signal intensity, and discharge pulse frequency are multiplied point by point with the normalized temperature change rate, load fluctuation amplitude, and humidity change rate, respectively. The absolute value of the product mean of the nine component combinations is taken and averaged to obtain the cross-correlation score corresponding to the current candidate offset. The above cross-correlation score calculation is repeated for all candidate offsets, and the candidate offset with the largest cross-correlation score is determined as the best matching offset.

[0037] The environmental change value is shifted according to the optimal matching offset. The shifted environmental change value is then paired window by window with the partial discharge statistics value at the original window position to obtain multiple matching feature records. Each matching feature record includes the partial discharge statistics value, environmental change value, and window start time at the corresponding window. All matching feature records are arranged in ascending order of window start time to obtain the feature matching sequence.

[0038] It should be noted that, for each current window in the feature matching sequence, the current window and the consecutive matching feature records preceding it are read to obtain a relevant calculation segment; the time length covered by the relevant calculation segment is the same as the time length corresponding to the preset window length during the initial feature extraction process. If the number of matching feature records preceding the current window is insufficient, all matching feature records from the start position of the feature matching sequence to the current window position are read as the relevant calculation segment; if the number of matching feature records in the relevant calculation segment is less than two, the window correlation value of the current window is postponed until the first relevant calculation segment containing at least two matching feature records is formed and then uniformly assigned a value.

[0039] Specifically, the mean value of the partial discharge signal intensity is paired with the rate of temperature change, the load fluctuation amplitude, and the humidity change rate, respectively; the standard deviation of the partial discharge signal intensity is paired with the rate of temperature change, the load fluctuation amplitude, and the humidity change rate, respectively; and the discharge pulse frequency is paired with the rate of temperature change, the load fluctuation amplitude, and the humidity change rate, respectively, to obtain nine component combinations. In this embodiment, the Pearson correlation coefficients of the nine component combinations within the relevant calculation segment are calculated, and the absolute values ​​of the nine Pearson correlation coefficients are taken and averaged to obtain the window correlation value corresponding to the current window.

[0040] It should be noted that the window start time corresponding to each window's relevant value is read, and all window-related values ​​are arranged in ascending order of window start time to obtain the time-series relevant values.

[0041] It is worth noting that in this embodiment, the running sequence and historical aging comparison data are read, the initial features are calculated according to the method of extracting initial features, and the historical time-series correlation values ​​are calculated according to the method described above in this step; the remaining life measured value and the rated design life are read from the historical aging comparison data, the remaining life measured value is divided by the rated design life to obtain the life ratio, and then the life ratio is subtracted from one to obtain the insulation aging index calibration value.

[0042] It should be noted that the preset weight mapping model adopts a single-input, single-output univariate linear regression model. The model input is the historical time-series correlation value, and the model output is the sample correction coefficient. The model parameters include an intercept parameter and a slope parameter. The intercept parameter represents the basic correction amount corresponding to the historical time-series correlation value being zero, and the slope parameter represents the magnitude of the correction amount change corresponding to the historical time-series correlation value changing by one unit. In this embodiment, a candidate correction coefficient set is established, and the value range of the candidate correction coefficient set is from zero to one. The interval between adjacent candidate correction coefficients in the candidate correction coefficient set is determined by the number of historical samples. In this embodiment, one is divided by the number of historical samples to obtain the interval. For each historical sample, each candidate correction coefficient is read sequentially, and the partial discharge statistics in the initial feature are weighted using the current candidate correction coefficient to obtain the candidate correction feature. The least squares method is used to fit the linear relationship between the candidate correction feature and the insulation aging index calibration value, and the fitting error is calculated. The candidate correction coefficient with the smallest fitting error is determined as the sample correction coefficient corresponding to that historical sample.

[0043] Specifically, the historical time-series correlation values ​​corresponding to each historical sample are used as training inputs, and the sample correction coefficients corresponding to each historical sample are used as training outputs to construct training sample pairs. With the optimization objective of minimizing the sum of squared errors between the training output and the model output in the training sample pairs, the intercept parameter and the slope parameter are solved using the least squares method to obtain an initial mapping model. The historical time-series correlation values ​​are input into the initial mapping model to obtain historical prediction correction coefficients, and the mean square error between the historical prediction correction coefficients and the sample correction coefficients is calculated. If the mean square error is not greater than a preset error threshold, the initial mapping model is determined as the preset weight mapping model. If the mean square error is greater than the preset error threshold, training sample pairs with absolute errors greater than the 95th percentile of the error distribution are removed, and the least squares method is used again to solve for the intercept parameter and the slope parameter until the mean square error is not greater than the preset error threshold, thus obtaining the preset weight mapping model.

[0044] The preset error threshold is determined by the output distribution of the training sample pairs. In this embodiment, the variance of the correction coefficients of all samples is calculated, and one-tenth of the variance is determined as the preset error threshold. For example, if the variance of the correction coefficients of all samples is 0.02, then in this embodiment, 0.002 is determined as the preset error threshold; if the mean square error corresponding to the initial mapping model obtained by training is 0.0015, then in this embodiment, the initial mapping model is determined as the preset weight mapping model.

[0045] In one implementation, this embodiment inputs the currently calculated time-series correlation value into the preset weight mapping model to obtain the model output value; if the model output value is less than zero, then zero is determined as the correction coefficient corresponding to the current time node; if the model output value is greater than one, then in this embodiment, one is determined as the correction coefficient corresponding to the current time node; if the model output value is between zero and one, then the model output value is determined as the correction coefficient corresponding to the current time node; the correction coefficients corresponding to each time node are arranged in chronological order to obtain the correction coefficient output in this step.

[0046] like Figure 4 As shown, the process of calculating the noise level based on the running sequence and performing a confidence mapping based on the noise level to obtain the noise weight includes: Wavelet decomposition is performed on the partial discharge signal in the running sequence to obtain high-frequency components; The amplitude dispersion of the high-frequency component is calculated to obtain the amplitude standard deviation of the high-frequency component, and the amplitude standard deviation is determined as the noise level; The noise level is matched with a preset noise range to determine the noise level corresponding to the noise level. Based on the noise level, the confidence value corresponding to the noise level is extracted from the preset confidence mapping table, and the confidence value is determined as the noise weight.

[0047] In one implementation, a partial discharge signal segment corresponding to each window is extracted from the running sequence; discrete wavelet decomposition is performed on each partial discharge signal segment, and the highest frequency detail coefficient sequence is extracted from the wavelet decomposition result to obtain the high frequency component corresponding to that window. The number of decomposition levels for the discrete wavelet decomposition is determined during the offline calibration stage. Partial discharge signal segments in the running sequence are read, and wavelet decomposition is performed on each candidate decomposition level, extracting the highest frequency detail coefficient sequence for each; the amplitude standard deviation of the highest frequency detail coefficient sequence under each candidate decomposition level is calculated, and the fluctuation amplitude of the amplitude standard deviation within the same stable running segment is statistically analyzed; the candidate decomposition level with the smallest fluctuation amplitude is determined as the wavelet decomposition level used in this step. The stable running segment is determined by the stable interval obtained from the offline calibration in the extracted initial features.

[0048] It should be noted that in this embodiment, the amplitude of each sampling point in the high-frequency component corresponding to the current window is read, and the mean of all amplitudes is calculated; the difference between the amplitude of each sampling point and the mean is calculated respectively, and the standard deviation of all differences is calculated to obtain the amplitude standard deviation of the high-frequency component; the amplitude standard deviation of the high-frequency component corresponding to the current window is determined as the noise level of the current window.

[0049] In this embodiment, the running sequence is read and divided into windows according to the preset window length and preset sliding step size determined in the initial feature extraction. For each historical window, the partial discharge signal is subjected to wavelet decomposition and amplitude standard deviation calculation to obtain multiple historical noise levels. All historical noise levels are sorted from smallest to largest value, and the first and second tertiary values ​​in the sorting results are read. The range of values ​​not greater than the first tertiary value is determined as the low noise interval, the range of values ​​greater than the first tertiary value but not greater than the second tertiary value is determined as the medium noise interval, and the range of values ​​greater than the second tertiary value is determined as the high noise interval. The low noise interval, the medium noise interval, and the high noise interval are written into a noise interval table, and the noise interval table is determined as the preset noise interval. For example, if the first tertile of the historical noise levels obtained during the offline calibration stage is 0.03 mV and the second tertile is 0.08 mV, then in this embodiment, the range less than or equal to 0.03 mV is defined as the low noise range, the range greater than 0.03 mV and less than or equal to 0.08 mV is defined as the medium noise range, and the range greater than 0.08 mV is defined as the high noise range.

[0050] In one implementation, this embodiment reads the noise level corresponding to the current window and compares the noise level with the low noise interval, the medium noise interval, and the high noise interval in sequence. When the noise level falls into the low noise interval, this embodiment determines the low noise level as the noise level corresponding to the current window. When the noise level falls into the medium noise interval, this embodiment determines the medium noise level as the noise level corresponding to the current window. When the noise level falls into the high noise interval, this embodiment determines the high noise level as the noise level corresponding to the current window.

[0051] It should be noted that in this embodiment, all historical noise levels obtained when establishing the preset noise interval are read, and the 95th percentile of all historical noise levels is calculated to obtain the noise reference upper limit; the historical noise levels corresponding to low, medium, and high noise levels are read respectively, and the median of the historical noise levels at each of the three noise levels is calculated; the median of the historical noise level corresponding to each noise level is divided by the noise reference upper limit to obtain the noise proportion corresponding to that noise level; the noise proportion is subtracted from one to obtain the confidence value corresponding to that noise level; if the confidence value is less than zero, this embodiment writes zero into the confidence value field corresponding to that noise level; if the confidence value is greater than one, this embodiment writes one into the confidence value field corresponding to that noise level. This embodiment writes the noise level field and the confidence value field into a mapping table to obtain the preset confidence mapping table.

[0052] It is worth noting that in this embodiment, the noise level corresponding to the current window is read, and the confidence value corresponding to the same noise level field is found in the preset confidence mapping table. In this embodiment, the found confidence value is written into the noise weight field corresponding to the current window to obtain the noise weight corresponding to the current window. In this embodiment, the noise level matching and confidence value extraction operations are repeated for all windows, and the noise weights corresponding to all windows are arranged in ascending order according to the window start time to obtain the noise weight output in this step.

[0053] like Figure 5 As shown, the process of obtaining corrected features by weighting and correcting based on initial features includes: Extract the partial discharge statistics and environmental change values ​​from the initial features; The mean value of partial discharge signal intensity, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency in the partial discharge statistics are corrected and weighted using the correction coefficients to obtain the corrected partial discharge statistics. The noise correction statistics are obtained by weighting the corrected partial discharge statistics with the noise weights to improve their reliability. The noise correction statistics and the environmental change values ​​are vector-concatenated according to a preset feature order to obtain the correction features.

[0054] It should be noted that in this embodiment, the initial features corresponding to each window are read, and the partial discharge statistics and environmental change values ​​of each window are read in ascending order of the window's start time.

[0055] Further, in this embodiment, the correction coefficient corresponding to the current window is read, and the mean value of the partial discharge signal intensity, the standard deviation of the partial discharge signal intensity, and the discharge pulse frequency corresponding to the same window are also read. In this embodiment, the mean value of the partial discharge signal intensity is multiplied by the correction coefficient to obtain the corrected mean value of the intensity; the standard deviation of the partial discharge signal intensity is multiplied by the correction coefficient to obtain the corrected standard deviation of the intensity; the discharge pulse frequency is multiplied by the correction coefficient to obtain the corrected pulse frequency; and the corrected mean value of the intensity, the corrected standard deviation of the intensity, and the corrected pulse frequency are determined as the corrected partial discharge statistical value.

[0056] Specifically, in this embodiment, the noise weight corresponding to the current window is read, and the mean of the correction intensity, the standard deviation of the correction intensity, and the correction pulse frequency corresponding to the same window are read; the mean of the correction intensity is multiplied by the noise weight to obtain the mean of the noise correction intensity; the standard deviation of the correction intensity is multiplied by the noise weight to obtain the standard deviation of the noise correction intensity; the correction pulse frequency is multiplied by the noise weight to obtain the noise correction pulse frequency; in this embodiment, the mean of the noise correction intensity, the standard deviation of the noise correction intensity, and the noise correction pulse frequency are determined as the noise correction statistics. For example, if the mean correction intensity of a certain window is 4.32 mV, the standard deviation of the correction intensity is 0.27 mV, the correction pulse frequency is 279 times per second, and the noise weight is 0.80, then in this embodiment, 4.32 mV is multiplied by 0.80 to get 3.46 mV, 0.27 mV is multiplied by 0.80 to get 0.22 mV, and 279 times per second is multiplied by 0.80 to get 223.20 times per second. The above three values ​​are determined as the noise correction statistics corresponding to the window.

[0057] It should be noted that in this embodiment, following the generation order of the initial features extracted from the initial features, the three fields corresponding to the partial discharge statistics are arranged first, followed by the three fields corresponding to the environmental change values. Since the partial discharge statistics have been replaced by the noise correction statistics, this embodiment determines the preset feature order by arranging the fields of the noise correction intensity mean, noise correction intensity standard deviation, noise correction pulse frequency, temperature change rate, load fluctuation amplitude, and humidity change rate. This embodiment writes the preset feature order into a feature configuration table and reads the feature configuration table when generating correction features in each window.

[0058] Specifically, in this embodiment, the mean noise correction intensity, standard deviation of noise correction intensity, noise correction pulse frequency, temperature change rate, load fluctuation amplitude, and humidity change rate corresponding to the current window are read; the above six values ​​are written into the same correction feature record according to the preset feature order, and the correction feature record is determined as the correction feature corresponding to the current window; the above vector splicing operation is repeated for all windows, and all correction features are arranged in ascending order according to the window start time to obtain the correction feature.

[0059] like Figure 6 As shown, the process of inputting the correction features into a preset aging model to obtain an aging sequence, and then performing trend calculations based on the aging sequence to obtain the aging features includes: The correction features are input into a preset aging model to obtain the aging index values ​​corresponding to each time window. The aging index values ​​are arranged in chronological order to obtain the aging sequence; Regression fitting is performed on the aging sequence to obtain a trend fitting curve; The slope value is calculated based on the trend fitting curve to obtain the aging trend slope; The aging rate is obtained by calculating the difference in aging index values ​​corresponding to adjacent time windows in the aging sequence. The aging trend slope and the aging rate are defined as the aging characteristics.

[0060] It should be noted that the preset aging model is a thermal-electric hybrid driven linear accumulation model, and its construction method and training process are as follows.

[0061] Construct thermal stress and electrical stress input terms. For the correction features of each window, obtain the original sampling segment corresponding to that window in the running sequence, extract all temperature data samples in that segment, calculate the average value of these samples, and obtain the average temperature of that window in K.

[0062] The thermal stress term for this window is calculated using the natural constant e as the base and the negative activation energy divided by the product of the Boltzmann constant and the average temperature as the exponent. The activation energy is preset to 0.8 eV, and the Boltzmann constant is taken as 8.617 × 10⁻⁵ eV / K. Simultaneously, the mean noise correction intensity and noise correction pulse frequency are extracted from the noise correction statistics of the same window. The mean noise correction intensity and the noise correction pulse frequency are multiplied, and the product is used as the electrical stress term for this window. The unit of the mean noise correction intensity is mV, the unit of the noise correction pulse frequency is pulses / s, and the unit of the electrical stress term is mV·pulses / s.

[0063] Construct a hybrid driving model expression. The output of the preset aging model is a windowed aging index value, obtained by multiplying the thermal stress term by the thermal stress weighting coefficient, adding the electrical stress term by the electrical stress weighting coefficient, and adding the basic aging intercept term. The thermal stress weighting coefficient, electrical stress weighting coefficient, and basic aging intercept term are the parameters that the model needs to solve for.

[0064] The model parameters are solved and verified. The running sequence is read, and historical correction features are generated according to the methods in steps S11 to S15, and the window start time of each window is recorded. Historical aging comparison data is read, which includes transformer equipment number, detection time, measured value of remaining life, and rated design life. For each detection time, in the window sequence corresponding to the historical correction feature, the window whose window start time is closest to that detection time is found. The historical thermal stress term and historical electrical stress term of that window are used as training inputs for that detection time, and the calibration value of the insulation aging index corresponding to that detection time is used as training output to construct training samples.

[0065] The calculation method for the insulation aging index calibration value is as follows: divide the measured remaining life by the rated design life to obtain the life ratio, and then subtract the life ratio from 1. All training samples are sorted in ascending order of detection time, with the first 80% of samples forming the training set and the last 20% forming the validation set.

[0066] The thermal stress weighting coefficient, electrical stress weighting coefficient, and basic aging intercept term are solved on the training set using the least squares method to obtain the initial model parameters. The historical thermal stress and electrical stress terms for each sample in the validation set are input into the initial model to obtain the validation aging index values. The mean absolute error between the validation aging index values ​​and the corresponding calibrated insulation aging index values ​​is calculated. The range of the calibrated insulation aging index values ​​in the validation set is calculated, and one-tenth of the range is determined as the model validation threshold. For example, when the range is 0.7, the model validation threshold is 0.07.

[0067] When the mean absolute error (MAE) is not greater than the model validation threshold, the current model parameters are retained, and the model containing these parameters is designated as the preset aging model. When the MAE is greater than the model validation threshold, the training absolute error for each sample in the training set is calculated, and samples with training absolute errors greater than the 95th percentile are removed to obtain an updated training set. The least squares method is then used again to solve for the model parameters on the updated training set. This iterative process is repeated until the MAE calculated on the validation set is not greater than the model validation threshold, thus obtaining the preset aging model.

[0068] For the transformer to be evaluated, the correction characteristics of each window are read, and the thermal stress and electrical stress terms for each window are generated in the same way as described above. The thermal stress and electrical stress terms are then input into the pre-trained aging model to calculate the aging index values ​​corresponding to each time window. All aging index values ​​are arranged in ascending order of the window start time to obtain the aging sequence.

[0069] The aging sequence was fitted using linear regression with the least squares method to obtain a trend curve. This trend curve is a linear function with window number as the independent variable and aging index value as the dependent variable; the slope of this linear function is the aging trend slope. The difference in aging index values ​​between each pair of adjacent time windows in the aging sequence was calculated. Subtracting the aging index value of the previous window from the value of the aging index in the later window yielded the aging rate corresponding to that adjacent window. The aging trend slope and the aging rate corresponding to each window were then defined as aging characteristics.

[0070] like Figure 7 As shown, the process of calculating residuals based on the aging characteristics to obtain a residual sequence, and then smoothing and correcting the aging characteristics based on the residual sequence to obtain stable characteristics includes: The aging trend slope and aging rate in the aging characteristics are fitted and calculated to obtain fitted characteristic values; The difference between the fitted feature value and the actual feature value in the aging feature is calculated to obtain the residual value; The residual values ​​are arranged in chronological order to obtain the residual sequence; The residual sequence is compared with a preset residual threshold to determine the deviation feature segment; The deviation feature segments are smoothed and corrected by local weighted regression to obtain stable features.

[0071] Specifically, in this embodiment, the aging characteristics are read, including the aging trend slope and the aging rate; continuous time numbers are generated according to the start time of the window corresponding to the aging rate, and the aging trend slope is read as the fitting slope; each time number is multiplied by the fitting slope to obtain the slope term corresponding to each time number; each aging rate is subtracted from the corresponding slope term to obtain multiple intercept candidate values, and the average value of all intercept candidate values ​​is calculated to obtain the fitting intercept; the slope term corresponding to each time number is added to the fitting intercept to obtain the fitting feature value corresponding to each time window.

[0072] It is worth noting that in this embodiment, the aging rate in the aging characteristics is determined as the actual characteristic value; for any time window, this embodiment reads the actual characteristic value and the fitted characteristic value corresponding to the time window, and subtracts the fitted characteristic value from the actual characteristic value to obtain the residual value corresponding to the time window.

[0073] It should be noted that in this embodiment, the window start time corresponding to each residual value is read, and all residual values ​​are arranged in ascending order of window start time to obtain the residual sequence. If the window start times corresponding to two adjacent residual values ​​are the same, this embodiment sorts them according to the order of the corresponding windows in the aging rate.

[0074] In one implementation, this embodiment reads the running sequence, generates historical aging features, and obtains historical residual values ​​according to the residual value calculation method in this step; this embodiment takes the absolute value of all historical residual values ​​to obtain the absolute value of historical residuals; this embodiment extracts historical samples with stable equipment insulation status from historical aging comparison data, wherein the historical samples with stable equipment insulation status refer to samples where the absolute value of the difference between the calibration values ​​of insulation aging indexes in two adjacent detection records is less than the 25th percentile of the historical insulation aging index difference; this embodiment extracts the absolute value of historical residuals corresponding to the historical samples with stable equipment insulation status and arranges them in ascending order of value; this embodiment reads the 95th percentile in the sorting result and determines the absolute value of historical residuals corresponding to the 95th percentile as the preset residual threshold.

[0075] Specifically, in this embodiment, the absolute value of each residual value in the residual sequence is taken to obtain the absolute residual value; each absolute residual value is compared with the preset residual threshold; when the absolute residual value corresponding to a certain time window is greater than the preset residual threshold, this embodiment determines that time window as a deviation window; this embodiment merges temporally continuous deviation windows to obtain a deviation feature segment. If there is a non-deviation window between two deviation windows, this embodiment divides the two deviation windows into different deviation feature segments.

[0076] It should be noted that, in this embodiment, for any deviation feature segment, all deviation windows within the deviation feature segment are read, and the non-deviation window preceding and following the deviation feature segment is also read; if the deviation feature segment is located at the beginning of the sequence, this embodiment only reads the first non-deviation window following the deviation feature segment; if the deviation feature segment is located at the end of the sequence, this embodiment only reads the first non-deviation window preceding the deviation feature segment; in this embodiment, all read windows are determined as the local regression window set corresponding to the current deviation feature segment.

[0077] In one implementation, this embodiment reads the time index and aging rate of each window in the local regression window set; for any window to be corrected within the deviation feature segment, this embodiment calculates the time distance between the time index of the window to be corrected and the time index of each window in the local regression window set, and reads the maximum value among all time distances as the distance benchmark; this embodiment divides each time distance by the distance benchmark to obtain the distance ratio; the result obtained by subtracting the cube of the distance ratio is then cubed to obtain the local weight corresponding to each window; this embodiment inputs the local weight, the time index of each window, and the aging rate of each window into the weighted least squares solution process to obtain the local linear fitting relationship; this embodiment substitutes the time index of the window to be corrected into the local linear fitting relationship to obtain the corrected aging rate corresponding to the window to be corrected; this embodiment repeats the above local weighted regression processing for all windows to be corrected within the deviation feature segment to obtain the corrected aging rate sequence.

[0078] It should be noted that in this embodiment, the aging rate within the deviation feature segment is replaced with the corresponding corrected aging rate to obtain the updated aging rate; the slope value of the updated aging rate is recalculated using the trend fitting method in the step of obtaining the aging feature to obtain the corrected trend slope; the corrected trend slope and the updated aging rate are written into the corrected feature record to obtain the corrected feature. The aging rate that does not fall into the deviation feature segment retains its original value.

[0079] In one implementation, this embodiment reads the correction trend slope and the updated aging rate from the correction feature, writes the correction trend slope into the trend field of the stable feature record, and writes the updated aging rate into the rate field of the stable feature record to obtain the stable feature.

[0080] like Figure 8 As shown, the process of inputting the stable features into a preset lifetime model to obtain a lifetime distribution, comparing the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrating and predicting the stable features based on the distribution state to obtain a prediction result includes: The stable characteristics are input into a preset lifetime model to calculate the dynamic changes in the insulation aging index. The lifetime distribution is obtained by performing distribution fitting based on the dynamic change values ​​and preset historical aging comparison data; The variance of the lifetime distribution is calculated by performing variance calculation on the lifetime distribution. The distribution state is obtained by comparing the variance of the lifetime distribution with a preset convergence threshold. When the variance of the lifetime distribution, which characterizes the distribution state, is higher than the convergence threshold, the stability feature is calibrated to obtain the calibration feature. The calibration features are re-input into the lifetime model to obtain the calibration lifetime distribution, and the mean lifetime corresponding to the calibration lifetime distribution is determined as the prediction result. When the variance of the lifetime distribution, as characterized by the distribution state, is lower than or equal to the convergence threshold, the mean lifetime corresponding to the lifetime distribution is determined as the prediction result.

[0081] In one implementation, this embodiment reads the running sequence and generates historical stable features; reads historical aging comparison data, which includes the transformer equipment number, detection time, measured remaining life, rated design life, and insulation aging index calibration values ​​at adjacent detection times; subtracts the insulation aging index calibration value of the previous detection time from the insulation aging index calibration value of the later detection time to obtain the change in insulation aging index; divides the change in insulation aging index by the time interval between two adjacent detection times to obtain the dynamic change calibration value of insulation aging index; and maps the historical stable features corresponding to the same detection time to the dynamic change calibration value of insulation aging index to obtain life model training samples.

[0082] In one implementation, the lifespan model employs a multiple linear regression model. The input to the multiple linear regression model is stable features, including a trend field and a rate field from the stable feature records. The output of the multiple linear regression model is the dynamic change value of the insulation aging index. The model parameters of the multiple linear regression model include trend field coefficients, rate field coefficients, and a lifespan model intercept parameter. Historical stable features from the lifespan model training samples are used as training input, and the corresponding dynamic change calibration values ​​of the insulation aging index are used as training output. In this embodiment, the least squares method is used to solve for the trend field coefficients, rate field coefficients, and lifespan model intercept parameter to obtain the trained multiple linear regression model, which is then determined as the preset lifespan model.

[0083] In one implementation, this embodiment arranges all lifetime model training samples in ascending order of detection time, and determines the first 80% of lifetime model training samples as the training set and the last 20% as the validation set. This embodiment uses the training set to solve for the trend field coefficient, the rate field coefficient, and the lifetime model intercept parameter. This embodiment inputs the historical stability features from the validation set into the preset lifetime model to obtain the validation dynamic change value. This embodiment calculates the mean absolute error between the validation dynamic change value and the dynamic change calibration value of the insulation aging index corresponding to the validation set. This embodiment calculates the mean absolute error between the validation dynamic change value and the dynamic change calibration value of the insulation aging index corresponding to the validation set. The range of the dynamic change calibration values ​​of the aging index is determined, and one-tenth of the range is set as the lifetime model verification threshold. If the mean absolute error is not greater than the lifetime model verification threshold, this embodiment retains the trend field coefficients, rate field coefficients, and lifetime model intercept parameters obtained from the current solution. If the mean absolute error is greater than the lifetime model verification threshold, this embodiment removes training samples whose absolute errors are greater than the 95th percentile of the verification set error, and re-solves the trend field coefficients, rate field coefficients, and lifetime model intercept parameters using the least squares method until the mean absolute error is not greater than the lifetime model verification threshold, thus obtaining the preset lifetime model.

[0084] Specifically, in this embodiment, the trend field and rate field in the stable feature are read; the trend field is multiplied by the trend field coefficient to obtain the trend contribution value; the rate field is multiplied by the rate field coefficient to obtain the rate contribution value; the trend contribution value, the rate contribution value and the lifetime model intercept parameter are added together to obtain the dynamic change value of the insulation aging index.

[0085] This embodiment reads historical maintenance records and historical operation records of transformers of the same model. The historical maintenance records include the transformer equipment number, inspection time, measured remaining life, and rated design life. The historical operation records include historical stability characteristics corresponding to the inspection time. This embodiment matches the historical maintenance records and historical operation records according to the inspection time. This embodiment calculates the historical dynamic change value for each historical operation record according to the method for calculating the dynamic change value of insulation aging index in this step. This embodiment writes the historical dynamic change value, measured remaining life, historical stability characteristics, and inspection time into the same historical aging record. This embodiment writes all historical aging records into the historical aging database in ascending order of inspection time to obtain the preset historical aging comparison data.

[0086] It is worth noting that in this embodiment, each historical dynamic change value and each measured remaining lifetime value are read from the preset historical aging comparison data; the absolute difference between the current dynamic change value and each historical dynamic change value is calculated respectively; the maximum value among all absolute differences is read as the difference benchmark; if the difference benchmark is zero, this embodiment assigns the same weight to all historical aging records; if the difference benchmark is not zero, this embodiment divides each absolute difference by the difference benchmark to obtain a normalized difference, and the result of dividing one by one and adding the normalized difference is determined as the lifetime sample weight of the historical aging record. This embodiment associates each measured remaining lifetime value with the corresponding lifetime sample weight to obtain a weighted remaining lifetime sample set; this embodiment determines the weighted remaining lifetime sample set as the lifetime distribution.

[0087] This embodiment reads all remaining lifetime measured values ​​and corresponding lifetime sample weights from the lifetime distribution; normalizes all lifetime sample weights using a weighting and normalization method to obtain normalized lifetime sample weights; multiplies each remaining lifetime measured value by its corresponding normalized lifetime sample weight, and sums all the product results to obtain the lifetime mean; calculates the squared difference between each remaining lifetime measured value and the lifetime mean, multiplies each squared difference by its corresponding normalized lifetime sample weight, and sums all the weighted squared differences to obtain the lifetime distribution variance.

[0088] In one implementation, this embodiment reads preset historical aging comparison data and divides the preset historical aging comparison data into training samples and validation samples according to the detection time. The training samples are the first 80% of historical aging records, and the validation samples are the last 20% of historical aging records. For each validation sample, it is used as the sample to be validated, and the remaining validation samples are used as historical reference samples. According to the method for obtaining the lifetime distribution in this step, a validation lifetime distribution is generated based on the historical dynamic change value of the sample to be validated and the historical reference samples. The variance of the validation lifetime distribution is calculated to obtain the variance of the validation lifetime distribution. The validation lifetime mean corresponding to the validation lifetime distribution is calculated, and the absolute error between the validation lifetime mean and the measured value of the remaining lifetime of the sample to be validated is calculated to obtain the validation prediction error. All validation prediction errors are arranged in ascending order of value, and the median in the sorting result is read. Validation samples with an absolute error not greater than the median are determined as stable prediction samples. The variance of the validation lifetime distribution corresponding to all stable prediction samples is read, and the maximum value among them is determined as the preset convergence threshold.

[0089] Specifically, when the lifetime distribution variance is higher than the preset convergence threshold, this embodiment marks the distribution state as non-converged; when the lifetime distribution variance is lower than or equal to the preset convergence threshold, the distribution state is marked as converged.

[0090] This embodiment reads each historical stable feature and each lifetime sample weight from the preset historical aging comparison data; this embodiment performs weighting and normalization processing on all lifetime sample weights, multiplies the trend field in each historical stable feature by the corresponding normalized lifetime sample weight, and adds all the product results to obtain the weighted historical trend field; multiplies the rate field in each historical stable feature by the corresponding normalized lifetime sample weight, and adds all the product results to obtain the weighted historical rate field; averages the trend field in the current stable feature with the weighted historical trend field to obtain the calibration trend field; averages the rate field in the current stable feature with the weighted historical rate field to obtain the calibration rate field; writes the calibration trend field and the calibration rate field into the calibration feature record to obtain the calibration feature.

[0091] Specifically, in this embodiment, following the method for calculating the dynamic change value of the insulation aging index in this step, the calibration feature is input into the lifetime model to obtain the calibration dynamic change value; following the method for obtaining the lifetime distribution in this step, the calibration dynamic change value and the preset historical aging comparison data are used to perform distribution fitting to obtain the calibration lifetime distribution; following the method for calculating the lifetime mean in this step, the lifetime mean corresponding to the calibration lifetime distribution is calculated, and the lifetime mean corresponding to the calibration lifetime distribution is determined as the prediction result. The lifetime mean is read and written into the prediction result record.

[0092] In summary, this invention discloses a method for predicting the remaining life of transformer insulation based on partial discharge signals. By aligning the time of multi-source monitoring data and dynamically correcting the partial discharge characteristics, and establishing a physical mapping relationship between partial discharge characteristics and insulation aging state based on a thermo-electric hybrid driven aging model, and then through residual smoothing and life distribution convergence calibration, a stable and accurate prediction of the remaining life of transformer insulation is finally achieved.

[0093] Example 2: like Figure 9 As shown, a transformer insulation life prediction system based on partial discharge signals is provided, comprising: The sampling preprocessing module is used to sample the monitoring data of the transformer and preprocess it to obtain the operating sequence. The preprocessing includes resampling and time alignment. The feature extraction module is used to perform sliding window segmentation on the running sequence and extract initial features, including partial discharge statistics and environmental change values. The correlation analysis module is used to calculate the time-series correlation value between the partial discharge statistics and the environmental change value based on the initial characteristics, and to perform weight mapping based on the time-series correlation value to obtain the correction coefficient; The noise assessment module is used to calculate the noise level based on the running sequence and perform a confidence mapping based on the noise level to obtain the noise weight; The feature correction module is used to perform weighted correction on the initial features according to the correction coefficient and the noise weight to obtain the corrected features; The aging analysis module is used to input the correction features into a preset aging model to obtain an aging sequence, and to perform trend calculations based on the aging sequence to obtain aging features. A smoothing correction module is used to perform residual calculation based on the aging characteristics to obtain a residual sequence, and to perform smoothing correction on the aging characteristics based on the residual sequence to obtain stable characteristics; The lifetime prediction module is used to input the stable features into a preset lifetime model to obtain a lifetime distribution, compare the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrate and predict the stable features based on the distribution state to obtain a prediction result.

[0094] It should be noted that the transformer insulation life prediction system based on partial discharge signal provided in this embodiment of the invention is used to execute all the process steps of the transformer insulation life prediction method based on partial discharge signal in the above embodiment. The working principle and beneficial effect of the two correspond one-to-one, so they will not be described again.

[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications may be made without departing from the technical solutions described in the claims.

Claims

1. A method for predicting the insulation life of a transformer based on partial discharge signals, characterized in that, Includes the following steps: The monitoring data of the sampled transformer is preprocessed to obtain the operating sequence. The preprocessing includes resampling and time alignment. The running sequence is segmented using a sliding window and initial features are extracted, including partial discharge statistics and environmental change values. The time-series correlation between partial discharge statistics and environmental change values ​​is calculated, and a weighted mapping is performed based on the time-series correlation values ​​to obtain the correction coefficients; The noise level is calculated based on the running sequence, and a confidence mapping is performed according to the noise level to obtain the noise weight; The initial features are weighted and corrected according to the correction coefficient and the noise weight to obtain the corrected features; The correction features are input into a preset aging model to obtain an aging sequence, and the trend is calculated based on the aging sequence to obtain aging features. The residuals are calculated based on the aging characteristics to obtain a residual sequence, and the aging characteristics are smoothed and corrected based on the residual sequence to obtain stable characteristics; The stable features are input into a preset lifetime model to obtain a lifetime distribution. The lifetime distribution is then compared with a preset convergence condition to obtain a distribution state. Based on the distribution state, the stable features are calibrated and predicted to obtain a prediction result.

2. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of sampling transformer monitoring data and preprocessing it to obtain the operating sequence includes: Acquire the transformer's partial discharge signal, temperature data, load data, and humidity data, along with the timestamps generated at the corresponding acquisition times; The partial discharge signal, temperature data, load data, humidity data, and timestamp are written into the sampling record to obtain the original monitoring sequence; Using the sampling time axis of the partial discharge signal as a reference, the original monitoring sequence is interpolated and resampled to obtain an environmental resampled sequence; The environmental resampling sequence and the partial discharge signal are time-aligned according to their corresponding timestamps to obtain the running sequence.

3. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of performing sliding window segmentation on the running sequence and extracting initial features includes: The running sequence is segmented into sliding windows according to a preset window length and a preset sliding step size to obtain the window sequence; Statistical calculations are performed on the partial discharge signals in the window sequence to obtain the mean intensity of partial discharge signals, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency. The mean value of the partial discharge signal intensity, the standard deviation of the partial discharge signal intensity, and the discharge pulse frequency are determined as the statistical value of partial discharge. The rate of change of the window sequence is calculated to obtain the rate of temperature change, the load fluctuation amplitude, and the rate of humidity change. The rate of temperature change, the magnitude of load fluctuation, and the rate of humidity change are defined as the environmental change values. The initial features are obtained by combining the partial discharge statistics and environmental change values ​​corresponding to the same window.

4. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of calculating the time-series correlation between partial discharge statistics and environmental change values, and then performing weighted mapping based on these correlation values ​​to obtain correction coefficients, includes: The initial features are arranged in a time sequence to obtain a feature time sequence that includes partial discharge statistics and environmental change values. By performing hysteresis cross-correlation analysis on the characteristic time series, the optimal matching offset between the partial discharge statistics and the environmental change values ​​can be obtained. Align the feature time sequence according to the optimal matching offset to obtain the feature matching sequence; The correlation between the partial discharge statistics and the environmental change values ​​in the feature matching sequence is calculated using the Pearson correlation coefficient algorithm to obtain the window correlation value corresponding to each window. The time-series correlation values ​​are obtained by arranging the window-related values ​​in chronological order. The time-series correlation value is input into a preset weight mapping model to obtain the correction coefficient.

5. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of calculating the noise level based on the running sequence and mapping the confidence level to obtain the noise weight includes: Wavelet decomposition is performed on the partial discharge signal in the running sequence to obtain high-frequency components; The amplitude dispersion of the high-frequency component is calculated to obtain the amplitude standard deviation of the high-frequency component, and the amplitude standard deviation is determined as the noise level; The noise level is matched with a preset noise range to determine the noise level corresponding to the noise level. Based on the noise level, the confidence value corresponding to the noise level is extracted from the preset confidence mapping table, and the confidence value is determined as the noise weight.

6. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of weighting and correcting the initial features according to the correction coefficient and the noise weight to obtain the corrected features includes: Extract the partial discharge statistics and environmental change values ​​from the initial features; The mean value of partial discharge signal intensity, the standard deviation of partial discharge signal intensity, and the discharge pulse frequency in the partial discharge statistics are corrected and weighted using the correction coefficients to obtain the corrected partial discharge statistics. The noise correction statistics are obtained by weighting the corrected partial discharge statistics with the noise weights to improve their reliability. The noise correction statistics and the environmental change values ​​are vector-concatenated according to a preset feature order to obtain the correction features.

7. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of inputting the correction features into a preset aging model to obtain an aging sequence, and then performing trend calculations based on the aging sequence to obtain the aging features includes: The correction features are input into a preset aging model to obtain the aging index values ​​corresponding to each time window. The aging index values ​​are arranged in chronological order to obtain the aging sequence; Regression fitting is performed on the aging sequence to obtain a trend fitting curve; The slope value is calculated based on the trend fitting curve to obtain the aging trend slope; The aging rate is obtained by calculating the difference in aging index values ​​corresponding to adjacent time windows in the aging sequence. The aging trend slope and the aging rate are defined as the aging characteristics.

8. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of calculating residuals based on the aging characteristics to obtain a residual sequence, and then smoothing and correcting the aging characteristics based on the residual sequence to obtain stable characteristics includes: The aging trend slope and aging rate in the aging characteristics are fitted and calculated to obtain fitted characteristic values; The difference between the fitted feature value and the actual feature value in the aging feature is calculated to obtain the residual value; The residual values ​​are arranged in chronological order to obtain the residual sequence; The residual sequence is compared with a preset residual threshold to determine the deviation feature segment; The deviation feature segments are smoothed and corrected by local weighted regression to obtain stable features.

9. The method for predicting transformer insulation life based on partial discharge signals according to claim 1, characterized in that, The process of inputting the stable features into a preset lifetime model to obtain a lifetime distribution, comparing the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrating and predicting the stable features based on the distribution state to obtain the prediction result includes: The stable characteristics are input into a preset lifetime model to calculate the dynamic changes in the insulation aging index. The lifetime distribution is obtained by performing distribution fitting based on the dynamic change values ​​and preset historical aging comparison data; The variance of the lifetime distribution is calculated by performing variance calculation on the lifetime distribution. The distribution state is obtained by comparing the variance of the lifetime distribution with a preset convergence threshold. When the variance of the lifetime distribution, which characterizes the distribution state, is higher than the convergence threshold, the stability feature is calibrated to obtain the calibration feature. The calibration features are re-input into the lifetime model to obtain the calibration lifetime distribution, and the mean lifetime corresponding to the calibration lifetime distribution is determined as the prediction result. When the variance of the lifetime distribution, as characterized by the distribution state, is lower than or equal to the convergence threshold, the mean lifetime corresponding to the lifetime distribution is determined as the prediction result.

10. A transformer insulation life prediction system based on partial discharge signals, characterized in that, include: The sampling preprocessing module is used to sample the monitoring data of the transformer and preprocess it to obtain the operating sequence. The preprocessing includes resampling and time alignment. The feature extraction module is used to perform sliding window segmentation on the running sequence and extract initial features, including partial discharge statistics and environmental change values. The correlation analysis module is used to calculate the time-series correlation value between the partial discharge statistics and the environmental change value based on the initial characteristics, and to perform weight mapping based on the time-series correlation value to obtain the correction coefficient; The noise assessment module is used to calculate the noise level based on the running sequence and perform a confidence mapping based on the noise level to obtain the noise weight; The feature correction module is used to perform weighted correction on the initial features according to the correction coefficient and the noise weight to obtain the corrected features; The aging analysis module is used to input the correction features into a preset aging model to obtain an aging sequence, and to perform trend calculations based on the aging sequence to obtain aging features. A smoothing correction module is used to perform residual calculation based on the aging characteristics to obtain a residual sequence, and to perform smoothing correction on the aging characteristics based on the residual sequence to obtain stable characteristics; The lifetime prediction module is used to input the stable features into a preset lifetime model to obtain a lifetime distribution, compare the lifetime distribution with a preset convergence condition to obtain a distribution state, and calibrate and predict the stable features based on the distribution state to obtain a prediction result.

Images