Lightning arrester fault detection method and system based on current data

Through high-frequency sampling, wavelet transformation and deep learning technology, accurate detection and prediction of subway lightning arrester faults is achieved, and the problems of misjudgment, misjudgment and lack of adaptability in the existing technology are solved, and the accuracy of fault detection is improved. and preventive maintenance capabilities.

CN118885885BActive Publication Date: 2024-12-03CCCC (TIANJIN) TECH INSPECTION CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411367539.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-29
Publication Date
2024-12-03
Estimated Expiration
2044-09-29

AI Technical Summary

Technical Problem

The prior art lightning arrester fault detection in subway systems has misjudgment, misjudgment and lack of adaptability, making it difficult to deal with complex faults and environmental changes, and cannot effectively predict the failure trend, affecting safety and reliability.

Method used

High-frequency sampling and wavelet transform noise reduction processing technology are adopted, combined with convolutional neural networks and long-term memory networks, multi-dimensional feature extraction and fault pattern recognition are carried out to achieve fault severity assessment and future state prediction.

Benefits of technology

It improves the accuracy and real-time nature of fault detection, reduces misjudgment and misjudgment, provides preventive maintenance support, reduces the risk of sudden failures, and improves the safety and reliability of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118885885B_ABST
    Figure CN118885885B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of fault detection, and discloses a lightning arrester fault detection method and system based on current data, which are used to improve the accuracy and efficiency of lightning arrester fault detection based on current data. The method includes: performing high-frequency sampling on the current signal of the subway lightning arrester grounding line to obtain the original current data; performing wavelet transform noise reduction processing on the original current data to obtain the noise-reduced current data; performing feature extraction on the noise-reduced current data to obtain a multi-dimensional feature vector; performing fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result; performing weighted scoring processing on the fault mode classification result to obtain a fault severity assessment score; performing fault prediction on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a lightning arrester fault detection method and system based on current data. Background Art

[0002] In modern urban rail transit systems, subways are important public transportation tools, and their safe operation is of vital importance. Lightning arresters in subway systems are key components for protecting power equipment and passenger safety, and are used to prevent lightning and overvoltage from damaging the system. At present, fault detection of subway lightning arresters mainly relies on regular manual inspections and basic electrical parameter monitoring. These methods usually use threshold judgments on basic electrical quantities such as voltage and current to evaluate the working status of lightning arresters. Some advanced subway systems also use online monitoring equipment to collect parameters such as leakage current and harmonic content of lightning arresters in real time, and identify possible faults through preset judgment rules.

[0003] However, these traditional methods have some obvious shortcomings in the subway environment. First, the subway has a long operating time and a complex environment. The periodicity and subjectivity of manual inspections make it difficult to detect hidden faults in a timely manner, which may affect the normal operation of trains. Secondly, the electromagnetic interference in the subway system is strong, and simple threshold judgments cannot effectively identify complex fault modes, which is prone to misjudgment and missed judgments, increasing maintenance costs. Furthermore, the preset judgment rules lack adaptive capabilities and are difficult to cope with the aging process of subway lightning arresters under high-frequency use and special environments, as well as the impact of seasonal and environmental changes. Finally, existing technologies generally lack the ability to predict the development trend of subway lightning arrester failures, and cannot provide strong support for preventive maintenance, potentially affecting the safety and reliability of the subway system. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and system for detecting a lightning arrester fault based on current data, so as to improve the accuracy and efficiency of detecting a lightning arrester fault based on current data.

[0005] The present invention provides a lightning arrester fault detection method based on current data, comprising: high-frequency sampling of a current signal of a subway lightning arrester grounding line to obtain original current data; performing wavelet transform noise reduction processing on the original current data to obtain noise-reduced current data, wherein the step of performing wavelet transform noise reduction processing on the original current data to obtain noise-reduced current data comprises: selecting a wavelet basis function on the original current data to obtain a wavelet basis function, and performing wavelet decomposition on the original current data according to the wavelet basis function to obtain wavelet coefficients; performing threshold calculation on the wavelet coefficients to obtain a noise reduction threshold, and performing soft threshold processing on the wavelet coefficients according to the noise reduction threshold to obtain noise-reduced wavelet coefficients; performing wavelet reconstruction on the noise-reduced wavelet coefficients to obtain preliminary noise reduction data, and performing residual noise evaluation on the preliminary noise reduction data to obtain a noise evaluation result; and performing residual noise evaluation on the preliminary noise reduction data according to the noise evaluation result. The preliminary noise reduction data is adaptively filtered to obtain filtered current data, and the signal-to-noise ratio of the filtered current data is calculated to obtain a signal-to-noise ratio index; a threshold judgment is performed on the signal-to-noise ratio index to obtain a judgment result, and the noise reduction effect of the filtered current data is confirmed according to the judgment result to obtain a noise reduction confirmation result; data calibration is performed on the filtered current data according to the noise reduction confirmation result to obtain the noise-reduced current data; feature extraction is performed on the noise-reduced current data to obtain a multi-dimensional feature vector; fault identification and classification processing is performed on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result; weighted scoring processing is performed on the fault mode classification result to obtain a fault severity assessment score; fault prediction is performed on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.

[0006] The present invention also provides a lightning arrester fault detection system based on current data, comprising:

[0007] The sampling module is used to perform high-frequency sampling on the current signal of the subway lightning arrester grounding line to obtain the original current data;

[0008] The noise reduction module is used to perform wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction. The step of performing wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction includes: selecting a wavelet basis function on the original current data to obtain a wavelet basis function, and performing wavelet decomposition on the original current data according to the wavelet basis function to obtain wavelet coefficients; performing threshold calculation on the wavelet coefficients to obtain a noise reduction threshold, and performing soft threshold processing on the wavelet coefficients according to the noise reduction threshold to obtain the wavelet coefficients after noise reduction; performing wavelet decomposition on the wavelet coefficients after noise reduction Reconstruct to obtain preliminary noise reduction data, and perform residual noise evaluation on the preliminary noise reduction data to obtain a noise evaluation result; adaptively filter the preliminary noise reduction data according to the noise evaluation result to obtain filtered current data, and perform signal-to-noise ratio calculation on the filtered current data to obtain a signal-to-noise ratio index; perform threshold judgment on the signal-to-noise ratio index to obtain a judgment result, and confirm the noise reduction effect of the filtered current data according to the judgment result to obtain a noise reduction confirmation result; perform data calibration on the filtered current data according to the noise reduction confirmation result to obtain the noise-reduced current data;

[0009] An extraction module, used for performing feature extraction on the noise-reduced current data to obtain a multi-dimensional feature vector;

[0010] An identification module is used to perform fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result;

[0011] A scoring module is used to perform weighted scoring processing on the fault mode classification results to obtain a fault severity assessment score;

[0012] The prediction module is used to perform fault prediction on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.

[0013] In the technical solution provided by the present invention, through high-frequency sampling and wavelet transform noise reduction processing, the method effectively improves the quality of the original current data, high-frequency sampling can capture transient current characteristics, and wavelet transform noise reduction effectively removes environmental noise and electromagnetic interference, making the fault characteristics more prominent. Secondly, the multi-dimensional feature extraction technology comprehensively considers the time domain, frequency domain, energy and nonlinear characteristics, greatly enhancing the characterization ability of the fault mode. This multi-angle feature extraction method can more comprehensively describe the working state of the arrester and help identify complex and hidden fault modes. Furthermore, the use of convolutional neural networks for fault identification and classification significantly improves the accuracy and robustness of identification. The multi-layer structure of the convolutional neural network can automatically learn the intrinsic characteristics of the data, adapt to the fault mode under different working conditions, and reduce the subjectivity and uncertainty of human experience judgment. In addition, the evaluation of the severity of the fault adopts a multi-factor weighted calculation and dynamic adjustment mechanism, which not only considers the current state, but also incorporates historical trends and system impacts, making the evaluation results more comprehensive and accurate. This evaluation method can provide a more reliable basis for maintenance decisions and help to formulate a more reasonable maintenance plan. Finally, the application of long short-term memory network realizes the accurate prediction of the future state of the arrester, providing strong support for preventive maintenance. Through deep learning of historical data, this method can capture long-term dependencies and potential fault development trends, and warn of possible faults in advance, thereby greatly reducing the risk of sudden failures. In general, this arrester fault detection method based on current data integrates advanced signal processing, machine learning and deep learning technologies, and realizes the intelligence of the whole process from data acquisition, feature extraction, fault identification to state prediction. It not only improves the accuracy and real-time performance of fault detection, but also realizes fault prediction and preventive maintenance, effectively reducing the downtime and maintenance cost of the subway system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 The present invention is a flowchart of a method for detecting lightning arrester faults based on current data in an embodiment of the present invention.

[0016] Figure 2 Schematic diagram of a lightning arrester fault detection system based on current data in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0019] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 FIG. 1 is a flow chart of a lightning arrester fault detection method based on current data according to an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0021] S101, performing high frequency sampling on the current signal of the subway lightning arrester grounding line to obtain original current data;

[0022] S102, performing wavelet transform noise reduction processing on the original current data to obtain noise-reduced current data;

[0023] Specifically, wavelet basis functions are selected for the original current data to obtain wavelet basis functions, and wavelet decomposition is performed on the original current data according to the wavelet basis functions to obtain wavelet coefficients; threshold calculation is performed on the wavelet coefficients to obtain noise reduction thresholds, and soft threshold processing is performed on the wavelet coefficients according to the noise reduction thresholds to obtain noise-reduced wavelet coefficients; wavelet reconstruction is performed on the noise-reduced wavelet coefficients to obtain preliminary noise reduction data, and residual noise evaluation is performed on the preliminary noise reduction data to obtain noise evaluation results; adaptive filtering is performed on the preliminary noise reduction data according to the noise evaluation results to obtain filtered current data, and signal-to-noise ratio calculation is performed on the filtered current data to obtain a signal-to-noise ratio index; threshold judgment is performed on the signal-to-noise ratio index to obtain a judgment result, and noise reduction effect is confirmed on the filtered current data according to the judgment result to obtain a noise reduction confirmation result; data calibration is performed on the filtered current data according to the noise reduction confirmation result to obtain noise-reduced current data;

[0024] S103, extracting features from the noise-reduced current data to obtain a multi-dimensional feature vector;

[0025] S104, performing fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result;

[0026] S105, performing weighted scoring processing on the fault mode classification results to obtain a fault severity assessment score;

[0027] S106. Fault prediction is performed on the pre-collected historical feature data and fault severity assessment scores through a long short-term memory network to obtain a future fault state prediction result.

[0028] Specifically, the current signal of the arrester grounding line is sampled at high frequency. A high-precision current sensor is used to collect the current signal at a sampling frequency of at least 10kHz to ensure that high-frequency transient characteristics are captured. For example, for a lightning arrester with a rated current of 5kA, 500 data points may be collected within 1ms to form an original current data sequence. Next, the original current data is subjected to wavelet transform denoising. A suitable wavelet basis function (such as Daubechies wavelet) is selected to perform multi-scale decomposition of the signal, and then the wavelet coefficients are processed by the soft threshold method, and finally the signal is reconstructed to obtain the denoised current data. Environmental noise and electromagnetic interference are effectively removed, and the accuracy of subsequent analysis is improved.

[0029] The next key step is to extract features from the noise-reduced current data. The extracted features include time domain features (such as mean, variance, peak factor, etc.), frequency domain features (such as spectrum energy distribution, harmonic content, etc.), wavelet packet energy features, and nonlinear features (such as approximate entropy, sample entropy, etc.). These features together constitute a multidimensional feature vector that fully reflects the working status of the arrester.

[0030] After feature extraction, the multi-dimensional feature vector is processed by convolutional neural network (CNN) for fault identification and classification. The structure of CNN includes multiple convolutional layers, pooling layers and fully connected layers. The convolutional layer extracts higher-level feature representations by learning the local correlation between features. The pooling layer reduces the data dimension and improves the computational efficiency of the model. The fully connected layer finally outputs the fault classification result. For example, CNN may classify the state of the arrester into four categories: normal, slightly degraded, severely degraded and failed.

[0031] Weighted scoring of fault mode classification results is the key to assessing fault severity. Weights are assigned according to the degree of harm of different fault types, and combined with the degree of abnormality of fault characteristics, a fault severity assessment score between 0 and 100 is calculated. For example, for a slightly degraded state, an assessment score of 30 points may be given, while for a severely degraded state, an assessment score of 80 points may be given. Finally, the long short-term memory network (LSTM) is used to predict faults based on the pre-collected historical feature data and fault severity assessment scores. The LSTM network can effectively capture long-term dependencies in time series data and is suitable for trend prediction. By inputting daily feature data and assessment scores over a period of time (such as 30 days), LSTM can predict the development trend of the arrester's fault state in the next week.

[0032] For example, a lightning arrester on a subway line showed signs of slight degradation during continuous monitoring. Through high-frequency sampling, 10,000 raw current data points were collected per second. After wavelet transform noise reduction processing, the noise level was reduced by about 80%. The feature extraction step extracted 50 features from the denoised data to form a 50-dimensional feature vector. The CNN model classified this feature vector as a "slightly degraded" state with a confidence level of 85%. The weighted scoring process gave a fault severity assessment score of 35 points. Based on the data from the past 30 days, the LSTM network predicted that the fault score of the lightning arrester might rise to 45 points in the next 7 days, indicating that its condition might deteriorate further.

[0033] It should be noted that when executing step S102, the wavelet basis function is selected for the original current data to obtain the wavelet basis function. The selection of the wavelet basis function directly affects the noise reduction effect. Commonly used wavelet basis functions include Daubechies wavelet, Symlets wavelet and Coiflets wavelet. When selecting, the characteristics of the arrester current signal, such as the smoothness and mutation of the signal, need to be considered. For example, for the current signal of the subway arrester, Daubechies wavelet (such as db4 or db6) is usually selected as the basis function because it has good localization characteristics and sensitivity to mutation signals. According to the selected wavelet basis function, the original current data is wavelet decomposed to obtain the wavelet coefficient. Wavelet decomposition is the process of decomposing a signal into wavelet coefficients of different scales and positions. For the current signal of the subway arrester, 3 to 5 layers of wavelet decomposition are usually performed to fully separate the high-frequency noise and low-frequency useful information in the signal. The wavelet coefficients obtained after decomposition include approximate coefficients (low-frequency part) and detail coefficients (high-frequency part). Next, the wavelet coefficients are threshold calculated to obtain the noise reduction threshold. The selection of threshold directly affects the noise reduction effect. Common threshold selection methods include the Universal threshold method and the SURE threshold method. For the current signal of the subway arrester, an adaptive threshold method based on the statistical characteristics of the wavelet coefficient is usually used to better adapt to different noise levels. The wavelet coefficients are soft-thresholded according to the calculated noise reduction threshold to obtain the noise-reduced wavelet coefficients. Soft threshold processing can effectively suppress noise while retaining the main features of the signal. The noise-reduced wavelet coefficients are reconstructed by wavelet to obtain preliminary noise reduction data. Wavelet reconstruction is the inverse process of wavelet decomposition. By recombining the processed wavelet coefficients, the noise-reduced time domain signal is restored. The preliminary noise reduction data is evaluated for residual noise to obtain the noise evaluation result. The residual noise evaluation usually uses indicators such as signal-to-noise ratio (SNR) or mean square error (MSE) to quantify the noise reduction effect.

[0034] According to the noise evaluation results, the preliminary noise reduction data is adaptively filtered to obtain the filtered current data. Adaptive filtering can further eliminate residual noise. Common methods include Wiener filtering and Kalman filtering. For the current signal of the subway arrester, an adaptive filter based on the minimum mean square error criterion is usually used to adapt to the non-stationary characteristics of the signal. The signal-to-noise ratio of the filtered current data is calculated to obtain the signal-to-noise ratio index. The signal-to-noise ratio calculation evaluates the data quality by comparing the signal power with the noise power. The signal-to-noise ratio index is threshold judged to obtain the judgment result. The threshold judgment compares the calculated signal-to-noise ratio with the preset threshold to evaluate whether the noise reduction effect meets the requirements. According to the judgment result, the noise reduction effect of the filtered current data is confirmed to obtain the noise reduction confirmation result. If the noise reduction effect does not meet the requirements, it may be necessary to readjust the wavelet basis function or threshold and repeat the above noise reduction process.

[0035] Finally, the filtered current data is calibrated according to the noise reduction confirmation result to obtain the noise-reduced current data. Data calibration aims to eliminate the deviation that may be introduced in the noise reduction process and ensure that the noise-reduced data accurately reflects the actual working status of the arrester.

[0036] For example, the raw current data collected by the arrester on a subway line during a lightning strike event contains 5000 sampling points with a sampling frequency of 50kHz. First, the db4 wavelet is selected as the basis function, and the raw data is decomposed by a 4-layer wavelet. After decomposition, an approximate coefficient vector with a length of 313 and four detail coefficient vectors with lengths of 313, 625, 1250, and 2500 are obtained. The adaptive threshold method based on the statistical characteristics of the wavelet coefficients is used to calculate the noise reduction threshold. For detail coefficients of different scales, the thresholds obtained are 0.15, 0.22, 0.31, and 0.45 (unit: kA). The wavelet coefficients are processed using a soft threshold function. For example, for the first layer of detail coefficients, the coefficients with an absolute value less than 0.15kA are set to 0, and the coefficients greater than 0.15kA are retained but their amplitudes are reduced by 0.15kA. The processed wavelet coefficients are reconstructed to obtain preliminary noise reduction data. The signal-to-noise ratio of the preliminary noise reduction data is calculated to be 25dB. Then, an adaptive filter based on the minimum mean square error criterion was used for further processing, and the signal-to-noise ratio after filtering was increased to 28dB. 28dB was compared with the preset 26dB threshold to confirm that the noise reduction effect met the requirements. Finally, data calibration was performed to correct the amplitude deviation of about 0.5% to obtain the final noise-reduced current data. The entire noise reduction process effectively removed the high-frequency noise and random disturbances in the original signal, while retaining the rapid rise and slow decay characteristics of the lightning current.

[0037] By executing the above steps, the method effectively improves the quality of the original current data through high-frequency sampling and wavelet transform noise reduction processing. High-frequency sampling can capture transient current characteristics, while wavelet transform noise reduction effectively removes environmental noise and electromagnetic interference, making the fault characteristics more prominent. Secondly, the multi-dimensional feature extraction technology comprehensively considers the time domain, frequency domain, energy and nonlinear characteristics, greatly enhancing the characterization ability of the fault mode. This multi-angle feature extraction method can more comprehensively describe the working state of the lightning arrester and help identify complex and hidden fault modes. Furthermore, the use of convolutional neural networks for fault identification and classification significantly improves the accuracy and robustness of identification. The multi-layer structure of the convolutional neural network can automatically learn the intrinsic characteristics of the data, adapt to the fault mode under different working conditions, and reduce the subjectivity and uncertainty of human experience judgment. In addition, the evaluation of the fault severity adopts a multi-factor weighted calculation and dynamic adjustment mechanism, which not only considers the current state, but also incorporates historical trends and system impacts, making the evaluation results more comprehensive and accurate. This evaluation method can provide a more reliable basis for maintenance decisions and help to formulate a more reasonable maintenance plan. Finally, the application of long short-term memory network realizes the accurate prediction of the future state of the arrester, providing strong support for preventive maintenance. Through deep learning of historical data, this method can capture long-term dependencies and potential fault development trends, and warn of possible faults in advance, thereby greatly reducing the risk of sudden failures. In general, this arrester fault detection method based on current data integrates advanced signal processing, machine learning and deep learning technologies, and realizes the intelligence of the whole process from data acquisition, feature extraction, fault identification to state prediction. It not only improves the accuracy and real-time performance of fault detection, but also realizes fault prediction and preventive maintenance, effectively reducing the downtime and maintenance cost of the subway system.

[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0039] (1) Install a current sensor on the grounding line of the subway lightning arrester to obtain a current acquisition device, and set the sampling frequency of the current acquisition device to obtain sampling parameters;

[0040] (2) Checking the numerical range of the sampling parameters to obtain valid sampling parameters, and calculating the sampling time interval of the current acquisition device according to the valid sampling parameters to obtain the sampling time interval;

[0041] (3) Evaluate the anti-interference capability of the sampling time interval to obtain an anti-interference evaluation result, and perform sampling trigger setting on the current acquisition device according to the anti-interference evaluation result to obtain a sampling trigger signal;

[0042] (4) synchronously calibrating the sampling trigger signal to obtain a calibrated trigger signal, and sampling the current of the current acquisition device according to the calibrated trigger signal to obtain analog current data;

[0043] (5) performing analog-to-digital conversion on the analog current data to obtain digitized current data, and adding a timestamp to the digitized current data to obtain current data with a timestamp;

[0044] (6) Performing data format standardization processing on the current data with timestamp to obtain current data in a standard format, and performing data integrity verification on the current data in the standard format to obtain original current data.

[0045] Specifically, a current sensor is installed on the grounding line of the subway lightning arrester to obtain a current acquisition device. This involves selecting a suitable type of current sensor, such as a Hall effect sensor or a Rogowski coil, and installing it on the grounding line of the lightning arrester. During installation, the measurement range, accuracy and anti-interference ability of the sensor must be considered to ensure that the working current of the lightning arrester can be accurately captured. Next, the sampling frequency of the current acquisition device is set to obtain the sampling parameters. The selection of the sampling frequency must take into account the frequency range of the lightning arrester's working characteristics and fault characteristics, and is usually set between 10kHz and 100kHz to capture high-frequency transient phenomena. Subsequently, the sampling parameters are checked for a numerical range to obtain valid sampling parameters. This is to ensure that the set sampling frequency is within the working range of the current acquisition device and satisfies the Nyquist sampling theorem, that is, the sampling frequency is at least twice the highest frequency of the signal. The sampling time interval of the current acquisition device is calculated based on the valid sampling parameters to obtain the sampling time interval. The sampling time interval is the reciprocal of the sampling frequency. For example, for a sampling frequency of 50kHz, the sampling time interval is 20 microseconds.

[0046] Next, the anti-interference ability of the sampling time interval is evaluated to obtain the anti-interference evaluation result. The anti-interference ability of the sampling scheme is evaluated by analyzing the relationship between the sampling time interval and the characteristic frequency of environmental noise and electromagnetic interference. According to the anti-interference evaluation result, the sampling trigger setting of the current acquisition device is performed to obtain a sampling trigger signal. The sampling trigger setting may include setting the trigger threshold, selecting the rising edge or falling edge trigger, etc., to ensure accurate data collection at the critical moment. The sampling trigger signal is synchronously calibrated to obtain a calibrated trigger signal. The purpose of synchronous calibration is to ensure that the data acquisition time between multiple acquisition channels or multiple lightning arresters is accurately synchronized, which is usually achieved by GPS time synchronization or network time protocol (NTP). According to the calibrated trigger signal, the current acquisition device is sampled to obtain analog current data. The data acquisition process is actually performed to convert the current signal of the lightning arrester into a processable analog voltage signal.

[0047] Subsequently, the analog current data is converted from analog to digital to obtain digitized current data. The analog-to-digital conversion is completed by an analog-to-digital converter (ADC), which converts the continuous analog signal into a discrete digital signal. The resolution and sampling rate of the ADC need to be considered during the conversion process to ensure that the digitized data can accurately represent the original analog signal. The digitized current data is timestamped to obtain current data with a timestamp. The addition of timestamps is crucial for subsequent data analysis and fault location, and usually uses millisecond or microsecond accuracy. Finally, the data format of the timestamped current data is standardized to obtain current data in a standard format. Standardization includes operations such as unified data format, unit conversion, and data alignment to ensure data consistency and comparability. The data integrity of the current data in the standard format is verified to obtain the original current data. Integrity verification includes checking data continuity, identifying outliers, and processing missing data to ensure that high-quality and reliable data is used for subsequent analysis.

[0048] For example, a lightning arrester with a rated current of 10kA is installed on a subway line, and a Hall effect current sensor is used for monitoring. The sampling frequency is set to 50kHz, and the value range is checked to be valid. The sampling time interval is calculated to be 20 microseconds. By analyzing the electromagnetic interference characteristics in the subway environment, the sampling trigger threshold is determined to be 100A, and the rising edge trigger is used. The GPS timing equipment is used for synchronization calibration, and the time accuracy reaches 100 nanoseconds. The analog-to-digital conversion uses a 16-bit ADC with a resolution of 0.3A. In a lightning strike event, the system successfully captured the transient change process of the current: starting from the trigger moment (recorded as t=0), the current quickly rose from 100A to 8000A within 40 microseconds, and then gradually decayed to 500A within 200 microseconds. 5000 data points were collected in the whole process, and each data point was timestamped with microsecond accuracy. After data format standardization and integrity verification, the final raw current data fully recorded the entire process of this lightning strike event.

[0049] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0050] (1) Calculate the time domain characteristics of the noise-reduced current data to obtain the time domain characteristic value, and perform frequency domain transformation on the noise-reduced current data to obtain frequency domain data;

[0051] (2) Calculate the frequency domain characteristics of the frequency domain data to obtain the frequency domain eigenvalues, and perform wavelet packet decomposition on the noise-reduced current data to obtain the wavelet packet coefficients;

[0052] (3) Extract energy features from wavelet packet coefficients to obtain energy eigenvalues, and perform empirical mode decomposition on the noise-reduced current data to obtain intrinsic mode functions;

[0053] (4) Calculate the nonlinear characteristics of the intrinsic mode function to obtain the nonlinear eigenvalue, and perform feature fusion on the time domain eigenvalue, frequency domain eigenvalue, energy eigenvalue and nonlinear eigenvalue to obtain a fused feature set;

[0054] (5) Perform principal component analysis on the fused feature set to obtain a feature set after dimensionality reduction, and sort the feature importance of the feature set after dimensionality reduction to obtain a sorting result;

[0055] (6) Perform feature selection on the reduced feature set based on the sorting results to obtain a multi-dimensional feature vector.

[0056] Specifically, the time domain characteristics of the current data after noise reduction are calculated to obtain the time domain characteristic values. The time domain characteristics include statistical quantities such as mean, variance, peak, peak factor, skewness and kurtosis, which can reflect the basic statistical characteristics of the current signal. The calculation method is as follows:

[0057] Mean:

[0058] variance:

[0059] Crest Factor:

[0060] in, N is the number of data points, For the The value of the data point.

[0061] Next, the noise-reduced current data is transformed in the frequency domain to obtain frequency domain data. The frequency domain transformation usually uses the fast Fourier transform (FFT) algorithm to convert the time domain signal into a frequency domain representation. The frequency domain characteristics of the frequency domain data are calculated to obtain the frequency domain characteristic values. The frequency domain characteristics include the amplitude, phase, harmonic distortion rate, etc. of the main frequency components. These characteristics can reflect the spectral characteristics of the current signal. The formula for calculating the harmonic distortion rate is:

[0062]

[0063] in, For the The amplitude of the subharmonics, is the fundamental amplitude.

[0064] Subsequently, the noise-reduced current data is subjected to wavelet packet decomposition to obtain wavelet packet coefficients. Wavelet packet decomposition is an extension of wavelet transform and can provide more sophisticated time-frequency analysis. Usually 3 to 5 layers of decomposition are selected to obtain multi-scale wavelet packet coefficients. Energy feature extraction is performed on the wavelet packet coefficients to obtain energy eigenvalues. The energy feature calculation formula is:

[0065]

[0066] in, For the The energy of the sub-band, For the The subband wavelet packet coefficients, is the number of coefficients of this subband. Next, empirical mode decomposition (EMD) is performed on the denoised current data to obtain the intrinsic mode function (IMF). EMD is an adaptive signal decomposition method that can decompose complex signals into a finite number of IMFs and a residue. Nonlinear features are calculated for the intrinsic mode function to obtain nonlinear eigenvalues. Nonlinear features include approximate entropy, sample entropy, etc. These features can reflect the complexity and irregularity of the signal. The calculation formula for approximate entropy is:

[0067]

[0068] in, is the pattern dimension, is the similarity tolerance, is the data length, for m The correlation in dimensional space, for Correlation in dimensional space.

[0069] Then, the time domain eigenvalues, frequency domain eigenvalues, energy eigenvalues ​​and nonlinear eigenvalues ​​are fused to obtain a fused feature set. Feature fusion comprehensively reflects the working status of the arrester by combining different types of features into a high-dimensional feature vector. Principal component analysis (PCA) is performed on the fused feature set to obtain a reduced-dimensional feature set. PCA converts possibly correlated variables into linearly uncorrelated variables through orthogonal transformation, which are called principal components. The calculation of principal components is based on eigenvalue decomposition:

[0070]

[0071] in, is the covariance matrix, is the eigenvector matrix, is a diagonal matrix of eigenvalues.

[0072] The feature set after dimensionality reduction is sorted by feature importance to obtain the sorting result. Feature importance can be evaluated by the variance contribution rate of the principal component. Finally, feature selection is performed on the feature set after dimensionality reduction based on the sorting result to obtain a multidimensional feature vector. Feature selection usually retains the principal components with a cumulative variance contribution rate of 85% to 95%.

[0073] For example, the noise-reduced current data collected by the arrester on a subway line during a lightning strike event contains 5000 sampling points with a sampling frequency of 50kHz. First, the time domain characteristics are calculated to obtain time domain eigenvalues ​​such as mean 0.5kA, variance 25kA², and peak factor 3.2. The frequency domain data is obtained through FFT, and the frequency domain eigenvalues ​​such as fundamental frequency 50Hz, amplitude 8kA, and total harmonic distortion rate 2.5% are calculated. A 4-layer wavelet packet decomposition is performed to obtain the wavelet packet coefficients of 16 sub-bands. The energy of each sub-band is calculated to obtain the energy eigenvalue vector [100, 80, 60, 50, 40, 30, 25, 20, 15, 12, 10, 8,6, 5, 4, 3] (unit: kJ). The empirical mode decomposition obtains 5 IMFs and 1 residual. The approximate entropy of each IMF is calculated to obtain the nonlinear eigenvalues ​​[0.8, 0.7, 0.6, 0.5, 0.4, 0.3]. All features are fused to obtain a 100-dimensional feature vector. PCA is used for dimensionality reduction, and the first 10 principal components are retained, with a cumulative variance contribution rate of 92%. Finally, the principal components are sorted and selected according to their variance contribution rates to obtain a 10-dimensional multidimensional feature vector as the input for subsequent fault identification.

[0074] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0075] (1) Standardize the arrester fault characteristics of the multi-dimensional feature vector to obtain a standardized feature vector, and perform data enhancement on the standardized feature vector to obtain an enhanced arrester current feature data set;

[0076] (2) Design a convolution layer for the enhanced arrester current feature data set to obtain an arrester fault feature convolution kernel, and perform a convolution operation on the enhanced arrester current feature data set using the arrester fault feature convolution kernel to obtain an arrester fault feature map;

[0077] (3) Performing pooling layer processing on the arrester fault feature map to obtain a reduced-dimensional arrester fault feature map, and performing nonlinear activation on the reduced-dimensional arrester fault feature map to obtain an activated arrester fault feature;

[0078] (4) Perform a fully connected layer mapping on the activated arrester fault features to obtain the arrester fault classification vector, and perform softmax normalization on the arrester fault classification vector to obtain the arrester fault probability distribution;

[0079] (5) Perform threshold judgment on the probability distribution of lightning arrester faults to obtain preliminary lightning arrester fault classification results, and perform confidence assessment on the preliminary lightning arrester fault classification results to obtain the confidence of lightning arrester fault classification;

[0080] (6) The preliminary lightning arrester fault classification result is corrected according to the lightning arrester fault classification confidence to obtain the fault mode classification result.

[0081] Specifically, the arrester fault feature standardization process is performed on the multidimensional feature vector to obtain a standardized feature vector. The standardization process usually uses the Z-score standardization method to convert each eigenvalue into a standard normal distribution with a mean of 0 and a standard deviation of 1. It can eliminate the dimensional differences between different features and improve the training effect of subsequent models. Subsequently, the standardized feature vector is data enhanced to obtain an enhanced arrester current feature data set. Data enhancement techniques include adding Gaussian noise, random scaling, time shifting and other methods, the purpose of which is to expand the training samples and improve the generalization ability of the model. For example, 10 enhanced samples are generated for each original sample, and the size of the data set is expanded by 10 times.

[0082] Next, the enhanced arrester current feature data set is convolved to obtain the arrester fault feature convolution kernel. The convolution layer design includes determining the number, size, and step size of the convolution kernel. For arrester fault feature extraction, a multi-layer convolution structure is usually used, such as designing three convolution layers, each layer uses 32, 64, and 128 3x3 convolution kernels, respectively, with a step size of 1. The enhanced arrester current feature data set is convolved with the arrester fault feature convolution kernel to obtain the arrester fault feature map. The convolution operation can effectively extract local features and is particularly effective in capturing fault modes in arrester current signals. The arrester fault feature map is processed by the pooling layer to obtain the arrester fault feature map after dimensionality reduction. The pooling operation usually uses maximum pooling or average pooling, the purpose of which is to reduce the dimension of the feature map and extract the most significant features. For example, using 2x2 maximum pooling with a step size of 2 can reduce the size of the feature map by half. The arrester fault feature map after dimension reduction is activated nonlinearly to obtain the activated arrester fault feature. Common activation functions include ReLU (Rectified Linear Unit) and Leaky ReLU, which can introduce nonlinearity and enhance the expression ability of the model.

[0083] Then, the activated arrester fault features are mapped by the fully connected layer to obtain the arrester fault classification vector. The fully connected layer maps the features extracted by the convolutional network to the fault category space, and usually contains multiple hidden layers. For example, two hidden layers are designed with 256 and 128 nodes respectively. The arrester fault classification vector is softmax normalized to obtain the arrester fault probability distribution. The softmax function converts the classification vector into the probability value of each fault category to ensure that the sum of all probabilities is 1. Next, the arrester fault probability distribution is threshold judged to obtain the preliminary arrester fault classification result. The threshold judgment usually selects the category with the highest probability as the classification result, and the minimum probability threshold (such as 0.5) can be set to filter out judgments with low confidence. The preliminary arrester fault classification result is confidence evaluated to obtain the arrester fault classification confidence. The confidence evaluation can be based on the gap between the highest probability value and the second highest probability value, or by using methods such as information entropy.

[0084] Finally, the preliminary arrester fault classification results are corrected according to the arrester fault classification confidence to obtain the fault mode classification results. The correction process may include rejecting low-confidence classification results or manually reviewing samples with low confidence. For example, the current data collected during the operation of a batch of arresters (a total of 100) on a subway line was pre-processed to obtain a 10-dimensional feature vector for each arrester. First, these feature vectors are Z-score standardized, and the mean of each feature is adjusted to 0 and the standard deviation is adjusted to 1. Then, through data enhancement technology, 9 enhanced samples are generated for each original sample, and an enhanced data set of 1,000 samples is obtained.

[0085] Next, a neural network structure with three convolutional layers was designed. The first layer used 32 3x3 convolution kernels, the second layer used 64 3x3 convolution kernels, and the third layer used 128 3x3 convolution kernels. Each convolution layer was followed by a 2x2 maximum pooling layer and ReLU activation function. After the convolution operation, the size of the feature map gradually changed from the initial 10x1 to 8x1, 6x1, and 4x1. After pooling and activation function processing, 128 2x1 feature maps were obtained. These feature maps were flattened and input into two fully connected layers with 256 and 128 nodes respectively. The last layer is the output layer, and the number of nodes is equal to the number of fault categories (assuming 4 types: normal, slightly degraded, severely degraded, and failed). The probability distribution of each category is obtained through the softmax function. For example, for a specific sample, the probability distribution of the model output is [0.05, 0.15, 0.75, 0.05], which means that the probability of the sample being classified as “severely degraded” is the highest, which is 75%.

[0086] The probability threshold is set to 0.6, so this sample is initially classified as "severely degraded". By calculating the difference between the highest probability and the second highest probability (0.75 - 0.15 = 0.6) as the confidence index, the classification confidence of this sample is evaluated as high. Finally, this classification result is accepted, and the arrester is judged to be in a state of severe deterioration and needs to be maintained or replaced in time.

[0087] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0088] (1) Perform arrester fault type analysis on the fault mode classification results to obtain the fault type identification, and query the preset arrester fault weight library according to the fault type identification to obtain the corresponding fault weight coefficient;

[0089] (2) Quantify the abnormality of the arrester current characteristics in the fault mode classification results to obtain the characteristic abnormality index, and perform weighted calculation based on the characteristic abnormality index and the fault weight coefficient to obtain a preliminary arrester fault severity score;

[0090] (3) Conducting historical data comparison and analysis on the preliminary arrester fault severity scores to obtain fault trend indicators, and adjusting the preliminary arrester fault severity scores according to the fault trend indicators to obtain adjusted fault severity scores;

[0091] (4) Compare the adjusted fault severity score with the arrester safe operation threshold to obtain an over-limit index, and correct the adjusted fault severity score based on the over-limit index to obtain a corrected fault severity score;

[0092] (5) Conduct arrester system impact assessment on the corrected fault severity score to obtain the system impact factor, and calibrate the corrected fault severity score based on the system impact factor to obtain the calibrated fault severity score;

[0093] (6) The calibrated fault severity score is normalized to obtain the fault severity assessment score.

[0094] Specifically, the arrester fault type is parsed for the fault mode classification result to obtain a fault type identifier. The fault type identifier is a unique code used to represent different fault modes, such as "F001" for insulation aging, "F002" for valve breakdown, etc. According to the fault type identifier, the preset arrester fault weight library is queried to obtain the corresponding fault weight coefficient. The fault weight coefficient reflects the degree of influence of different fault types on the performance and safety of the arrester, and is usually derived from expert experience and historical data analysis. Next, the degree of abnormality of the arrester current characteristics in the fault mode classification result is quantified to obtain a characteristic abnormality index. The degree of abnormality quantification is to compare the current characteristics with the reference value under normal conditions and calculate the degree of deviation. The characteristic abnormality index can be calculated by the following formula:

[0095]

[0096] in, It is The abnormal indicators of the features, is the current eigenvalue, and are the mean and standard deviation of the feature under normal conditions. A weighted calculation is performed based on the feature abnormality index and the fault weight coefficient to obtain a preliminary arrester fault severity score. The calculation formula is as follows:

[0097]

[0098] in, is the preliminary fault severity score, It is The fault weight coefficient corresponding to the feature is: is the number of features. The preliminary arrester fault severity scores are compared and analyzed with historical data to obtain the fault trend index. The fault trend index reflects the rate of change of fault severity and can be calculated by the following formula:

[0099]

[0100] in, is the failure trend indicator, is the current preliminary fault severity score, is the fault severity score for the same period in history, is the time interval. The preliminary arrester fault severity score is adjusted according to the fault trend index to obtain the adjusted fault severity score. The adjustment formula is:

[0101]

[0102] in, is the adjusted fault severity score, is the trend influence coefficient.

[0103] The adjusted fault severity score is compared with the arrester safe operation threshold to obtain the excess degree index. The excess degree index calculation formula is:

[0104]

[0105] in, is an indicator of the degree of excess. is the safe operating threshold.

[0106] The adjusted fault severity score is corrected according to the over-limit index to obtain the corrected fault severity score. The correction formula is:

[0107]

[0108] in, is the corrected fault severity score, is the over-limit influence coefficient.

[0109] The modified fault severity score is evaluated for the arrester system impact to obtain the system impact factor. The system impact factor takes into account the importance of the arrester in the entire system and the chain reaction that the fault may cause. The modified fault severity score is calibrated according to the system impact factor to obtain the calibrated fault severity score. The calibration formula is:

[0110]

[0111] in, is the calibrated fault severity score, is the system influence coefficient, is the system influencing factor. Finally, the calibrated fault severity score is normalized to obtain the fault severity assessment score. The normalization formula is:

[0112]

[0113] in, is the final fault severity assessment score, and are the theoretical minimum and maximum values ​​of the score respectively.

[0114] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0115] (1) Reconstruct the time series of the pre-collected historical characteristic data and fault severity assessment scores to obtain the arrester fault prediction input sequence, and perform data standardization on the arrester fault prediction input sequence to obtain a standardized arrester fault prediction sequence;

[0116] (2) Optimize the input gate parameters of the long short-term memory network for the standardized arrester fault prediction sequence to obtain the optimized input gate parameters, and perform feature selection on the standardized arrester fault prediction sequence based on the optimized input gate parameters to obtain the key arrester fault feature sequence;

[0117] (3) Calculate the forget gate threshold of the long short-term memory network for the key arrester fault feature sequence to obtain the forget gate threshold, and filter the historical information of the key arrester fault feature sequence according to the forget gate threshold to obtain the filtered arrester fault feature sequence;

[0118] (4) Selecting the output gate activation function of the long short-term memory network for the screened arrester fault feature sequence to obtain the output gate activation function, and performing nonlinear mapping on the screened arrester fault feature sequence through the output gate activation function to obtain the arrester fault state prediction value;

[0119] (5) Perform error back propagation on the arrester fault state prediction value to obtain the network parameter update amount, and optimize the parameters of the long short-term memory network based on the network parameter update amount to obtain the optimized arrester fault prediction model;

[0120] (6) The optimized arrester fault prediction model is used to predict the operating status of the arrester in the future time period to obtain the future fault status prediction results.

[0121] Specifically, the pre-collected historical feature data and fault severity assessment scores are reconstructed in time series to obtain the arrester fault prediction input sequence. Time series reconstruction is to organize discrete historical data points into a continuous sequence in chronological order, and usually a fixed-length time window (such as 30 days) is selected as a prediction unit. The arrester fault prediction input sequence is subjected to data standardization to obtain a standardized arrester fault prediction sequence. The standardization process uses the Z-score method to convert each feature into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features. Next, the standardized arrester fault prediction sequence is optimized for the input gate parameters of the long short-term memory network to obtain the optimized input gate parameters. The input gate parameter optimization is to improve the sensitivity of the network to important features by adjusting the weight matrix and bias term of the input gate. The optimization process uses the stochastic gradient descent method to minimize the prediction error. According to the optimized input gate parameters, feature selection is performed on the standardized arrester fault prediction sequence to obtain the key arrester fault feature sequence. Feature selection is based on the importance score of the input gate parameters, retaining features with higher scores and removing redundant or irrelevant features, thereby reducing computational complexity and improving prediction accuracy.

[0122] The forget gate threshold of the long short-term memory network is calculated for the key arrester fault feature sequence to obtain the forget gate threshold. The forget gate threshold determines the degree to which the network retains historical information. By dynamically adjusting the threshold, the short-term and long-term memories can be balanced. The threshold calculation takes into account the temporal correlation and importance of the feature, and an adaptive algorithm is used for dynamic adjustment. The key arrester fault feature sequence is screened for historical information according to the forget gate threshold to obtain the screened arrester fault feature sequence. During the historical information screening process, the historical information below the threshold is attenuated or discarded, and the historical information that has an important impact on the current state prediction is retained. Then, the output gate activation function of the long short-term memory network is selected for the screened arrester fault feature sequence to obtain the output gate activation function. The selection of the activation function affects the nonlinear expression ability of the network. Commonly used activation functions include Sigmoid, Tanh, and ReLU. The selection process takes into account the distribution characteristics of the arrester fault features and the requirements of the prediction task. The screened arrester fault feature sequence is nonlinearly mapped by the output gate activation function to obtain the arrester fault state prediction value. Nonlinear mapping transforms the output of the LSTM unit into a final prediction result, such as a fault severity score at a future point in time.

[0123] The error back propagation is performed on the predicted value of the arrester fault state to obtain the network parameter update. Error back propagation is to calculate the difference between the predicted value and the actual value, and propagate this error back along the network structure to calculate the contribution of each parameter to the error. According to the network parameter update amount, the parameters of the long short-term memory network are optimized to obtain the optimized arrester fault prediction model. The parameter optimization adopts the gradient descent method, and the network parameters are continuously adjusted through multiple iterations to minimize the prediction error.

[0124] Finally, the optimized arrester fault prediction model is used to predict the arrester operating status in the future time period to obtain the future fault status prediction results. During the prediction process, the latest historical data is input into the optimized model to obtain the fault status prediction value in the future period (such as 7 days, 30 days).

[0125] For example, the daily fault severity scores and 10 key current features of a lightning arrester on a subway line for the past 90 days are used for fault prediction. First, the data are organized into a time series of 90 time steps and 11 features (10 current features plus 1 score) per step. Each feature is Z-score standardized, for example, the original score sequence [45.2, 46.8, 48.5, ...] is converted to [-0.8, -0.3, 0.2, ...]. Next, a network layer containing 64 LSTM units is designed. The input gate parameters are optimized by stochastic gradient descent, and the initial learning rate is set to 0.01. After 1000 iterations, the optimized input gate weight matrix is ​​obtained. Based on the optimized weights, the importance score of each feature is calculated, and the top 8 features with the highest scores are selected to obtain the key feature sequence.

[0126] Then, the forget gate threshold is calculated, the initial value is set to 0.5, and it is dynamically adjusted during the training process. For example, for the state at time t, if the forget gate output value of a feature is lower than 0.5, the historical information weight of the feature will be halved. For the filtered feature sequence, ReLU is selected as the output gate activation function, and nonlinear mapping is performed to obtain the predicted value.

[0127] During the training process, the data from the first 80 days were used for model training, and the data from the last 10 days were used for validation. The gradient was calculated by the back-propagation algorithm, and the parameters were updated using the Adam optimizer, with an initial learning rate of 0.001. After 100 epochs of training, the mean absolute error of the model on the validation set was reduced to 2.5 points (out of 100 points). Finally, the optimized model was used to predict the fault status for the next 7 days. The data from the last 90 days was input, and the model output the predicted values ​​of the daily fault severity scores for the next 7 days [52.3, 53.1, 54.0, 55.2, 56.5, 58.1, 60.0]. This prediction result shows that the fault status of the arrester has a trend of gradual deterioration, especially the score growth in the last two days is faster, suggesting that maintenance personnel need to conduct timely inspections and preventive maintenance to avoid possible serious failures.

[0128] The embodiment of the present invention also provides a lightning arrester fault detection system based on current data, such as Figure 2 As shown, the arrester fault detection system based on current data specifically includes:

[0129] The sampling module 201 is used to perform high-frequency sampling on the current signal of the subway lightning arrester grounding line to obtain original current data;

[0130] The noise reduction module 202 is used to perform wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction. The steps of performing wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction include: selecting a wavelet basis function on the original current data to obtain a wavelet basis function, and performing wavelet decomposition on the original current data according to the wavelet basis function to obtain wavelet coefficients; performing threshold calculation on the wavelet coefficients to obtain a noise reduction threshold, and performing soft threshold processing on the wavelet coefficients according to the noise reduction threshold to obtain the wavelet coefficients after noise reduction; performing wavelet decomposition on the wavelet coefficients after noise reduction Reconstruct to obtain preliminary noise reduction data, and perform residual noise evaluation on the preliminary noise reduction data to obtain a noise evaluation result; adaptively filter the preliminary noise reduction data according to the noise evaluation result to obtain filtered current data, and perform signal-to-noise ratio calculation on the filtered current data to obtain a signal-to-noise ratio index; perform threshold judgment on the signal-to-noise ratio index to obtain a judgment result, and confirm the noise reduction effect of the filtered current data according to the judgment result to obtain a noise reduction confirmation result; perform data calibration on the filtered current data according to the noise reduction confirmation result to obtain the noise-reduced current data;

[0131] An extraction module 203 is used to extract features from the noise-reduced current data to obtain a multi-dimensional feature vector;

[0132] The identification module 204 is used to perform fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result;

[0133] Scoring module 205, used to perform weighted scoring processing on the fault mode classification result to obtain a fault severity assessment score;

[0134] The prediction module 206 is used to perform fault prediction on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.

[0135] Through the collaborative work of the above modules, high-frequency sampling and wavelet transform noise reduction processing, this method effectively improves the quality of the original current data. High-frequency sampling can capture transient current characteristics, while wavelet transform noise reduction effectively removes environmental noise and electromagnetic interference, making the fault characteristics more prominent. Secondly, the multi-dimensional feature extraction technology comprehensively considers the time domain, frequency domain, energy and nonlinear characteristics, greatly enhancing the characterization ability of the fault mode. This multi-angle feature extraction method can more comprehensively describe the working state of the arrester and help identify complex and hidden fault modes. Furthermore, the use of convolutional neural networks for fault identification and classification significantly improves the accuracy and robustness of identification. The multi-layer structure of the convolutional neural network can automatically learn the intrinsic characteristics of the data, adapt to the fault mode under different working conditions, and reduce the subjectivity and uncertainty of human experience judgment. In addition, the evaluation of the fault severity adopts a multi-factor weighted calculation and dynamic adjustment mechanism, which not only considers the current state, but also incorporates historical trends and system impacts, making the evaluation results more comprehensive and accurate. This evaluation method can provide a more reliable basis for maintenance decisions and help to formulate a more reasonable maintenance plan. Finally, the application of long short-term memory network realizes the accurate prediction of the future state of the arrester, providing strong support for preventive maintenance. Through deep learning of historical data, this method can capture long-term dependencies and potential fault development trends, and warn of possible faults in advance, thereby greatly reducing the risk of sudden failures. In general, this arrester fault detection method based on current data integrates advanced signal processing, machine learning and deep learning technologies, and realizes the intelligence of the whole process from data acquisition, feature extraction, fault identification to state prediction. It not only improves the accuracy and real-time performance of fault detection, but also realizes fault prediction and preventive maintenance, effectively reducing the downtime and maintenance cost of the subway system.

[0136] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A lightning arrester fault detection method based on current data, characterized in that: include: Perform high-frequency sampling on the current signal of the subway arrester grounding line to obtain the original current data; The original current data is subjected to wavelet transform denoising processing to obtain the current data after denoising. The step of performing wavelet transform denoising processing on the original current data to obtain the current data after denoising comprises: performing wavelet basis function selection on the original current data to obtain the wavelet basis function, and performing wavelet decomposition on the original current data according to the wavelet basis function to obtain the wavelet coefficient; performing threshold calculation on the wavelet coefficient to obtain the denoising threshold, and performing soft threshold processing on the wavelet coefficient according to the denoising threshold to obtain the wavelet coefficient after denoising; and performing wavelet reconstruction on the wavelet coefficient after denoising. Obtaining preliminary noise reduction data, and performing residual noise evaluation on the preliminary noise reduction data to obtain a noise evaluation result; performing adaptive filtering on the preliminary noise reduction data according to the noise evaluation result to obtain filtered current data, and performing signal-to-noise ratio calculation on the filtered current data to obtain a signal-to-noise ratio index; performing threshold judgment on the signal-to-noise ratio index to obtain a judgment result, and confirming the noise reduction effect on the filtered current data according to the judgment result to obtain a noise reduction confirmation result; performing data calibration on the filtered current data according to the noise reduction confirmation result to obtain the noise-reduced current data; Performing feature extraction on the noise-reduced current data to obtain a multi-dimensional feature vector; Performing fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result; Performing weighted scoring processing on the fault mode classification results to obtain a fault severity assessment score; Fault prediction is performed on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.

2. The lightning arrester fault detection method based on current data according to claim 1, characterized in that: The step of performing high-frequency sampling on the current signal of the subway lightning arrester grounding line to obtain the original current data comprises: Installing a current sensor on the subway arrester grounding line to obtain a current acquisition device, and setting a sampling frequency of the current acquisition device to obtain a sampling parameter; Performing a numerical range check on the sampling parameter to obtain a valid sampling parameter, and performing a sampling time interval calculation on the current acquisition device according to the valid sampling parameter to obtain a sampling time interval; Performing an anti-interference capability evaluation on the sampling time interval to obtain an anti-interference evaluation result, and performing a sampling trigger setting on the current acquisition device according to the anti-interference evaluation result to obtain a sampling trigger signal; Synchronously calibrating the sampling trigger signal to obtain a calibrated trigger signal, and sampling the current of the current acquisition device according to the calibrated trigger signal to obtain analog current data; Performing analog-to-digital conversion on the analog current data to obtain digitized current data, and adding a timestamp to the digitized current data to obtain current data with a timestamp; The current data with the timestamp is subjected to data format standardization processing to obtain current data in a standard format, and the current data in the standard format is subjected to data integrity verification to obtain the original current data.

3. The lightning arrester fault detection method based on current data according to claim 1, characterized in that: The step of extracting features from the noise-reduced current data to obtain a multi-dimensional feature vector comprises: Performing time domain feature calculation on the noise-reduced current data to obtain time domain feature values, and performing frequency domain transformation on the noise-reduced current data to obtain frequency domain data; Performing frequency domain feature calculation on the frequency domain data to obtain frequency domain feature values, and performing wavelet packet decomposition on the noise-reduced current data to obtain wavelet packet coefficients; Performing energy feature extraction on the wavelet packet coefficients to obtain energy feature values, and performing empirical mode decomposition on the noise-reduced current data to obtain intrinsic mode functions; Performing nonlinear feature calculation on the intrinsic mode function to obtain nonlinear feature values, and performing feature fusion on the time domain feature values, frequency domain feature values, energy feature values ​​and nonlinear feature values ​​to obtain a fused feature set; Performing principal component analysis on the fused feature set to obtain a feature set after dimensionality reduction, and sorting the feature set after dimensionality reduction by feature importance to obtain a sorting result; Feature selection is performed on the feature set after dimensionality reduction according to the sorting result to obtain the multi-dimensional feature vector.

4. The arrester fault detection method based on current data according to claim 1, characterized in that: The step of performing fault identification and classification processing on the multidimensional feature vector through a convolutional neural network to obtain a fault mode classification result includes: Performing arrester fault feature standardization processing on the multidimensional feature vector to obtain a standardized feature vector, and performing data enhancement on the standardized feature vector to obtain an enhanced arrester current feature data set; Performing a convolution layer design on the enhanced arrester current feature data set to obtain an arrester fault feature convolution kernel, and performing a convolution operation on the enhanced arrester current feature data set through the arrester fault feature convolution kernel to obtain an arrester fault feature graph; Performing pooling layer processing on the arrester fault feature graph to obtain a reduced-dimensional arrester fault feature graph, and performing nonlinear activation on the reduced-dimensional arrester fault feature graph to obtain an activated arrester fault feature; Performing full connection layer mapping on the activated arrester fault feature to obtain an arrester fault classification vector, and performing softmax normalization on the arrester fault classification vector to obtain an arrester fault probability distribution; Performing a threshold judgment on the arrester fault probability distribution to obtain a preliminary arrester fault classification result, and performing a confidence evaluation on the preliminary arrester fault classification result to obtain an arrester fault classification confidence; The preliminary lightning arrester fault classification result is corrected according to the lightning arrester fault classification confidence to obtain the fault mode classification result.

5. The lightning arrester fault detection method based on current data according to claim 1, characterized in that: The step of performing weighted scoring processing on the fault mode classification result to obtain a fault severity assessment score comprises: The arrester fault type is parsed on the fault mode classification result to obtain a fault type identifier, and a preset arrester fault weight library is queried according to the fault type identifier to obtain a corresponding fault weight coefficient; quantifying the abnormality degree of the arrester current characteristics in the fault mode classification result to obtain a characteristic abnormality index, and performing weighted calculation based on the characteristic abnormality index and the fault weight coefficient to obtain a preliminary arrester fault severity score; Performing historical data comparison and analysis on the preliminary arrester fault severity score to obtain a fault trend index, and adjusting the preliminary arrester fault severity score according to the fault trend index to obtain an adjusted fault severity score; Comparing the adjusted fault severity score with the arrester safe operation threshold to obtain an over-limit index, and correcting the adjusted fault severity score according to the over-limit index to obtain a corrected fault severity score; Performing arrester system impact assessment on the corrected fault severity score to obtain a system impact factor, and calibrating the corrected fault severity score according to the system impact factor to obtain a calibrated fault severity score; The calibrated fault severity score is normalized to obtain the fault severity assessment score.

6. The lightning arrester fault detection method based on current data according to claim 5, characterized in that: The step of performing fault prediction on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result includes: Performing time series reconstruction on the pre-collected historical characteristic data and the fault severity assessment score to obtain a lightning arrester fault prediction input sequence, and performing data standardization processing on the lightning arrester fault prediction input sequence to obtain a standardized lightning arrester fault prediction sequence; Optimizing the input gate parameters of the long short-term memory network on the standardized arrester fault prediction sequence to obtain optimized input gate parameters, and performing feature selection on the standardized arrester fault prediction sequence according to the optimized input gate parameters to obtain a key arrester fault feature sequence; Performing a forget gate threshold calculation of a long short-term memory network on the key arrester fault feature sequence to obtain a forget gate threshold, and filtering historical information of the key arrester fault feature sequence according to the forget gate threshold to obtain a filtered arrester fault feature sequence; Selecting the output gate activation function of the long short-term memory network for the screened arrester fault feature sequence to obtain the output gate activation function, and performing nonlinear mapping on the screened arrester fault feature sequence through the output gate activation function to obtain the arrester fault state prediction value; Performing error back propagation on the arrester fault state prediction value to obtain a network parameter update amount, and performing parameter optimization on the long short-term memory network according to the network parameter update amount to obtain an optimized arrester fault prediction model; The optimized arrester fault prediction model is used to predict the arrester operating state in a future time period to obtain the future fault state prediction result.

7. A lightning arrester fault detection system based on current data, used to execute the lightning arrester fault detection method based on current data according to any one of claims 1 to 6, characterized in that: include: The sampling module is used to perform high-frequency sampling on the current signal of the subway lightning arrester grounding line to obtain the original current data; The noise reduction module is used to perform wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction. The step of performing wavelet transform noise reduction processing on the original current data to obtain the current data after noise reduction includes: selecting a wavelet basis function on the original current data to obtain a wavelet basis function, and performing wavelet decomposition on the original current data according to the wavelet basis function to obtain wavelet coefficients; performing threshold calculation on the wavelet coefficients to obtain a noise reduction threshold, and performing soft threshold processing on the wavelet coefficients according to the noise reduction threshold to obtain the wavelet coefficients after noise reduction; performing wavelet decomposition on the wavelet coefficients after noise reduction Reconstruct to obtain preliminary noise reduction data, and perform residual noise evaluation on the preliminary noise reduction data to obtain a noise evaluation result; adaptively filter the preliminary noise reduction data according to the noise evaluation result to obtain filtered current data, and perform signal-to-noise ratio calculation on the filtered current data to obtain a signal-to-noise ratio index; perform threshold judgment on the signal-to-noise ratio index to obtain a judgment result, and confirm the noise reduction effect of the filtered current data according to the judgment result to obtain a noise reduction confirmation result; perform data calibration on the filtered current data according to the noise reduction confirmation result to obtain the noise-reduced current data; An extraction module, used for performing feature extraction on the noise-reduced current data to obtain a multi-dimensional feature vector; An identification module is used to perform fault identification and classification processing on the multi-dimensional feature vector through a convolutional neural network to obtain a fault mode classification result; A scoring module is used to perform weighted scoring processing on the fault mode classification results to obtain a fault severity assessment score; The prediction module is used to perform fault prediction on the pre-collected historical feature data and the fault severity assessment score through a long short-term memory network to obtain a future fault state prediction result.

Citation Information

Patent Citations

  • Equipment fault intelligent prediction method based on small samples

    CN108802525A

  • Transmission line fault diagnosis method based on wavelet transform-multilayer perceptron

    CN116559728A

  • Intelligent bearing fault diagnosis method based on adaptive filtering

    CN118296335A

  • Method and apparatus for reducing noise in mass signal

    US20120298859A1