A method for diagnosing partial discharge insulation defects of a high voltage cable

Abstract

The application discloses a kind of high-voltage cable partial discharge insulation defect model diagnostic methods, to solve the deficiency that existing technology is difficult to find in the early stage of high-voltage cable insulation failure.The present application comprises the following steps: input the original voltage sampling signal data of high-voltage cable to be detected;Noise in signal is cleaned and partial discharge peak is screened by wavelet transform;Multi-dimensional features of partial discharge peak are extracted to form measurement group fusion feature vector;The feature vector is input into the pre-trained partial discharge insulation deterioration model to obtain the insulation deterioration probability;Based on the preset calibration standard, the state of high-voltage cable is determined by logical judgment.The present application uses wavelet transform and inflection point detection algorithm to improve signal detection accuracy, improves diagnostic accuracy through multi-dimensional feature fusion and LightGBM integrated learning framework, reduces false positive rate by combining sliding time window verification mechanism, and provides reliable technical support for high-voltage cable insulation state monitoring.

Description

Technical Field

[0001] This invention relates to the field of high-voltage cable discharge defect identification, and more specifically, to a method for diagnosing partial discharge insulation defects in high-voltage cables. Background Technology

[0002] High-voltage cables, as key equipment in urban power grids, directly affect the safety and stability of the power system. With economic development and continuous growth in electricity load, the scale of high-voltage cable equipment has expanded rapidly. The operating status of high-voltage cables directly impacts the safety and stability of the power system. Insulation failure in high-voltage cables essentially means the insulation medium is broken down, losing its ability to isolate and conduct electricity. This can trigger a series of serious consequences, from equipment damage to power grid accidents, personal safety issues, and social impacts. When insulation fails, high voltage directly breaks down the insulation, forming a conductive path and causing a violent electric arc, which can break down the cable and its connections, affecting or even destroying their function. When one phase of the cable is grounded, it causes a single-phase ground fault in the power grid, resulting in a surge in zero-sequence current and voltage deviation. Furthermore, when the insulation completely fails, all three phases become completely conductive, representing one of the most serious power faults, potentially affecting the power supply of the entire city and causing secondary fires and personal accidents.

[0003] Existing insulation testing methods are insufficient for timely detection of faults in high-voltage cables at their initial stages. There is an urgent need for a diagnostic method for partial discharge insulation defects in high-voltage cables, capable of timely detection and intervention at the onset of partial discharge symptoms and in their early stages, thus preventing serious consequences from insulation failure. Summary of the Invention

[0004] This invention overcomes the problem of existing technologies making it difficult to detect partial discharge defects in the early stages of high-voltage cable faults in a timely manner. It provides a diagnostic method for partial discharge insulation defects in high-voltage cables, which can accurately diagnose the insulation defect status of high-voltage cables, detect and intervene in a timely manner, and avoid serious consequences caused by insulation failure.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A diagnostic method for partial discharge insulation defects in high-voltage cables includes the following steps: S1. Input the original voltage sampling signal data of the high-voltage cable to be tested; S2. Clean the noise signal in the original voltage sampling signal data and filter out the partial discharge peak value by wavelet transform; S3. Extract the multi-dimensional features of the partial discharge peak to form a measurement group fusion feature vector; S4. Input the fused feature vector of the measurement group into the pre-trained partial discharge insulation degradation model to obtain the insulation degradation probability; S5. Based on preset verification standards, the condition of the high-voltage cable is determined by logical judgment of the insulation degradation probability.

[0006] Preferably, step S2 includes: S21. Wavelet function is used to perform a 5-level tower decomposition on the original voltage sampling signal data, which is decomposed into approximation coefficients and detail coefficients. S22. Apply a soft threshold to the detail coefficients, wherein the threshold is adaptively determined by the noise standard deviation; S23. Based on the processed wavelet coefficients, the reconstructed denoised signal is obtained by inverse wavelet transform. S24. Take the absolute value of the reconstructed denoised signal and identify candidate peak points using a local maximum detection algorithm; S25. Separate the signal and noise using the inflection point detection algorithm to obtain the partial discharge peak value.

[0007] Preferably, in step S24, the local maximum point detection adopts a local maximum detection algorithm with a window length of 25.

[0008] As a preferred embodiment, the inflection point detection algorithm in step S25 includes: arranging the candidate peaks in descending order of amplitude, calculating the amplitude gradient between adjacent peaks, performing convolution smoothing on the gradient sequence with a length of 9, using the plateau detection method to identify the inflection point of the noise floor, and retaining the peaks before the inflection point as partial discharge peaks.

[0009] Preferably, the measurement group fusion feature vector in step S3 includes: the amplitude ratio of adjacent sampling points, the root mean square error of sawtooth wave fitting, frequency domain phase alignment features, and statistical aggregation features. The measurement group-level fusion feature vector is formed by vector concatenation. The statistical aggregation features include the statistical values ​​of the number of peaks in the phase intervals of 0°-180° and 180°-360°, the mean and standard deviation of peak height, the mean amplitude ratio of adjacent sampling points, the mean root mean square error of sawtooth wave fitting, and the average offset of the distance between the peak and the zero-crossing point of the inverted phase.

[0010] Preferably, the frequency domain phase alignment feature obtains the phase angle by calculating the 50Hz fundamental frequency Fourier coefficients, maps the peak position to the phase space and divides it into four quadrants to achieve phase resolution analysis.

[0011] Preferably, the partial discharge insulation degradation model in step S4 is trained through the following steps: S1. Obtain training data composed of signal-level data, group the signal-level data based on the measurement group identifier, perform logical or rule processing based on the insulation degradation label of the signal-level data within the measurement group, construct the insulation degradation label at the measurement group level, and form a measurement group discrimination feature vector. S2. Based on the discriminative feature vectors of the measurement group, construct and train an ensemble learning framework based on LightGBM.

[0012] Preferably, in step S5, when the probability of insulation degradation is greater than a preset threshold, the high-voltage cable is determined to be in a degraded state. This also includes a sliding time window verification mechanism, which only determines a latent fault and triggers an early warning when the probability of insulation defects in the same measurement group is continuously higher than the optimal threshold within a preset time period. The duration threshold is set to 3-5 monitoring cycles.

[0013] As a preferred option, the ensemble learning framework generates 125 independent LightGBM classifiers through random seed iterations and is trained using three-fold cross-validation. The model is configured with a learning rate of 0.01, 80 leaf nodes, a feature sampling rate of 0.8, a sample sampling rate of 0.8, a maximum iteration count of 10,000, and an early stopping patience value of 100.

[0014] As a preferred method, the insulation degradation probability is obtained through a probabilistic ensemble method, which uses the arithmetic mean method to fuse the predicted probabilities of 125 models. The objective function of the model adopts the binary log loss and uses the Matthews correlation coefficient as the threshold optimization index. The probability threshold that maximizes the MCC is searched on the training set for the final classification decision.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) By using wavelet transform and inflection point detection algorithm, the detection accuracy of partial discharge signal is effectively improved; (2) The use of multi-dimensional feature fusion and LightGBM integrated learning framework improves the accuracy and reliability of insulation defect diagnosis; (3) The sliding time window verification mechanism reduces the false alarm rate and improves the stability of diagnosis. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of the diagnostic method of the present invention; Figure 2 This is a schematic diagram illustrating the wavelet transform process of the original voltage signal of the monitoring cable according to the present invention. Figure 3 This is a schematic diagram showing the result of wavelet transform on the original voltage signal of the monitoring cable according to the present invention. Detailed Implementation

[0017] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0018] 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.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0020] Example: like Figure 1 As shown, a diagnostic method for a partial discharge insulation defect model in a high-voltage cable includes the following steps: S1. Input the original voltage sampling signal data of the high-voltage cable to be tested; S2. Clean the noise signal in the original voltage sampling signal data by wavelet transform and filter out partial discharge peaks; S3. Extract the multi-dimensional features of the partial discharge peak to form a measurement group fusion feature vector; S4. Input the fused feature vector of the measurement group into the pre-trained partial discharge insulation degradation model to obtain the insulation degradation probability; S5. Based on preset verification standards, the state of the high-voltage cable is determined by logical judgment of the insulation degradation probability.

[0021] The original sampling signal data of the high-voltage cable is collected by an intelligent inspection robot. The robot moves within the pipeline and, upon reaching a preset sampling location, collects voltage sampling signal data using its robotic arm. The collected voltage sampling signal data is then processed locally or uploaded to a host computer for diagnostic purposes to determine if the high-voltage cable has partial discharge insulation defects, facilitating intervention.

[0022] A high-frequency voltage sensor on the robotic arm collects signals, each signal consisting of 800,000 high-speed samples within a 20ms monitoring period, at a sampling frequency of 40MHz. This sampling data is sufficient to effectively reflect the degradation of the high-voltage cable during the monitoring period. The data is compressed and stored in int8 format. The sampled data is organized in a columnar storage format using Parquet files, supporting on-demand loading and chunked processing, significantly reducing memory usage and improving data reading efficiency.

[0023] The high-frequency voltage sensor acquisition signal is then processed to remove noise and filter out the true partial discharge peaks.

[0024] As shown in step S2, this application uses wavelet transform for denoising and filters out partial discharge peaks.

[0025] Step S2 includes: S21. Wavelet function is used to perform a 5-level tower decomposition on the original voltage sampling signal data, which is decomposed into approximation coefficients and detail coefficients. S22. Apply a soft threshold to the detail coefficients, wherein the threshold is adaptively determined by the noise standard deviation; S23. Based on the processed wavelet coefficients, the reconstructed denoised signal is obtained by inverse wavelet transform. S24. Take the absolute value of the reconstructed denoised signal and identify candidate peak points using a local maximum detection algorithm; S25. Separate the signal and noise using the inflection point detection algorithm to obtain the partial discharge peak value.

[0026] The collected raw voltage sampling signal data of the high-voltage cable includes the actual voltage signal and the noise signal. According to the observation, the noise signal is mainly high-frequency random signal. The wavelet coefficients obtained by decomposition are processed by the threshold form of wavelet transform, and then the signal is denoised by wavelet reconstruction.

[0027] like Figure 2 As shown, in some embodiments, wavelet functions are used to perform a 5-level pyramid decomposition on the signal, decomposing it into approximation coefficients and detail coefficients, the expression of which is:

[0028] In the formula, a is the frequency. This refers to the time offset. By performing time-frequency analysis on the signal, it is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients.

[0029] The soft thresholding process in step S22 involves processing based on the frequency values ​​of the detail coefficients. A strong threshold is applied to the lower-level detail coefficients to suppress ambient white noise, while a weak threshold is applied to the higher-level detail coefficients to preserve partial discharge signal characteristics. The processed result is as follows: Figure 3 As shown.

[0030] Specifically, threshold It is determined adaptively through the noise standard deviation, and its expression is: ,in, This is an estimate of the noise standard deviation. This is the signal length.

[0031] The denoised signal is reconstructed using inverse wavelet transform to obtain the reconstructed denoised signal. The signal is mainly characterized by the noise-reduced partial discharge signal.

[0032] Step S24 reconstructs the denoised signal This process involves identifying specific points worthy of analysis and requiring further processing. The specific steps are as follows: The signal is converted into absolute value form, and a local maximum detection algorithm is used to identify candidate peak points. The local maximum detection uses a local maximum detection algorithm with a window length of 25. If the amplitude of a point is greater than the amplitudes of the 25 adjacent points on either side, it is determined to be a local maximum.

[0033] After capturing the location of local maxima, a corner point detection algorithm is used for filtering and selection to retain the partial discharge peak value. Its expression is:

[0034] in This represents the peak amplitude. The peak index is used. After smoothing the gradient sequence by convolution of length 9, the plateau detection method is used to identify the inflection point of the noise floor. The peak value before the inflection point is retained as the effective partial discharge peak value, effectively separating the signal from the noise.

[0035] After denoising and identifying the location of partial discharge peaks, the feature information of the local peaks is extracted to form a feature vector for model recognition. This vector is used to identify whether the local peak has deteriorated and to evaluate the quality of the signal where the local peak is located.

[0036] To avoid missing feature information and dimensionality explosion, this embodiment, through experimentation, defines a measurement group fusion feature vector that includes: the amplitude ratio of adjacent sampling points, the root mean square error of sawtooth wave fitting, frequency domain phase alignment features, and statistical aggregation features. These are formed by concatenating vectors to create a measurement group-level fusion feature vector. The statistical aggregation features include statistical values ​​of peak counts in the 0°-180° and 180°-360° phase intervals, the mean and standard deviation of peak heights, the mean amplitude ratio of adjacent sampling points, the mean root mean square error of sawtooth wave fitting, and the average offset between the peak and the zero-crossing point of the out-of-phase phase. In some embodiments, the frequency domain phase alignment feature obtains the phase angle by calculating the 50Hz fundamental frequency Fourier coefficients, mapping the peak position to the phase space and dividing it into four quadrants to achieve phase-resolved analysis.

[0037] In principle, the fusion feature vector of this measurement group specifically includes time-domain morphological features, frequency-domain morphological features, and statistical aggregation features.

[0038] Specifically, the temporal morphological characteristics include the amplitude ratio characteristics of adjacent sampling points and the root mean square error characteristics of the sawtooth wave fitting. For each effective peak, the amplitude ratio between it and its left and right adjacent peaks is calculated.

[0039] This ratio characterizes the steepness and attenuation characteristics of the discharge pulse. The root mean square error (RMSE) characteristic of the sawtooth wave fitting is calculated to determine the RMS error between the 25-point segments before and after the peak and the sawtooth wave template.

[0040] in, For the j-th point of the sawtooth wave template, For template length, For template radius, This represents the peak value sequence. Template matching is used to quantify the standard deviation of the partial discharge waveform. A smaller error indicates a more regular waveform that better matches typical partial discharge characteristics.

[0041] Frequency domain phase characteristics: Calculate the 50Hz fundamental frequency Fourier coefficients to obtain phase alignment information. For each signal Calculate the phase angle of its 50Hz component. And the peak position Mapping to phase space:

[0042] in The sampling interval is used. The peak value is divided into four quadrants based on the phase angle. This enables phase-resolved analysis.

[0043] Statistical aggregation features: Signal-level peak characteristics are aggregated into group-level statistics based on measurement group identifiers. This includes the statistical feature of peak quantity. , respectively representing the total number of peaks and Number of peak values ​​in phase interval; statistical characteristics of peak height The discharge energy intensity and dispersion are characterized by the distance offset statistics, which characterize the average offset of the discharge pulse from the zero-crossing point of the reverse phase.

[0044] The aforementioned feature information is concatenated column by column to form a fused feature vector. This fused feature vector comprehensively characterizes the time-frequency domain characteristics, morphological features, and statistical distribution patterns of partial discharge events.

[0045] Next, steps S4 and S5 are executed, inputting the obtained measurement group fusion feature vector into the pre-trained partial discharge insulation degradation model to obtain the probability of whether the partial discharge event of the measurement group has deteriorated. Combined with preset verification standards, it is then determined whether deterioration has occurred.

[0046] Specifically, by pre-setting a threshold for the robot or cloud device, which is a specific probability, if the result obtained by the model is higher than the threshold, it is determined that there is an abnormal discharge and an alarm is triggered; otherwise, it is determined to be a normal state, thereby enabling routine monitoring and diagnosis.

[0047] In some embodiments, a sliding time window verification mechanism is adopted to avoid false alarms caused by instantaneous probability fluctuations. For multiple consecutive monitoring cycles, a latent insulation degradation defect is finally determined and an early warning is triggered only when the probability of insulation degradation defects in the same measurement group is continuously higher than the threshold within a preset time period (specifically set to 3-5 detection cycles in this application).

[0048] Next, this application also discloses the training process of the aforementioned partial discharge insulation degradation model.

[0049] The first step involves processing and labeling the training data. Training data consisting of signal-level data is acquired, grouped based on measurement group identifiers, and then logically or rule-basedly processed based on the insulation degradation labels of the signal-level data within each measurement group. This constructs insulation degradation labels at the measurement group level and generates a measurement group-discriminative feature vector.

[0050] The data is generated for each measurement group, yielding corresponding signals. Hierarchically, the data includes the upper-level measurement group and the signal level data under each measurement group. Each signal level data is pre-labeled with a corresponding insulation degradation tag. Based on its measurement group, it is processed according to logic or rules, enabling the corresponding measurement group to obtain the appropriate insulation degradation tag based on its subordinate signal level data.

[0051] Specifically, in this embodiment, the logic is set such that if the insulation degradation label of any signal level data in the measurement group is "yes", then the insulation degradation label of the measurement group is "yes"; otherwise, it is "no".

[0052] The various feature values ​​in this feature vector are normalized to convert each feature value into a value between 0 and 1, avoiding proportional differences between features and ensuring the training effect of the model.

[0053] In some embodiments, after constructing the insulation degradation labels for the measurement group, a data index alignment check is performed to verify that the id_measurement index of the feature data frame is completely consistent with the label sequence index. An assertion check is used to ensure a one-to-one correspondence between samples and to avoid training bias caused by data misalignment.

[0054] In some embodiments, to ensure the training results of the model, it is also necessary to consider the ratio between true and negative samples in the training set. Positive samples refer to measurement groups with insulation degradation, while negative samples refer to measurement groups without insulation degradation. If the proportion of positive samples is too low, the model's training performance will be poor, making it difficult to learn the features of insulation degradation. Therefore, when the proportion of positive samples in the training set is typical of an imbalanced dataset, this application adjusts the classification threshold strategy to optimize the minority class recognition capability.

[0055] After the labeling is completed, the measurement group data is processed using the aforementioned denoising and feature extraction methods to obtain the measurement group fused feature vector.

[0056] Next, we combine the discriminative feature vectors of the measurement group to construct and train an ensemble learning framework based on LightGBM.

[0057] The ensemble learning framework constructs 125 independent LightGBM classifiers and performs probabilistic ensemble; the objective function of the model is binary log loss, and the Matthews correlation coefficient is used as the threshold optimization index to find the optimal judgment threshold for the final classification decision.

[0058] The model was trained using a three-fold cross-validation strategy with multiple random seed iterations. Specifically, the random seed iterations were set to 25, and the model was trained using three-fold cross-validation. The model configuration was as follows: learning rate 0.01, leaf nodes 80, feature sampling rate 0.8, sample sampling rate 0.8, maximum iterations 10,000, and early stopping patience value 100.

[0059] In some embodiments, the fused feature vector F obtained from the aforementioned training set data is input into the model, and the model outputs the probability of insulation degradation defects existing in each measurement group during the monitoring period:

[0060] in, This is the sigmoid function, used to map the model output to a range of 0-1, representing the probability of an insulation degradation defect. For the t-th LightGBM classifier, the input feature vector The predicted value.

[0061] The probability of insulation degradation defects predicted by 125 models during training was ensembled using an arithmetic mean:

[0062] The result is the degradation probability obtained through this model.

[0063] 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 diagnosing partial discharge insulation defects in high-voltage cables, characterized in that, Includes the following steps: S1. Input the original voltage sampling signal data of the high-voltage cable to be tested; S2. Clean the noise signal in the original voltage sampling signal data and filter out the partial discharge peak value by wavelet transform; S3. Extract the multi-dimensional features of the partial discharge peak to form a measurement group fusion feature vector, wherein the measurement group fusion feature vector includes: the amplitude ratio of adjacent sampling points, the root mean square error of sawtooth wave fitting, frequency domain phase alignment features and statistical aggregation features. The measurement group level fusion feature vector is formed by vector concatenation. The statistical aggregation features include the statistical values ​​of the number of peaks in the phase intervals of 0°-180° and 180°-360°, the mean and standard deviation of peak height, the mean amplitude ratio of adjacent sampling points, the mean root mean square error of sawtooth wave fitting, and the average offset of the distance between the peak and the zero-crossing point of the inverted phase. S4. Input the fused feature vector of the measurement group into the pre-trained partial discharge insulation degradation model to obtain the insulation degradation probability. The partial discharge insulation degradation model is trained through the following steps: S41. Obtain training data composed of signal-level data, group the signal-level data based on the measurement group identifier, perform logical or rule processing based on the insulation degradation label of the signal-level data within the measurement group, construct the insulation degradation label at the measurement group level, and form a measurement group discrimination feature vector. S42. Based on the discriminative feature vectors of the measurement group, construct and train an ensemble learning framework based on LightGBM; S5. Based on preset verification standards, the state of the high-voltage cable is determined by logical judgment of the insulation degradation probability.

2. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 1, characterized in that, Step S2 includes: S21. Wavelet function is used to perform a 5-level tower decomposition on the original voltage sampling signal data, which is decomposed into approximation coefficients and detail coefficients. S22. Apply a soft threshold to the detail coefficients, wherein the threshold is adaptively determined by the noise standard deviation; S23. Based on the processed wavelet coefficients, the reconstructed denoised signal is obtained by inverse wavelet transform. S24. Take the absolute value of the reconstructed denoised signal and identify candidate peak points using a local maximum detection algorithm; S25. Separate the signal and noise using the inflection point detection algorithm to obtain the partial discharge peak value.

3. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 2, characterized in that, In step S24, the local maximum point detection adopts a local maximum detection algorithm with a window length of 25.

4. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 2, characterized in that, The inflection point detection algorithm in step S25 includes: arranging the candidate peaks in descending order of amplitude, calculating the amplitude gradient between adjacent peaks, performing convolution smoothing on the gradient sequence with a length of 9, using the plateau detection method to identify the inflection point of the noise base, and retaining the peaks before the inflection point as partial discharge peaks.

5. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 1, characterized in that, The frequency domain phase alignment feature obtains the phase angle by calculating the 50Hz fundamental frequency Fourier coefficients, maps the peak position to the phase space and divides it into four quadrants to achieve phase resolution analysis.

6. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 1, characterized in that, In step S5, when the probability of insulation degradation is greater than the preset threshold, the high-voltage cable is determined to be in a degraded state. This also includes a sliding time window verification mechanism. Only when the probability of insulation defects in the same measurement group is continuously higher than the optimal threshold within a preset time period is it finally determined to be a latent fault and an early warning is triggered. The duration threshold is set to 3-5 monitoring cycles.

7. The diagnostic method for partial discharge insulation defects in high-voltage cables according to claim 1, characterized in that, The ensemble learning framework generates 125 independent LightGBM classifiers through random seed iterations and trains them using three-fold cross-validation. The model is configured with a learning rate of 0.01, 80 leaf nodes, a feature sampling rate of 0.8, a sample sampling rate of 0.8, a maximum iteration count of 10,000, and an early stopping patience value of 100.

8. The method for diagnosing partial discharge insulation defects in high-voltage cables according to claim 1, characterized in that, The insulation degradation probability was obtained through a probabilistic ensemble method, which used the arithmetic mean method to fuse the predicted probabilities of 125 models. The objective function of the model adopted a binary log loss and used the Matthews correlation coefficient as the threshold optimization index. The probability threshold that maximizes the MCC was searched on the training set for the final classification decision.

Images