Overhead line fault diagnosis system based on LSTM and attention mechanism
The overhead line fault diagnosis system uses LSTM and attention mechanism to improve fault diagnosis accuracy by preprocessing signals and capturing long-term dependencies, addressing inefficiencies in existing methods through enhanced feature extraction and timely fault detection.
Patent Information
- Application Number
- CN202411287182.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2024-12-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing overhead line fault diagnosis methods rely heavily on manual inspections and expert experience, leading to inefficiencies and inaccuracies due to insufficient feature extraction of pulse sequence data, resulting in low fault diagnosis accuracy and reliability.
An overhead line fault diagnosis system based on LSTM and attention mechanism, including signal acquisition and preprocessing, feature extraction, fault diagnosis, and warning output modules, which utilizes phase correction, wavelet threshold denoising, and a pre-trained fault diagnosis model to capture long-term dependencies and adaptively adjust feature importance.
Enhances fault diagnosis accuracy by effectively preprocessing signals to remove noise, retaining critical features, and providing timely fault alarms through a comprehensive analysis of signal characteristics, enabling quick maintenance responses.
Smart Images
Figure CN119125767A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of power system fault diagnosis and relates to deep learning technology, specifically an overhead line fault diagnosis system based on LSTM and attention mechanism. Background Art
[0002] Overhead lines are an important part of the power transmission system, responsible for transmitting electricity from power plants to various users. However, due to the influence of the natural environment, such as wind, rain, lightning, freezing, etc., and aging caused by long-term operation, overhead lines are prone to failures, such as disconnection, short circuit, poor contact, etc., which not only affect the normal supply of electricity, but also may cause serious economic losses and safety hazards.
[0003] Traditional overhead line fault diagnosis methods mostly rely on manual inspections and expert experience, which is not only inefficient but also difficult to cover all potential fault points. Partial discharge patterns are key early warning signals for early faults in medium-voltage insulated overhead lines. By accurately recognizing patterns and extracting features from partial discharge information in the original signal, early fault diagnosis of overhead lines can be more effectively supported. However, in the actual signal acquisition process, especially under the interference of noise pulses, the quality of the observed data is often affected. Therefore, current methods usually use a threshold setting method to filter noise pulses, which can easily lead to incomplete noise filtering and even misjudgment. In addition, the existing methods do not fully extract the features of pulse sequence data and fail to fully reveal the fault characteristics in the signal, further affecting the accuracy and reliability of fault diagnosis. Summary of the invention
[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an overhead line fault diagnosis system based on LSTM and attention mechanism, which is used to solve the technical problem that the feature extraction of pulse sequence data in the existing method is insufficient, resulting in low fault diagnosis accuracy.
[0005] To achieve the above object, the present invention provides an overhead line fault diagnosis system based on LSTM and attention mechanism, including: a signal acquisition and preprocessing module, a feature extraction module, a fault diagnosis module and a warning output module;
[0006] The signal acquisition and preprocessing module is used to collect signal data in a real environment at a sampling point, perform preprocessing operations on the signal data, and obtain a preprocessed partial discharge pulse signal; wherein each signal data includes a signal ID;
[0007] The feature extraction module is used to sort the pre-processed partial discharge pulse signal into pulse sequence data in time sequence, and perform feature selection and calculation on the pulse sequence data to obtain a signal feature value set;
[0008] The fault diagnosis module is used to input the signal feature value set into a pre-trained fault diagnosis model and output a diagnosis result; wherein the diagnosis result includes normal or fault; the fault diagnosis model is constructed based on LSTM and attention mechanism;
[0009] The warning output module is used to start the warning function according to the diagnosis result and output the signal ID corresponding to the signal whose diagnosis result is a fault.
[0010] Based on the above technical solution, through the signal acquisition and preprocessing module, it is possible to accurately collect partial discharge pulse signals from complex environments, and perform effective data preprocessing on them, remove noise and irrelevant information, and retain signal features that are critical to fault diagnosis; based on commonly used feature selection, a new feature combination is designed to improve the characterization ability of signal features; a pre-trained model based on LSTM and attention mechanism is used to perform fault diagnosis on input data, effectively capturing the long-term dependency of time series data and adaptively adjusting the importance of features, which can provide more accurate fault diagnosis results; finally, the early warning output module is used to initiate an early warning based on the real-time diagnosis results, and the real-time location of the fault is achieved based on the signal ID, helping operation and maintenance personnel to respond quickly and perform necessary maintenance work.
[0011] Further, the signal data is preprocessed, comprising the following steps:
[0012] The first step is to perform phase correction on the signal data through discrete Fourier transform;
[0013] The second step is to use Savitzky-Golay filter to fit the low-frequency signal in the signal data and remove the low-frequency signal from the phase-corrected signal data;
[0014] The third step is to denoise the signal data after removing the low-frequency signal through the SureShrink threshold wavelet of the soft threshold function;
[0015] In the fourth step, according to the preset threshold and the preset minimum distance, a fixed amplitude threshold is used to mark the peak signals of all points in the denoised signal data whose amplitudes and amplitudes are higher than the preset threshold and whose distances from each other are greater than the minimum distance, and further set constraints to constrain the maximum distance between pulse peak pairs, the height difference between adjacent peaks, the maximum peak amplitude, and the maximum number of peaks. Peaks that do not meet the constraints are marked as error peaks, and the error peaks are screened out from the peak signal to obtain the preprocessed local discharge pulse signal.
[0016] The three-phase signals in the signal data are processed separately by phase correction and wavelet threshold denoising to reduce interference. In order to extract the partial discharge signal, the signal data is further subjected to signal detrending, pulse detection and error pulse screening, which effectively eliminates external interference and background noise, improves signal quality, and lays the foundation for subsequent accurate feature extraction and fault diagnosis.
[0017] Further, the signal feature value set includes: approximate entropy, Katz fractal dimension, fractal dimension, mean, standard deviation, floating upper limit, floating lower limit, block amplitude, permutation singular entropy index, sample Pete entropy index, mean standard deviation ratio and seven percentiles;
[0018] It should be noted that, among them, the approximate entropy, Katz fractal dimension, fractal dimension, mean, standard deviation, floating upper limit, floating lower limit, and block amplitude can be calculated by Python programming software on the pulse sequence data.
[0019] Furthermore, the method for obtaining the permutation singular entropy index includes:
[0020] A1, data segmentation: segment the pulse sequence data to obtain subsequences of different lengths;
[0021] A2, calculate the permutation entropy: by formula PE = -∑p i log 2 (p i ) calculates the permutation entropy PE of the pulse sequence data, where p i is the probability of the i-th subsequence appearing, n is the number of subsequences;
[0022] A3, calculate the singular entropy: by formula The singular entropy SE of the pulse sequence data is calculated, where f i is the frequency of occurrence of the ith subsequence, and n is the number of subsequences;
[0023] A4, calculate the permutation singular entropy index: by formula The permutation singular entropy index PESI was calculated.
[0024] It should be noted that in the present invention, the permuted singular entropy index is used as one of the input features of the fault diagnosis model, which not only includes the randomness and structural information of the signal, but also reflects the stability and dynamic changes of the signal under different operating conditions, and can more comprehensively describe the characteristics of the signal.
[0025] Furthermore, the method of obtaining the sample Petr entropy includes:
[0026] B1, data segmentation: segment the pulse sequence data to obtain subsequences of different lengths;
[0027] B2, calculate sample entropy: by formula The sample entropy SampEn of the pulse sequence data is calculated, where C is the number of matches between two consecutive subsequences of the same length, U is the number of matches between two consecutive subsequences of different lengths, and m is the delay time or similarity threshold;
[0028] B3, calculate the Petrosian fractal dimension: by formula The Petrosian fractal dimension Dp of the pulse sequence data was calculated, where σ represents the area of the smallest square covering the signal, and C represents the number of smallest rectangles containing the square;
[0029] B4, calculate the sample Peter entropy: through the formula The sample Peter entropy SPEI is calculated;
[0030] It should be noted that in the present invention, the sample Petr entropy is used as one of the input features of the fault diagnosis model, which can comprehensively evaluate the complexity and irregularity of time series data, describe the roughness of the time series in the time dimension, help capture signal changes, and identify the difference between normal operating state and fault state.
[0031] Furthermore, the method for obtaining the mean standard deviation ratio includes:
[0032] C1, calculate the mean and standard deviation: by formula and Calculate the mean μ and standard deviation σ of the pulse sequence data, where x i (i=1,2,....n) represents a set of pulse sequence data, and n represents the total number of data points;
[0033] C2, calculate the mean standard deviation ratio: by formula The mean standard deviation MSDR of the pulse sequence data is calculated;
[0034] It should be noted that in the present invention, the mean-standard deviation ratio is used as one of the input features of the fault diagnosis model, which can reflect the degree of deviation of a single data point relative to the average level of the data set and improve the stability of the model in processing abnormal signal data.
[0035] Furthermore, the seven percentiles are obtained by respectively extracting the values at seven positions of 0%, 1%, 25%, 50%, 75%, 99%, and 100% between the maximum value and the minimum value in the above data.
[0036] By calculating characteristic values such as approximate entropy, Katz fractal dimension, and block amplitude, the complexity of the pulse sequence can be analyzed in detail; and by introducing the permutation singular entropy index, sample Pett entropy index, and calculating the mean-standard deviation ratio, the essential characteristics of the signal can be revealed through a comprehensive analysis of multiple dimensions, helping the fault diagnosis model to analyze data characteristics more deeply and enhancing the model's ability to judge abnormal signals.
[0037] Furthermore, the pre-trained fault diagnosis model is constructed based on deep learning, including the following steps:
[0038] D1, model construction: construct a fault diagnosis model based on LSTM network structure and attention mechanism;
[0039] D2, data preprocessing: perform preprocessing operations on some signal data in the open source data set to obtain preprocessed signal data;
[0040] D3, feature engineering and feature extraction: sorting the preprocessed signal data into pulse sequence data in chronological order, and dividing the pulse sequence data into several groups, performing feature selection and calculation on each group to obtain a feature value set;
[0041] D4, model training: the feature value set is divided into a training set and a validation set according to a 5-fold cross-validation method, and input into the fault diagnosis model for training and validation to obtain a pre-trained fault diagnosis model.
[0042] It should be noted that the preprocessing operation in D2 is the same as the preprocessing operation in the signal acquisition and preprocessing module; the feature selection and calculation in D3 are the same as the operations in the feature extraction module.
[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: through the signal acquisition and preprocessing module, the signal data collected from the complex environment can be effectively preprocessed, the interference of background noise and other pulse signals can be reduced, and the important signal characteristics can be retained; based on the commonly used feature selection method, a new feature combination scheme is designed, including the permutation singular entropy index, the sample Pete entropy index and the mean standard deviation ratio, which enhances the model's ability to judge the degree of signal abnormality, and reveals the essential characteristics of the signal through a comprehensive analysis of multiple dimensions; fault diagnosis is performed using a pre-trained model based on the LSTM network structure and the attention mechanism, which effectively captures the long-term dependency of time series data, and adaptively adjusts the importance of features, providing more accurate fault diagnosis results; finally, an early warning output module is added to achieve timely fault alarms, helping operation and maintenance personnel to respond quickly and perform necessary maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0045] Figure 1 A schematic flow chart of an overhead line fault diagnosis system based on LSTM and attention mechanism provided by the present invention;
[0046] Figure 2 A schematic flow chart of the signal data preprocessing operation provided by the present invention. DETAILED DESCRIPTION
[0047] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than 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.
[0048] See also Figure 1-Figure 2 , an embodiment of the present invention provides an overhead line fault diagnosis system based on LSTM and attention mechanism, including: a signal acquisition and preprocessing module, a feature extraction module, a fault diagnosis module and a warning output module which are progressively connected;
[0049] The signal acquisition and preprocessing module is used to collect signal data in a real environment at a sampling point, perform preprocessing operations on the signal data, and obtain a preprocessed partial discharge pulse signal; wherein each signal data includes a signal ID;
[0050] The feature extraction module is used to sort the pre-processed partial discharge pulse signal into pulse sequence data in time sequence, and perform feature selection and calculation on the pulse sequence data to obtain a signal feature value set;
[0051] The fault diagnosis module is used to input the signal feature value set into a pre-trained fault diagnosis model and output a diagnosis result; wherein the diagnosis result includes normal or fault; the fault diagnosis model is constructed based on LSTM and attention mechanism;
[0052] The warning output module is used to start the warning function according to the diagnosis result and output the signal ID corresponding to the signal whose diagnosis result is a fault.
[0053] Before conducting early fault diagnosis of partial discharge on insulated overhead lines, it is critical to preprocess the collected signal data. Since the data is collected in a real outdoor environment, there will be many interference factors and background noise in the signal, which will cause the signal phase to change, increase the difficulty of partial discharge detection, and reduce the accuracy of subsequent fault diagnosis. Therefore, it is necessary to preprocess the collected signal data to extract clean partial discharge signals;
[0054] In this embodiment, the collected signal data needs to be preprocessed by phase correction, signal detrending, wavelet threshold denoising, pulse detection, and error pulse screening;
[0055] Since overhead lines generally use a three-phase three-wire system, three-phase data is used simultaneously, so the collected signal data is composed of the three-phase signals of the medium-voltage insulated overhead line. The three-phase signal may be affected by various factors during the collection and transmission process, causing the signal phase to change, which will have an adverse effect on the analysis and processing of the signal, thereby affecting the subsequent extraction of three-phase features and the final fault diagnosis results. Therefore, phase correction processing must be performed first. In this embodiment, the original signal data is phase corrected by digital signal processing, and the steps are as follows:
[0056] S1, decoupling three-phase signals: first extract the three-phase signals separately so that each phase can be processed independently later;
[0057] S2, perform 50 Hz Fourier transform: perform 50 Hz Fourier transform on the signal of each phase through the Fourier transform algorithm to obtain the amplitude and phase information of the 50 Hz component;
[0058] S3, select reference phase: select a 50 Hz component of one phase in the three-phase signal as a reference;
[0059] S4, calculating the phase difference: taking the selected reference phase as a reference, calculating the phase difference of the other two phase signals relative to the reference phase;
[0060] S5, phase correction: using the calculated phase difference as the phase offset, and performing phase correction on the other two phase signals by adjusting the phase of the signal or introducing a phase shifter;
[0061] S6, resynthesize three-phase signal: after phase correction, resynthesize the three phase signals into a three-phase signal, maintaining a phase difference of 120 degrees.
[0062] In order to effectively analyze the partial discharge signal in the data and eliminate or weaken the power frequency interference in the collected data, it is necessary to perform signal detrending on the phase-corrected data to remove the power frequency waveform in the signal. In this embodiment, the original data is processed by signal smoothing technology to fit the low-frequency part of the signal, including the power frequency waveform and its possible harmonic components, and then the low-frequency signal is subtracted from the phase-corrected signal. The processed signal mainly contains high-frequency partial discharge pulses, which reduces the interference of the power frequency signal, making these high-frequency pulses more prominent in the waveform, which is convenient for the subsequent observation of the pulse signal and the extraction of the characteristics of the partial discharge signal.
[0063] Specifically, a Savitzky-Golay filter is used to fit the low-frequency part of the signal. The Savitzky-Golay filter mainly includes two key parameters: the filter window length w and the polynomial fitting order k. If the window length w is too long, the effective data signal will be lost; if w is too short, the noise reduction performance of the filter will be reduced. When the k value is too high, new noise may appear in the signal; when the k value is too low, the signal may be distorted. In this embodiment, the filter window length is set to 99, and the filter fitting polynomial order is set to 3. The data after phase correction is then input into the set Savitzky-Golay filter, and a high-frequency pulse signal is obtained after fitting and filtering inside the filter.
[0064] Ideally, partial discharge patterns can be classified by analyzing the specific combination of partial discharge phase distribution, pulse amplitude, and time variation. However, in the partial discharge measurement process, external noise interference will make feature extraction difficult, thereby reducing the sensitivity of partial discharge measurement detection;
[0065] Therefore, in this embodiment, the wavelet threshold denoising algorithm is used to further denoise the signal data after trend removal. Assume that the noisy signal model of the signal data is: f(n) = s(n) + δ·e(n), where f(n) is the noisy signal; s(n) is the real signal; δ is the standard deviation of the noise coefficient; e(n) is the noise signal. The specific steps of using wavelet threshold denoising are as follows:
[0066] (1) Wavelet decomposition of noisy signals. According to the characteristics of the signal data, select the appropriate wavelet basis function and the number of wavelet decomposition layers, perform wavelet decomposition on the signal, and obtain the wavelet coefficients as follows: ω s (j,k)=ω x (j,k)+ω n (j,k), where j represents the number of decomposition levels, k represents the time shift, and ω s (j,k) represents the noisy signal, ω x (j,k) represents a pure signal, ω n(j,k) represents the wavelet coefficient of the noise signal located on the jth layer;
[0067] (2) Threshold denoising. Select appropriate wavelet threshold and threshold function, process the wavelet coefficients after wavelet decomposition, so as to set the coefficients of the noise in the signal to zero, extract the coefficients of the useful part of the signal, and obtain new wavelet coefficients to achieve the purpose of denoising;
[0068] (3) Wavelet reconstruction: The wavelet coefficients after denoising are reconstructed by wavelet inverse transform, and finally the denoised signal is obtained.
[0069] The signal data collected in the real environment is often complex, containing multiple frequency components and noise. In fault diagnosis, the pulse signal may correspond to a specific fault mode of the equipment, so accurate detection of pulses is crucial for extracting information and diagnosing faults. After wavelet threshold denoising, by setting the peak threshold to 0.6, the amplitude threshold to 4.0, and the minimum distance between peaks to 0, all points in the signal with amplitudes higher than the threshold and distances greater than the minimum distance are considered as peaks, and the pulse signal is found.
[0070] Due to the pulse interference in the overhead line partial discharge noise interference, an erroneous peak will be generated in the time domain of the original signal, which will be mistaken for a partial discharge pulse signal and interfere with the subsequent experimental results. Therefore, it is necessary to traverse the pulse signal checked above, and by constraining the maximum distance between pulse peak pairs, the height difference between adjacent peaks, the maximum peak amplitude, and the maximum number of peaks, the erroneous pulse signal is screened out and eliminated. In this embodiment, the peaks with a distance between adjacent peaks less than 10, a maximum length of a pulse train less than 20, and a peak height of no more than 0.75 are set as erroneous peaks. After selection, they are screened out from the peak signal to obtain the final pre-processed partial discharge pulse signal, thereby enhancing the reliability of signal analysis and ensuring the accuracy of the analysis results.
[0071] Extraction of characteristic values of partial discharge signals is a key step in early fault diagnosis, and the effect of feature extraction will directly affect the results of fault diagnosis. The pre-processed partial discharge pulse signal will be further input into the feature extraction module, sorted into pulse sequence data in chronological order, and the signal characteristic value set will be obtained by feature selection and calculation of the pulse sequence data;
[0072] In this embodiment, the characteristics of the pulse sequence data are selected as the input characteristics of the fault diagnosis model. The discrete distribution degree of the time series is better expressed through the time series entropy and the percentiles of each proportion, reflecting the data characteristics. Among them, the commonly used characteristic indicators include approximate entropy, Katz fractal dimension, fractal dimension, mean, standard deviation, floating upper limit, floating lower limit and block amplitude, which can be calculated by Python programming software for the pulse sequence data. At the same time, the permutation singular entropy index, sample Pete entropy index and mean standard deviation ratio are added to enrich the input characteristics, helping the fault diagnosis model to analyze the data characteristics more deeply;
[0073] Specifically, the permutation singular entropy index is calculated by permutation entropy and singular entropy. First, the pulse sequence data is segmented to obtain subsequences of different lengths; through the formula PE = -∑p i log 2 (p i ) calculates the permutation entropy PE of the pulse sequence data, where p i is the probability of the occurrence of the ith subsequence, n is the number of subsequences; by the formula The singular entropy SE of the pulse sequence data is calculated, where f i is the frequency of occurrence of the ith subsequence, and n is the number of subsequences; thus reflecting the stability and dynamic changes of the signal under different operating conditions;
[0074] The sample bit entropy is calculated by the sample entropy and the Petrosian fractal dimension. First, the pulse sequence data is segmented to obtain subsequences of different lengths; The sample entropy SampEn of the pulse sequence data is calculated, where C is the number of matches between two consecutive subsequences of the same length, U is the number of matches between two consecutive subsequences of different lengths, and m is the delay time or similarity threshold; The Petrosian fractal dimension Dp of the pulse sequence data is calculated, where σ represents the area of the smallest square covering the signal, and C represents the number of smallest rectangles containing the square; The sample Peter entropy SPEI is calculated to comprehensively evaluate the complexity and irregularity of time series data, describe the roughness of the time series in the time dimension, help the model capture signal changes, and better diagnose fault signals;
[0075] The mean-standard deviation ratio is obtained by calculating the ratio of the mean μ and standard deviation σ of the pulse sequence data. The calculation formula is: MSDR stands for mean standard deviation ratio, which reflects the degree of deviation of a single data point from the average level of the data set and improves the stability of the model in processing abnormal signal data;
[0076] Finally, the values at seven positions, 0%, 1%, 25%, 50%, 75%, 99%, and 100%, between the maximum and minimum values of the above data are extracted respectively to form percentiles of seven proportions, providing more comprehensive data distribution information.
[0077] The eigenvalues calculated above are combined into a eigenvalue set and input into the pre-training constructed using the deep learning algorithm to obtain the diagnosis result. Among them, the signal with the diagnosis result of fault will further activate the warning function of the warning output module, and output the signal ID corresponding to the fault signal, so as to realize timely fault alarm and help the operation and maintenance personnel to respond quickly and perform necessary maintenance work.
[0078] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.
[0079] Working principle of the present invention:
[0080] S1, collects signal data from the real environment through the signal acquisition and preprocessing module, and performs preprocessing operations on these signals to extract clean partial discharge signals, reduce noise and interference, and improve the accuracy of subsequent fault diagnosis;
[0081] S2, through the feature extraction module, the pre-processed partial discharge pulse signal is sorted into pulse sequence data in time sequence, and various eigenvalues are calculated to form a eigenvalue set, which comprehensively reflects the different characteristics of the signal and provides rich feature information for fault diagnosis;
[0082] S3, through the fault diagnosis module, inputs the feature value set into the pre-trained fault diagnosis model based on LSTM and attention mechanism to identify the normal and fault states of the signal;
[0083] S4, the early warning function is activated according to the detected fault signal through the early warning output module, and the signal ID corresponding to the fault signal is output so that the operation and maintenance personnel can respond quickly and take necessary maintenance measures.
[0084] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.
Claims
1. Overhead line fault diagnosis system based on LSTM and attention mechanism, It is characterized in that Signal acquisition and preprocessing module, feature extraction module, fault diagnosis module and early warning output module; The signal acquisition and preprocessing module is used to collect signal data in a real environment at a sampling point, perform preprocessing operations on the signal data, and obtain a preprocessed partial discharge pulse signal; wherein each signal data includes a signal ID; The feature extraction module is used to sort the pre-processed partial discharge pulse signal into pulse sequence data in time sequence, and perform feature selection and calculation on the pulse sequence data to obtain a signal feature value set; The fault diagnosis module is used to input the signal feature value set into a pre-trained fault diagnosis model and output a diagnosis result; wherein the diagnosis result includes normal or fault; the fault diagnosis model is constructed based on LSTM and attention mechanism; The warning output module is used to start the warning function according to the diagnosis result and output the signal ID corresponding to the signal with the diagnosis result of fault.
2. According to the overhead line fault diagnosis system based on LSTM and attention mechanism according to claim 1, It is characterized in that The signal data is preprocessed, comprising the following steps: In the first step, the signal data is phase corrected by discrete Fourier transform; The second step is to use Savitzky-Golay filter to fit the low-frequency signal in the signal data and remove the low-frequency signal from the phase-corrected signal data; The third step is to denoise the signal data after removing the low-frequency signal through the SureShrink threshold wavelet of the soft threshold function; In the fourth step, according to the preset threshold and the preset minimum distance, a fixed amplitude threshold is used to mark the peak signals of all points in the denoised signal data whose amplitudes and amplitudes are higher than the preset threshold and whose distances from each other are greater than the minimum distance, and further set constraints to constrain the maximum distance between pulse peak pairs, the height difference between adjacent peaks, the maximum peak amplitude, and the maximum number of peaks. Peaks that do not meet the constraints are marked as error peaks, and the error peaks are screened out from the peak signal to obtain the preprocessed local discharge pulse signal.
3. According to the overhead line fault diagnosis system based on LSTM and attention mechanism according to claim 1, It is characterized in that The signal feature value set includes: approximate entropy, Katz fractal dimension, fractal dimension, mean, standard deviation, floating upper limit, floating lower limit, block amplitude, permutation singular entropy index, sample Pett entropy index, mean standard deviation ratio and seven percentiles.
4. According to claim 3, the overhead line fault diagnosis system based on LSTM and attention mechanism, It is characterized in that The method for obtaining the permutation singular entropy index includes: A1, data segmentation: segment the pulse sequence data to obtain subsequences of different lengths; A2, calculate the permutation entropy: by formula PE = -∑p i log 2 (p i ) calculates the permutation entropy PE of the pulse sequence data, where p i is the probability of the i-th subsequence appearing, n is the number of subsequences; A3, calculate the singular entropy: by formula The singular entropy SE of the pulse sequence data is calculated, where f i is the frequency of occurrence of the ith subsequence, and n is the number of subsequences; A4, calculate the permutation singular entropy index: by formula The permutation singular entropy index PESI was calculated.
5. According to claim 3, the overhead line fault diagnosis system based on LSTM and attention mechanism, It is characterized in that The method of obtaining the sample Petr entropy includes: B1, data segmentation: segment the pulse sequence data to obtain subsequences of different lengths; B2, calculate sample entropy: by formula The sample entropy SampEn of the pulse sequence data is calculated, where C is the number of matches between two consecutive subsequences of the same length, U is the number of matches between two consecutive subsequences of different lengths, and m is the delay time or similarity threshold; B3, calculate the Petrosian fractal dimension: by formula The Petrosian fractal dimension Dp of the pulse sequence data was calculated, where σ represents the area of the smallest square covering the signal, and C represents the number of the smallest rectangles containing the square; B4, calculate the sample Peter entropy: through the formula The sample Petr entropy SPEI is calculated.
6. The overhead line fault diagnosis system based on LSTM and attention mechanism according to claim 3, It is characterized in that The method for obtaining the mean standard deviation ratio includes: C1, calculate the mean and standard deviation: by formula and Calculate the mean μ and standard deviation σ of the pulse sequence data, where x i (i=1,2,....n) represents a set of pulse sequence data, and n represents the total number of data points; C2, calculate the mean standard deviation ratio: by formula The mean standard deviation MSDR of the pulse sequence data was calculated.
7. The overhead line fault diagnosis system based on LSTM and attention mechanism according to claim 3, It is characterized in that The seven percentiles are obtained by respectively extracting the values at seven positions, 0%, 1%, 25%, 50%, 75%, 99%, and 100%, between the maximum value and the minimum value in the above data.
8. The overhead line fault diagnosis system based on LSTM and attention mechanism according to claim 1, It is characterized in that The pre-trained fault diagnosis model is constructed based on deep learning and includes the following steps: D1, model construction: construct a fault diagnosis model based on LSTM network structure and attention mechanism; D2, data preprocessing: perform preprocessing operations on some signal data in the open source data set to obtain preprocessed signal data; D3, feature engineering and feature extraction: sorting the preprocessed signal data into pulse sequence data in chronological order, and dividing the pulse sequence data into several groups, performing feature selection and calculation on each group to obtain a feature value set; D4, model training: the feature value set is divided into a training set and a validation set according to a 5-fold cross-validation method, and input into the fault diagnosis model for training and validation to obtain a pre-trained fault diagnosis model.