Power transmission cable ground fault adaptive diagnosis method and system

Abstract

The invention discloses a power transmission cable ground fault adaptive diagnosis method and system, and the method comprises the steps: collecting multi-modal data, and carrying out the preprocessing and extraction of the multi-modal data, and forming a statistical feature vector; the method comprises the following steps: adaptively decomposing sheath current data by adopting OVMD, calculating segmented TKEO energy operators, and realizing fault early warning through a ratio criterion of maximum values of TKEO derivatives of adjacent segments; inputting the real-time acquisition time sequence working condition class feature vector into a trained LSTM to obtain a fault probability, dynamically updating a fault probability threshold based on LSTM output, and judging that a fault occurs when the fault probability exceeds a continuous set period of the fault probability threshold; and inputting the statistical class feature vector into ANN, inputting the sheath current data into 1D-CNN for fault classification, and performing adaptive weighted voting integrated decision on the output of the two types of networks to identify the fault type. According to the invention, the problems of low fault detection sensitivity, poor positioning precision, lack of early warning capability and the like in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-voltage cable fault diagnosis and intelligent monitoring, and particularly relates to a power transmission cable grounding fault adaptive diagnosis method and system. BACKGROUND

[0002] With the rapid development of regional power grids, high-voltage cables are increasingly widely used in power transmission systems. According to statistics, by the end of 2024, the total length of 110kV voltage level power cable lines in China has exceeded 600,000 kilometers, and is growing at a rate of more than 15% per year. High-voltage cable lines usually adopt a cross-connection grounding mode to limit the shielded layer induced voltage, but this complex grounding structure also increases the difficulty of fault diagnosis.

[0003] According to the statistics of State Grid Corporation of China, cable faults account for more than 35% of the total number of power transmission system faults, and more than 60% of permanent faults are caused by early weak faults gradually developing. There are the following technical difficulties in the current high-voltage cable fault diagnosis: 1. Single function of traditional grounding box: The existing grounding box only has basic grounding protection function, lacks fault detection and diagnosis capability, and cannot monitor the cable operation state in real time. More than 90% of the grounding boxes on the market do not have intelligent monitoring function, which leads to late fault discovery and the average fault handling time is more than 8 hours.

[0004] 2. Difficult fault location: The fault location of the cable under the cross-connection grounding mode is more complex than that of the single-end grounding or two-end direct grounding. The traditional traveling wave method, impedance method and other fault location methods have a positioning error of 5%-10% in the cross-connection system. For a 30km cable line, the positioning error can reach 1500-3000m, which seriously affects the fault repair efficiency.

[0005] 3. Insufficient early fault warning: Studies have shown that more than 70% of permanent cable faults will have early fault characteristics such as intermittent discharge and local overheating within 2-48 hours before the fault occurs. However, these early fault signals are weak, usually less than 5% of normal signals, and the existing monitoring system cannot effectively detect them, resulting in missed best repair opportunity.

[0006] 4. Limited data processing capability: Traditional methods rely on manual setting of fixed thresholds. If the threshold is set too high, it will lead to missed reports, and if it is set too low, it will cause frequent false alarms. According to statistics, the false alarm rate of the fault detection method based on fixed threshold is as high as 30%, and the missed report rate reaches 15%. Moreover, the fixed threshold cannot adapt to dynamic conditions such as seasonal changes and load fluctuations.

[0007] 5. Lack of intelligent diagnosis capability: The existing system is mainly for simple data collection and threshold alarm, and lacks intelligent analysis capability for fault type, fault severity and development trend, and cannot provide decision support for operation and maintenance personnel.

[0008] In the prior art, Chinese patent CN108646144A discloses a high-voltage single-core cable short-circuit fault offline ranging method, device and system, but only the sheath current can be monitored, and the fault type cannot be identified; Chinese patent CN115173093B discloses a cable intelligent grounding box with a fault positioning function, which adopts GPS time service and high-frequency transient current detection, but lacks early fault warning capability and intelligent diagnosis function. SUMMARY

[0009] The purpose of the present application is to provide a power transmission cable grounding fault adaptive diagnosis method and system to solve the problems of low fault detection sensitivity, poor positioning accuracy, lack of early warning capability and the like in the prior art, and to realize intelligent detection, section-level positioning, type identification and early warning of high-voltage cable faults, thereby significantly improving the operation and maintenance efficiency and safety of cable lines.

[0010] The technical solution of the present application is as follows: On the one hand, the present application provides a power transmission cable grounding fault adaptive diagnosis method, comprising the following steps: Collecting multi-modal data including sheath current, three-phase core current, three-phase voltage, temperature and humidity, vibration and partial discharge, and performing preprocessing to extract time-domain features, frequency-domain features, time-frequency domain features and environmental and state features to form a statistical feature vector; Adopting OVMD to adaptively decompose the sheath current data, calculating the segmented TKEO energy operator, and realizing early warning of faults through the maximum value ratio criterion of the derivatives of adjacent segments TKEO; Collecting time-series working condition feature vectors as input vectors of LSTM for training, inputting the real-time collected time-series working condition feature vectors into the trained LSTM and mapping the fault probability through the Sigmoid function, and dynamically updating the fault probability threshold based on the LSTM output, and determining that a fault has occurred when the fault probability exceeds the fault probability threshold for a continuous set period; Inputting the statistical feature vector into ANN and the sheath current data into 1D-CNN for fault classification respectively, and adopting adaptive weighted voting integrated decision for the outputs of the two types of networks to identify the fault type.

[0011] Preferably, under the trigger of GPS / Beidou time service signal, all multi-modal data are synchronously collected by sensors, and the preprocessing of the collected multi-modal data specifically includes: De-noising: wavelet threshold de-noising is adopted, the mother wavelet is db4, the decomposition layer is 5 layers, and soft threshold processing is adopted; filtering: notch filter is adopted to remove power frequency interference; normalization: Min-Max normalization is adopted to normalize the multi-modal data to the interval [0, 1]; The statistical feature vector is composed of time domain features, frequency domain features, time-frequency domain features and environment and state features extracted from the pretreated multi-modal data, and is specifically: The preliminary time domain features including effective value, peak value, peak factor, waveform factor, pulse factor, margin factor, skewness and kurtosis are extracted from the multi-modal data, and 8 types of time domain features are selected from all the preliminary time domain features according to the importance analysis and fault sensitivity sorting to form the time domain features of the feature vector; The time domain signals of the multi-modal data are converted into frequency domain signals by fast Fourier transform, and the frequency domain features including fundamental amplitude, harmonic content and total harmonic distortion rate are extracted from the frequency domain signals; The multi-modal data is decomposed by 5 layers of db4 wavelet, and the preset number of frequency bands with the largest energy proportion in the wavelet packet energy distribution are selected as the time-frequency domain features; The environment and state feature vector including three-phase core current effective value and sheath current effective value is extracted.

[0012] Preferably, the OVMD adaptive decomposition of the sheath current data is adopted to calculate the segmented TKEO energy operator, and the early fault warning is realized by the maximum value ratio criterion of adjacent segment TKEO derivatives, and is specifically: The OVMD adaptive decomposition of the sheath current data is adopted to obtain a plurality of IMFs, the pulse factor of each IMF and the cosine similarity of the original sheath current data are calculated, and the two IMFs with the maximum pulse factor and the highest similarity are selected for reconstruction to obtain a reconstructed signal; The TKEO energy operator is calculated for the reconstructed signal:

[0013] Wherein, TKEO energy operator; is the sampling point index of the discrete time sequence, value is 1, 2, … , is the total number of sampling points of the sheath current signal; represents the instantaneous amplitude of the reconstructed signal at the sampling point, and the unit is ampere; represents the instantaneous amplitude of the reconstructed signal at the sampling point; represents the instantaneous amplitude of the reconstructed signal at the sampling point; The calculated TKEO energy operator is composed of a TKEO energy operator sequence, the TKEO energy operator sequence is segmented by time, each segment corresponds to half a power frequency period, and the maximum value of the TKEO derivative of each segment is calculated; The maximum ratio of TKEO derivatives of adjacent segments is calculated, and when the maximum ratio of TKEO derivatives of adjacent segments is greater than a set threshold value, early failure is determined, and early failure is graded and warned according to a preset threshold interval based on the ratio.

[0014] Preferably, the collection of the time sequence working condition class feature vector is specifically: The current change of the sheath current is calculated:

[0015] In the formula, Indicates the current sampling time; Indicates the power frequency period; Indicates the effective value of the first Road sheath current at time; Indicates the effective value of the first Road sheath current at time; Indicates the periodic variation of the first Road sheath current; The three-phase core current is decomposed into symmetrical positive, negative and zero sequence components by the symmetrical component method; The unbalance degree of the three-phase core current is calculated; The time sequence working condition class feature vector is composed of the normalized sheath current change, the positive, negative and zero sequence components of the three-phase core current, the unbalance degree of the three-phase core current, the effective value of the three-phase voltage, the sheath voltage to ground, the temperature, the humidity, the vibration amplitude, the partial discharge intensity and the power factor.

[0016] Preferably, the fault probability obtained by the Sigmoid function mapping is specifically:

[0017] In the formula, Indicates the original output value of the LSTM at time, reflecting the deviation degree of the current time from the historical normal mode, which is close to 0 in normal operation and significantly deviates in abnormal operation; Indicates the fault probability at time, with a value range of [0, 1]; Indicates normal operation, Indicates high fault probability; The dynamic updating of the fault probability threshold value is specifically:

[0018] Wherein, is the fault probability threshold value at time; Indicates the historical fault probability mean, For the corresponding weight factor; denotes the standard deviation of the historical failure probability, For the corresponding weight factor; For the corresponding weight factor.

[0019] Preferably, the method further comprises: based on the differential of the circulating current, the preliminary positioning of the fault section is carried out, the 1D-CNN is used to calculate the membership matrix of each section, the three criteria of the membership matrix, the correlation of adjacent sections and the historical failure probability are fused to determine the fault section, and the fault is limited in the single intelligent grounding box interval section; a comprehensive diagnostic report containing the fault type, the section where the fault is located, the severity and the development trend is generated, uploaded to the cloud platform and triggers an automatic work order.

[0020] Preferably, based on the differential of the circulating current, the preliminary positioning of the fault section is carried out, the 1D-CNN is used to calculate the membership matrix of each section, the three criteria of the membership matrix, the correlation of adjacent sections and the historical failure probability are fused to determine the fault section, and the fault is limited in the single intelligent grounding box interval section, specifically: based on the differential principle of the sheath circulating current, the fault section is preliminarily judged by comparing the current difference at both ends of each section; according to the geographical position sequence of the intelligent grounding box, the sheath current data from the line head end to the end of the intelligent grounding box is spliced into a one-dimensional sequence, and the 1D-CNN is input to obtain the membership matrix of each section; based on the membership matrix output by the 1D-CNN, the fault probability of each section is extracted as one of the criteria, the circulating current correlation mutation variable of each section is calculated based on the Pearson correlation coefficient as the second criterion, and the historical failure probability of each section is calculated based on the historical failure record as the third criterion, the fault possibility score of each section is calculated based on the weighted sum of the three criteria, the section with the highest score is selected as the final fault section, and the confidence of the section is calculated through the ratio of the highest score to the second highest score:

[0021] wherein, denotes the confidence of the final fault section; denotes the highest score in the sorting result of the section fault possibility score; denotes the second highest score in the sorting result of the section fault possibility score.

[0022] In another aspect, the present application provides a power cable grounding fault adaptive diagnosis system, which is used to implement the power cable grounding fault adaptive diagnosis method as described in any embodiment of the present application, and comprises a multi-modal data acquisition and preprocessing module, an early fault warning module, an adaptive threshold-free fault detection module, and a fault type intelligent identification module. The multi-modal data acquisition and preprocessing module is used to acquire multi-modal data including sheath current, three-phase core current, three-phase voltage, temperature and humidity, vibration, and partial discharge, and to preprocess and extract time-domain features, frequency-domain features, time-frequency domain features, and environmental and state features to form a statistical feature vector. The early fault warning module is used to adopt OVMD to adaptively decompose the sheath current data, calculate a segmented TKEO energy operator, and realize early warning of the fault through a maximum value ratio criterion of adjacent segment TKEO derivatives. The adaptive threshold-free fault detection module is used to acquire a time-series working condition feature vector as an input vector of the LSTM for training, input a real-time acquired time-series working condition feature vector into the trained LSTM, and map a fault probability through a Sigmoid function, and dynamically update a fault probability threshold based on the LSTM output, and determine that a fault occurs when the fault probability exceeds the fault probability threshold for a set period of time. The fault type intelligent identification module is used to input the statistical feature vector into an ANN and input the sheath current data into a 1D-CNN for fault classification respectively, and adopt adaptive weighted voting integrated decision for the outputs of the two types of networks to identify the fault type.

[0023] In still another aspect, the present application further provides an intelligent grounding box, which adopts a separate compartment structure, and comprises a primary grounding compartment and a secondary monitoring compartment, which are separated from each other; the primary grounding compartment comprises a grounding terminal system and an intelligent vacuum switch matrix; the secondary monitoring compartment comprises a multi-modal sensor array, an intelligent control unit, and a power cable grounding fault adaptive diagnosis system, which is the power cable grounding fault adaptive diagnosis system as described in claim 8. The grounding terminal system is provided with 6-9 grounding terminals corresponding to single-loop or double-loop cables, and a single-core cable is grounded through the grounding terminals, the grounding terminals are connected through cross-interconnected copper bars, and an intelligent vacuum switch matrix is configured. The multi-modal sensor array comprises a sheath current sensor, a high-frequency current sensor, a cable core current sensor, a voltage sensor, a temperature and humidity sensor, a vibration sensor, and a partial discharge sensor. The intelligent control unit comprises a main controller, a data acquisition module, a GPS / Beidou timing module, a data storage module, and a communication module.

[0024] In still another aspect, the application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the power cable grounding fault adaptive diagnosis method according to any one of the embodiments of the application.

[0025] Compared with the prior art, the application has the following technical effects: 1. Clear four-level diagnosis system: The progressive diagnosis process of "early warning → fault starting → type identification → section positioning" is adopted, and the functions of each link are clear and do not overlap; the early warning finds potential risks 2-48 hours in advance, the fault starting ensures millisecond-level rapid response, the type identification has an accuracy of 99.2%, and the fault is limited within a single intelligent grounding box section.

[0026] 2. Fault detection with double time scales: The early warning works on the hour-level time scale to detect weak fault precursor signals; the fault starting works on the millisecond-level time scale to monitor sudden faults in real time; the two complement each other to form a complete fault detection system, which can both prevent and respond to emergencies.

[0027] 3. Fault positioning capability: The intelligent section positioning algorithm based on 1D-CNN is adopted to fuse multiple criterion decisions and accurately lock the fault within the target section divided by the intelligent grounding box, which is more than 20 times higher than the traditional method and fully meets the operation and maintenance requirements.

[0028] 4. Advantage of multi-modal data fusion: Seven types of parameters such as current, voltage, temperature, humidity, vibration, and partial discharge are collected at the same time, and in particular, the high-frequency transient traveling wave current sensor can capture characteristic signals at the moment of fault, providing rich data support for accurate diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is the overall flowchart of the power cable grounding fault adaptive diagnosis method according to the application; Figure 2 is the overall structure diagram of the intelligent grounding box according to the application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the application clearer, the technical scheme of the application will be described clearly and completely below by combining the specific embodiments of the application and referring to the drawings.

[0031] Embodiment One This embodiment provides a power cable grounding fault adaptive diagnosis method, as shown in Figure 1 The method comprises the following steps: The multi-modal data including shield current, three-phase core current, three-phase voltage, temperature and humidity, vibration, and partial discharge are collected, and time domain features, frequency domain features, time-frequency domain features, and environmental and state features are extracted to form a statistical feature vector. Specifically, the shield current is the induced current and fault current flowing through the metal sheath or metal shielding layer of the high-voltage single-core cable metal sheath.

[0032] As a preferred embodiment of the present embodiment, under the triggering of GPS / Beidou timing signal, all multi-modal data are synchronously collected by sensors, and the preferred sampling frequency is 1 kHz for power frequency signal, 100 kHz for high frequency signal, and 1 Hz for environmental parameters. The pre-processing of the collected multi-modal data specifically includes: De-noising: wavelet threshold de-noising is adopted, db4 is used as the mother wavelet, 5 layers of decomposition are adopted, and soft threshold processing is adopted; filtering: notch filter (preferably 50 Hz) is used to remove power frequency interference; normalization: Min-Max normalization is used to normalize the multi-modal data to the interval [0, 1].

[0033] The pre-processed multi-modal data are extracted to form a statistical feature vector including time domain features, frequency domain features, time-frequency domain features, and environmental and state features. Due to the different sensitivities of different data to faults, after completing the extraction of each type of feature, the most representative features are selected through the feature selection and weighting fusion mechanism, and finally a fixed feature vector is formed. Specifically: The preliminary time domain features including effective value, peak value, peak factor, waveform factor, pulse factor, margin factor, skewness, and kurtosis are extracted from the multi-modal data, and according to the feature importance analysis and fault sensitivity sorting, 8 representative time domain features with the highest correlation to faults are selected from all preliminary time domain features and included in the time domain features of the feature vector.

[0034] Specifically, let the sampling sequence be , , and the number of sampling points is n, then the multi-modal data are calculated as follows: RMS (Root Mean Square): This parameter effectively reflects the energy level of the signal, and RMS will change significantly when a fault occurs.

[0035] Peak (Peak Value): This parameter reflects the maximum instantaneous assignment value of the signal.

[0036] CF (Peak Factor): Under normal circumstances, CF is about 1.414, showing the characteristics of a sine wave, and CF increases when a fault occurs, reflecting the trend of increased impact of the signal.

[0037] FF (Waveform Factor): This parameter describes the degree to which the waveform deviates from a sine wave.

[0038] Impulse Factor IF: , which reflects the sensitivity to the impulse signal, and IF increases significantly in early failure.

[0039] Margin Factor MF: , which is a failure-sensitive indicator, and the early failure detection effect is better than the peak factor.

[0040] Skewness: , , which is the standard deviation, and the parameter describes the symmetry of the distribution, which is close to 0 in normal times, and skewed in failure.

[0041] Kurtosis: , which describes the sharpness of the distribution, which is about 3 (Gaussian distribution) in normal times, and increases in failure.

[0042] The time domain signal of the multi-modal data is converted into a frequency domain signal by fast Fourier transform, and the frequency domain features including fundamental amplitude, harmonic content, and total harmonic distortion rate are extracted from the frequency domain signal, specifically: Fast Fourier transform (FFT) is performed on the signal, and the time domain signal is converted into a frequency domain signal : .

[0043] Fundamental amplitude : Extract the amplitude of the 50Hz component, i.e. , which reflects the strength of the power frequency signal.

[0044] Harmonic content: Extract the 3rd harmonic and the 5th harmonic , and the harmonic content increases in failure.

[0045] Total harmonic distortion rate THD: , where is the amplitude of the th harmonic, is the fundamental amplitude. THD<5% in normal times, and can reach more than 20% in failure.

[0046] The fundamental amplitude, 3rd harmonic, 5th harmonic, and THD of the power frequency are extracted from the spectrum, and after feature selection, a preset number (4 in this embodiment) of frequency domain features are retained as the final input frequency domain features.

[0047] The multi-modal data is decomposed by 5 layers using db4 wavelet, and the preset number of frequency bands with the largest energy proportion in the wavelet packet energy distribution are selected as the time-frequency domain features, specifically: The db4 wavelet is used for 5-layer decomposition, and 32 frequency bands are obtained. The energy proportion of each frequency band is calculated:

[0048] wherein, is the energy proportion of the i-th frequency band; is the wavelet coefficient of the i-th frequency band. A preset number of frequency bands with the largest energy proportion are selected as the features (in this embodiment, the preset number is preferably 3).

[0049] In addition, the Hilbert-Huang transform (HHT) instantaneous frequency can be calculated to assist in judging the feature mode (not included in the final frequency domain feature dimension), the signal is subjected to empirical mode decomposition (EMD) to obtain an intrinsic mode function (IMF), and then subjected to Hilbert transform to extract the instantaneous frequency mean :

[0050] wherein, is the instantaneous phase after the Hilbert transform.

[0051] The environmental and state feature vectors including the three-phase core current effective value and the sheath current effective value are extracted.

[0052] The OVMD is an adaptive signal decomposition method, which can decompose a complex signal into multiple intrinsic mode functions (IMF). Each IMF represents a specific frequency component of the signal.

[0053] As a preferred embodiment of this embodiment, the OVMD adaptive decomposition of the sheath current data, the calculation of the segmented TKEO energy operator, and the realization of the early fault warning through the maximum value ratio criterion of the adjacent segment TKEO derivative are specifically as follows: The determination of the decomposition layer number K is the key of OVMD. Too few layers will cause modal aliasing, and too many layers will produce false modes. The center frequency method is used in this embodiment to adaptively determine the optimal decomposition layer number. By analyzing the center frequency distribution of each mode under different layer numbers, when modes with similar frequencies appear, the optimal layer number is determined. The center frequency method is used to determine the decomposition layer number K, by analyzing the center frequency distribution of each mode under different layer numbers, when modes with similar frequencies appear, the optimal layer number is determined, and K is usually between 2 and 8.

[0054] ​​The OVMD adaptive decomposition coating current data is used to obtain a plurality of IMFs, and the pulse factor of each IMF and the cosine similarity of the original coating current data are calculated. The pulse factor reflects the impact characteristics of the signal, and the similarity reflects the correlation degree of the mode and the original signal. Two IMFs with the maximum pulse factor and the highest similarity are selected for reconstruction to obtain a reconstructed signal; The TKEO (Teager-Kaiser) energy operator is calculated for the reconstructed signal. TKEO is a method for estimating instantaneous energy, which is very sensitive to transient changes in the signal. The TKEO sequence is segmented by time, and each segment corresponds to half a power frequency period. The maximum value of the derivative of each segment TKEO is calculated, which reflects the energy change rate of the segment:

[0055] wherein, is the TKEO energy operator; is the sampling point index of the discrete time sequence, value is 1, 2, … , is the total number of sampling points of the coating current signal; represents the instantaneous amplitude of the reconstructed signal at the sampling point, and its unit is ampere; represents the instantaneous amplitude of the reconstructed signal at the sampling point (i.e. the previous sampling point); represents the instantaneous amplitude of the reconstructed signal at the sampling point (i.e. the next sampling point).

[0056] Through the TKEO operator, the difference between the square value of the current sampling point and the product of the adjacent two sampling points can effectively amplify the transient change of the signal, and has high sensitivity to the weak energy mutation caused by early faults. Compared with simple amplitude detection, the sensitivity of TKEO to energy change is increased by about 10-15 times.

[0057] The calculated TKEO energy operator is composed of a TKEO energy operator sequence, and the TKEO energy operator sequence is segmented by time, and each segment corresponds to half a power frequency period. The maximum value of the derivative of each segment TKEO is calculated; The ratio of the maximum values of the derivatives of the adjacent segments TKEO is calculated, and when the ratio of the maximum values of the derivatives of the adjacent segments TKEO is greater than a set threshold (such as > 2.0), it is determined as an early fault, and based on the ratio, the early fault is classified and warned according to a preset threshold interval. The warning level can be set based on the specific application scenario, such as setting the mild warning as 2.0 < R ≤ 5.0, the moderate warning as 5.0 < R ≤ 10.0, and the severe warning as R > 10.0.

[0058] This embodiment provides a calculation example as follows: Assume the sheath current values ​​at three consecutive sampling points are... It is 2.1A. It is 2.8A. If it is 2.3A, then: For a signal with a sampling frequency of 1kHz, each power frequency cycle (20ms) contains 20 sampling points, requiring the calculation of 20 TKEO values ​​to form a TKEO sequence: The TKEO sequence is divided into segments of half a period (10ms, corresponding to 10 sampling points), and the maximum value of the TKEO derivative for each segment is calculated. That is, the maximum value of the TKEO derivative for the k-th segment is calculated as follows: Calculation of the ratio between adjacent segments: ,in This indicates the maximum value of the TKEO derivative in the current segment. This represents the maximum value of the derivative of the preceding segment. During normal operation, the energy changes gradually between adjacent segments, with R ≈ 1.0-1.5. When an early fault occurs, the energy of the faulty segment increases significantly, causing R to suddenly increase. Therefore, when the ratio of the maximum TKEO derivative values ​​of adjacent segments is R > 2.0, it is considered a sign of an early fault.

[0059] As a preferred embodiment, the Weibull distribution can also be used to predict the fault development trend and estimate the remaining safe operating time, providing a basis for formulating maintenance plans. The Weibull distribution is a widely used lifetime distribution model in reliability engineering, particularly suitable for describing wear-type faults in electrical equipment. For early-stage cable faults, the development process follows an increasing failure rate, meaning the probability of failure increases over time. The cumulative failure probability function of the two-parameter Weibull distribution is:

[0060] Where t represents the time variable, and the unit is hours; This represents the cumulative failure probability up to time t, i.e., the probability of a failure occurring. This represents the scale parameter, also known as the characteristic lifetime. The shape parameter determines the trend of failure rate variation.

[0061] From the early warning trigger time Initially, continuously monitor the ratio of the maximum value of the TKEO derivative. And record its time series: ,in, Indicates the first The time for each sampling; Indicates the first The TKEO ratio of the second sample, when When R = 2.0, it is a light early warning, When R = 20.0, it represents a permanent failure threshold.

[0062] At the same time, the rate of change of the ratio, i.e. the failure development speed, is calculated:

[0063] The above Weibull parameter estimation method is to estimate the parameters by using Maximum Likelihood Estimation (MLE) And For the historical failure case database, the time from early warning triggering to permanent failure is recorded The likelihood function is constructed:

[0064] And the parameter estimation value is obtained by maximizing : Set the light early warning condition: 2 hours; the moderate early warning condition: 5 hours; the severe early warning condition: 10 hours; for most shielded ground faults, .

[0065] According to the above results, the remaining life prediction can be carried out, i.e. the current time After the early warning trigger Time, the measured , the remaining time to reach permanent failure (R = 20.0) is predicted , wherein the current failure probability is known:

[0066] The target is the permanent failure corresponding to the 90% failure probability:

[0067] Solved by Weibull distribution:

[0068] Wherein, The estimated result according to the current R(t) value is obtained by table lookup or empirical formula, which is as follows: R = 2.0 -> F ≈ 0.05, R = 5.0 -> F ≈ 0.20, R = 10.0 -> F ≈ 0.50, R = 15.0 -> F ≈ 0.75, R = 20.0 -> F ≈ 0.90. When F = 0.9, it is considered in engineering that power must be cut off at this time.

[0069] Considering the uncertainty of failure development, the development speed correction factor is introduced in this embodiment:

[0070] wherein, represents the historical average development speed, about 0.5-1.0 / hour; represents the current development speed, calculated in real time, if the current development speed is faster than the average value, the remaining time is shortened after correction.

[0071] By tracking the trend of the ratio, it can be judged whether the fault is deteriorating or recovering. Based on historical data and fault development model, the time of fault development into permanent fault can be predicted, providing basis for making maintenance plan.

[0072] Specifically, the embodiment respectively gives calculation cases of mild early warning and severe early warning for illustration, which are as follows: (1) Mild early warning: Triggering moment: R=2.3 (mild early warning). Current moment: 6 hours after early warning, R=3.1. Development speed: ≈0.13 / hour; ≈0.10, =0.9; =36 hours, =2.5; ≈35 hours. Correction: the development speed is slower, ≈40 hours. Prediction: there are about 40 hours of safe operation time, which can be handled in the next maintenance window.

[0073] (2) Severe early warning: Triggering moment: R=12.5 (severe early warning). Current moment: 2 hours after early warning, R=15.8. Development speed: ≈1.65 / hour; ≈0.77, =0.9; =6 hours, =2.5; ≈3 hours. Correction: the development speed is very fast, ≈2 hours. Prediction: only about 2 hours are left, which must be immediately arranged for power-off treatment.

[0074] Collect the time sequence working condition class feature vector as the input vector of LSTM for training, input the real-time collected time sequence working condition class feature vector into the trained LSTM and map the fault probability through the Sigmoid function, and update the fault probability threshold value based on the LSTM output, and determine that the fault occurs when the fault probability exceeds the fault probability threshold value for a set period. LSTM is especially suitable for processing time sequence data and can learn the time correlation characteristics of the cable operation state.

[0075] Furthermore, the relevant network parameters of the LSTM can be set according to the actual application scenario. In this embodiment, the LSTM preferably adopts a sliding time window mechanism, with each time window containing 100 time steps (corresponding to 0.1 seconds). Its network structure includes: Input layer: receiving time-series condition feature vectors, with an input dimension of (batch_size, time_steps, features), where time_steps=100, corresponding to a 0.1-second data window, and features=24; LSTM layer: two stacked LSTM layers, with 128 hidden units in the first layer and 64 hidden units in the second layer, using the tanh activation function, and the forget gate bias initialized to 1.0; Dropout layer: dropout rate of 0.3 to prevent overfitting; Fully connected layer: 64 neurons, ReLU activation; Output layer: outputting the fault probability; Training uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs; Dynamic threshold generation: based on the statistical characteristics of historical data, dynamically adjusted using the 3σ principle combined with LSTM prediction values. LSTM learns the data distribution characteristics under normal operating conditions to establish a baseline model of normal cable operation. The training process uses the backpropagation algorithm to optimize network parameters by minimizing prediction error.

[0076] In a preferred embodiment of this invention, the acquisition of the time-series operating condition feature vectors specifically involves: Calculate the sheath current abrupt change (6 dimensions). Under normal operation, the sheath current is stable. A; When a fault occurs, the sheath current changes drastically. The range of 50-200A is one of the main bases for LSTM to determine the occurrence of faults.

[0077] In the formula, Indicates the current sampling time; Indicates the power frequency period, Optimal 20ms; Indicates the first Road cover current The effective value at any given time; Indicates the first Road cover current The effective value at a given time (one period ago); Indicates the first The periodic variation in the current of the road cover layer reflects the degree of abrupt change in the circulating current.

[0078] The three-phase conductor current is decomposed into symmetrical positive-sequence, negative-sequence, and zero-sequence components (3-dimensional) using the symmetrical component method, specifically as follows: Positive order components: , negative sequence component: , zero sequence component: . Wherein, respectively represent the complex form (including amplitude and phase) of three-phase core current; represents the rotation factor; is the positive sequence component, representing the three-phase symmetrical load current; is the negative sequence component, representing the three-phase imbalance degree, and I2≈0 under normal conditions; is the zero sequence component, representing the ground fault current, and I0≈0 under normal conditions. Under normal operation, each sequence component is kept , constraint.

[0079] When a fault occurs, each sequence component has the following situation: single-phase ground fault, significantly increased ( ), moderate increase; interphase short circuit fault, significantly increased ( ), small.

[0080] Calculate the unbalance degree (1 dimension) of three-phase core current , which is defined as follows:

[0081] Wherein, respectively represent the effective value of three-phase core current; represents the average value of three-phase core current. Under normal operation %, single-phase fault about 30%-80%. This feature, together with the above three-phase sequence components, realizes the improvement of fault judgment accuracy.

[0082] The sheath current mutation variable, the positive sequence, negative sequence and zero sequence components of the three-phase core current, the unbalance degree of the three-phase core current, the effective value of the three-phase voltage (3 dimensions), the sheath voltage to ground (6 dimensions), temperature (1 dimension), humidity (1 dimension), vibration amplitude (1 dimension), partial discharge intensity (1 dimension) and power factor (1 dimension) are normalized to form a 24-dimensional time sequence working condition characteristic vector, which is arranged in time sequence to form an input matrix of (100, 24) input LSTM.

[0083] As a preferred embodiment of the present embodiment, the LSTM outputs a real value , and the fault probability is obtained by mapping through the Sigmoid function, which is specifically:

[0084] In the formula, represents the LSTM in The raw output value at any given time reflects the degree of deviation between the current time and the historical normal pattern. When normal, it is close to 0, and when abnormal, it deviates significantly. express The probability of failure at time t, with a value range of [0,1]. This indicates that the system is operating normally. Indicates a high probability of failure; The fault probability threshold is dynamically updated so that it can be adaptively adjusted according to the actual operating status of the cable. Specifically:

[0085] in, for The fault probability threshold at any given time; This represents the average historical failure probability. for The corresponding weighting factor; This represents the standard deviation of historical failure probabilities. for The corresponding weighting factor indicates that it is based on The coefficient of the principle, according to the normal distribution theory, It can cover 99.7% of normal data. Therefore, setting k=3 can control the false alarm rate below 0.3%. for The corresponding weighting factor allows the threshold to be fine-tuned based on the real-time output of the LSTM. When the LSTM detects an abnormal trend, As the threshold increases, the sensitivity is improved. Based on the test results of the examples, the preferred value is 0.2. Partly based on statistical principles, it adapts to the average level and fluctuation characteristics of the system. Partly based on deep learning, it can capture complex nonlinear trends. The combination of the two parts provides both the robustness of statistics and the sensitivity of deep learning.

[0086] Furthermore, This represents the historical average failure probability and reflects the baseline level of normal system operation.

[0087] in, Based on data from the past hour, a sampling frequency of 1Hz is preferred. =3600.

[0088] The standard deviation of historical failure probabilities reflects the degree of data fluctuation, especially in systems with large load variations. The value is relatively large, requiring a higher threshold to avoid false alarms.

[0089] wherein, is the failure probability at time t.

[0090] In addition, the failure probability threshold is also constrained to prevent abnormal fluctuations in the threshold value and limit the threshold value range:

[0091] And the threshold value update period is set to update every minute, balancing its calculation efficiency and adaptive ability.

[0092] At the same time, when the following conditions are met, it is determined that a failure has occurred and the diagnosis process is started: When That is, the failure probability exceeds the dynamic threshold and remains in this state for a continuous number of set values (preferably 3) power frequency periods (60 ms). The failure probability determination uses a multi-cycle confirmation mechanism. When the failure probability exceeds the dynamic threshold for a continuous number of set values, it is determined that a failure has occurred and the subsequent diagnosis process is immediately started. This mechanism effectively avoids false positives caused by transient interference, improving the reliability of failure detection. Compared with the fixed threshold method, the false positive rate is reduced from 30% to 4.5%, and the false negative rate is reduced from 15% to 1%.

[0093] Immediately after triggering, the following instructions are executed: switch to a high-speed sampling mode with a frequency of 1 kHz; record the waveforms before and after the failure for 100 ms; start fault type identification; start fault section positioning.

[0094] The statistical feature vector is input into the ANN, and the shield current data is input into the 1D-CNN for fault classification, and the outputs of the two types of networks are integrated by adaptive weighted voting decision to identify the fault type, including: Type 1: single-phase ground fault of the core (distinguishing between metallic ground and high-impedance ground); Type 2: core interphase short circuit fault; Type 3: single-point ground fault of the shield; Type 4: multi-point ground fault of the shield; Type 5: early fault of insulation deterioration. This step uses two different architectures of neural networks: artificial neural network (ANN) and one-dimensional convolutional neural network (1D-CNN). ANN is good at processing structured feature data, while 1D-CNN is good at automatically extracting features from raw waveform data. By combining the advantages of the two networks, more accurate and robust fault identification can be achieved.

[0095] Specifically, the dimension of the statistical feature vector input to the ANN can be limited according to the actual application scenario requirements. In this embodiment, a 24-dimensional statistical feature vector is preferred, and the four types of features include: time domain features (8 dimensions): including effective value, peak value, peak factor, waveform factor, pulse factor, margin factor, skewness, and kurtosis; frequency domain features (4 dimensions): 4 frequency domain features are reserved after feature selection from the fundamental amplitude, 3rd harmonic, 5th harmonic, and total harmonic distortion (THD) of the frequency spectrum; time-frequency domain features (3 dimensions): the 3 frequency bands with the highest energy proportion are selected from the wavelet packet energy distribution; environmental and state features (9 dimensions): including three-phase core current effective value (3 dimensions) and six-way shield current effective value (6 dimensions).

[0096] Further, the network structure of the ANN can be set according to the actual application scenario. In this embodiment, the network structure of the ANN adopts a fully connected feedforward neural network structure, including: an input layer of 24 neurons; a hidden layer: 3 fully connected layers with neuron numbers of 256, 128, and 64 respectively, using ReLU activation function, and the number of neurons decreases layer by layer, forming a structure of gradually abstracted features; an output layer: 5 neurons corresponding to 5 fault types, using Softmax activation and cross-entropy loss function.

[0097] Similarly, the network structure of the 1D-CNN can be set according to the actual application scenario. In this embodiment, the network structure of the 1D-CNN is a combination of multiple convolutional layers and pooling layers, including: an input layer: receiving the original waveform data of the shield current, the data source is the time series signal collected by the shield current sensor (Rogowski coil sensor), the sampling window is 100 ms, the sampling frequency is 1 kHz, forming a one-dimensional sequence of 100 sampling points; a convolutional layer: including 3 convolutional layers with filter numbers of 32, 64, and 128 respectively, convolution kernel size of 3, and step of 1, realizing automatic extraction of local features of the waveform; a pooling layer: 2 maximum pooling layers with pooling size of 2, reducing the data dimension and extracting main features; a fully connected layer: after global average pooling, 2 fully connected layers are connected, with 128 and 5 neurons respectively; an output layer: 5 neurons outputting the probability of each fault type. Different sizes of convolution kernels are used in the convolutional layer to capture features of different time scales. The pooling layer is used to reduce the data dimension and extract main features. After convolution and pooling operations, the extracted features are mapped to the fault type through the fully connected layer.

[0098] The outputs of the two types of networks are integrated by adaptive weighted voting, and the calculation formula for identifying the fault type is:

[0099] wherein, is the identified fault type; and are the output probabilities of the two types of networks respectively. and are the corresponding weights of the two types of networks, and The weights are adaptively adjusted by the validation set performance, and the network with better performance obtains a larger weight. The final fault type is determined by the class with the largest probability after fusion.

[0100] This embodiment realizes complementary advantages by processing different forms of data through two networks. Specifically, the ANN branch processes the extracted statistical features to capture global characteristics, and the 1D-CNN branch directly processes the original waveform to automatically learn local time sequence features. Integrated learning improves the fault recognition accuracy from 96% of a single network to 99.2%.

[0101] As a preferred embodiment of the present embodiment, the method further comprises: Based on the differential current loop, the fault section is preliminarily located, the 1D-CNN is used to calculate the membership matrix of each section, the membership matrix, the correlation of adjacent sections and the historical fault probability are fused to determine the fault section, and the fault is limited in a single intelligent grounding box section. The concept of section refers to the cable section divided by two adjacent intelligent grounding boxes. For a typical 110kV transmission line, the length of each loop is 15-30km, and 5-8 intelligent grounding boxes are usually installed along the line to divide the line into 4-7 sections, each with a length of about 3-5km. In other words: for a complete 110kV cable line, W intelligent grounding boxes are installed along the line, where W is usually 4-10 and depends on the length of the line, which naturally divides the line into W-1 cable sections. Each section is composed of the cable section between two adjacent intelligent grounding boxes, and the section number is 1 to W-1. The following is given as an example: for example, for a line with 5 intelligent grounding boxes, 4 sections are formed: section 1 (between grounding box 1 and grounding box 2), section 2 (between grounding box 2 and grounding box 3), section 3 (between grounding box 3 and grounding box 4), and section 4 (between grounding box 4 and grounding box 5).

[0102] A comprehensive diagnostic report containing fault type, section, severity, and fault development trend is generated, uploaded to the cloud platform and triggered to automatically generate a work order, and alarm information is pushed through SMS / APP, such as recommending immediate power outage for severe faults, scheduling planned maintenance for minor faults, and strengthening monitoring and preparing for emergencies for early faults.

[0103] For long-distance cable lines, quickly locating the fault section is crucial for quick repair. As a preferred embodiment of the present embodiment, based on the differential current loop, the fault section is preliminarily located, the 1D-CNN is used to calculate the membership matrix of each section, the membership matrix, the correlation of adjacent sections and the historical fault probability are fused to determine the fault section, and the fault is limited in a single intelligent grounding box section, specifically: Firstly, the rough positioning is performed, and the possible fault section is quickly determined based on the differential principle of the sheath current. When a section fails, the sheath currents at both ends of the section will have a significant difference. By comparing the current difference values at both ends of each section, the fault section can be preliminarily judged.

[0104] According to the geographical position sequence of the intelligent grounding boxes, the sheath current data from the line head end to the end of the intelligent grounding boxes is spliced into a one-dimensional sequence, and a membership matrix of each section is obtained by inputting the one-dimensional sequence into the 1D-CNN. Specifically, each intelligent grounding box regularly monitors: each grounding box collects the sheath current effective value at a frequency of 1 Hz, which is used for early warning and trend analysis. When a fault occurs, the high-speed sampling mode with a frequency of 1 kHz is automatically switched, and the waveform data 100 ms before and 100 ms after the fault are recorded. Under the synchronization of the GPS second pulse signal, the 0.2-second waveform data of 200 sampling points before and after the fault are uploaded after being time-stamped. The GPS / Beidou time service (accuracy ≤ 30 ns) ensures the time synchronization of the data of each intelligent grounding box, and the synchronization error is less than 1 . The sheath current waveforms collected by the N grounding boxes corresponding to the N-1 sections are aligned in time and spliced in turn to form a one-dimensional sequence with a length of N×200. Specifically, let the sheath current waveform collected by the i th grounding box be a vector , then the input sequence is: , and the total length of the sequence is N×200 dimensions.

[0105] Further, the network structure of the 1D-CNN can be set according to the actual application scenario. In this embodiment, the network structure of the 1D-CNN preferably includes: the first convolutional layer uses 32 filters, the kernel size is 5, the step size is 1, and the ReLU activation is used; the first pooling layer uses maximum pooling with a pooling size of 2; the second convolutional layer uses 64 filters, the kernel size is 3, the step size is 1, and the ReLU activation is used; the second pooling layer uses maximum pooling with a pooling size of 2; the flattening layer; the fully connected layer uses 128 neurons, the ReLU activation is used, and the Dropout (0.5) is used; and the output layer is an N×2-dimensional membership matrix U, which represents the fault / non-fault probability of each section, wherein U [i,0] represents the membership (probability) of the i th section belonging to the fault class, U [i,1] represents the membership belonging to the non-fault class, and U [i,0] + U [i,1]= 1. The waveform of each section contains data before and after the fault, which can reflect the dynamic characteristics of the fault. The convolutional layer extracts local features through sliding windows, the pooling layer reduces the data dimension, and finally the fault probability of each section is output through the fully connected layer. Each element of the membership matrix represents the probability of the corresponding section belonging to the fault class or non-fault class. By analyzing the membership matrix, not only the most likely fault section can be determined, but also the confidence evaluation can be given. When the fault probabilities of multiple sections are close, the system will give multiple possible fault locations for reference.

[0106] The membership matrix based on the 1D-CNN output extracts the fault probability of each section as one of the criteria, the current correlation mutation variable of each section based on the Pearson correlation coefficient is calculated as the second criterion, and the historical fault probability of each section based on the historical fault record is calculated as the third criterion. For each section, the weighted sum of the three criteria is calculated to obtain the fault possibility score, the section with the highest score is selected as the final fault section, and the confidence of the section is calculated by the ratio of the highest score to the second highest score.

[0107] Specifically, the current correlation mutation variable of each section is calculated based on the Pearson correlation coefficient: The fault section will have an abnormal mutation in the sheath current, which will cause a mutation in the current correlation between the section and the adjacent section. The current correlation between normal sections is high (correlation coefficient > 0.9), while the correlation before and after the fault section will be broken. The Pearson correlation coefficient between adjacent intelligent grounding boxes is calculated:

[0108] wherein, represents the Pearson correlation coefficient function, represents the covariance, represents the standard deviation of the th waveform, represents the correlation between the grounding boxes and , corresponding to the section . In normal times, > 0.9 (highly correlated); when a fault occurs, the Pearson correlation coefficient of the fault section will decrease significantly and be <0.6.

[0109] For section (bounded by grounding boxes and ), the correlation mutation variable is: . Finally, the score value is normalized: . The higher the value, the more likely the area is a fault section.

[0110] Specifically, the historical fault probability of each section is calculated based on the historical fault record: The number of faults in each section, the number of ground box actions, the abnormality record of the shielded layer circulating current, the cable partial discharge alarm, the operation and maintenance inspection report, and the GIS / TTS system count are recorded to construct a historical fault database in the past 12 months. The criterion indicates that some sections have a higher probability of failure due to harsh environment (high temperature, high humidity, and large mechanical stress), manufacturing defects, or long running time. The historical fault data provide statistical information on the vulnerability of the section, and the calculation method is: wherein, represents the number of faults in the th section in the past 12 months, represents the total number of faults in all sections.

[0111] The failure probability of each section is calculated by weighted summation based on three criteria: wherein, represents the failure probability; represents the correlation score of adjacent sections (based on the correlation coefficient of the shielded layer current); represents the historical failure probability (based on the operation and maintenance database statistics); , , are the weight factors corresponding to the three criteria, preferably set to 0.5, 0.3, and 0.2.

[0112] The final fault section and its confidence are output:

[0113] wherein, represents the confidence of the final fault section; represents the highest score in the sorted results of the section failure probability score; represents the second highest score in the sorted results of the section failure probability score.

[0114] The entire diagnostic process described in this embodiment is fully automatic, and only takes seconds from fault occurrence to output of the diagnostic result, greatly improving the fault handling efficiency. At the same time, this embodiment has self-learning ability, and can continuously optimize the diagnostic model according to the actual running data to improve the diagnostic accuracy.

[0115] Embodiment Two Correspondingly, the embodiment provides a self-adaptive diagnosis system for grounding faults of power transmission cables, which is used to realize the method as described in Embodiment One, and includes a multi-modal data acquisition and preprocessing module, an early fault warning module, a self-adaptive threshold-free fault detection module, and a fault type intelligent identification module.

[0116] The multimodal data acquisition and preprocessing module is used to acquire multimodal data including sheath current, three-phase conductor current, three-phase voltage, temperature and humidity, vibration, and partial discharge, and to preprocess the data to extract time-domain features, frequency-domain features, time-frequency-domain features, and environmental and state features to form statistical feature vectors.

[0117] The early fault warning module is used to adaptively decompose sheath current data using OVMD, calculate segmented TKEO energy operators, and achieve early fault warning by judging the ratio of the maximum value of the TKEO derivative of adjacent segments.

[0118] The adaptive threshold-free fault detection module is used to collect time-series operating condition feature vectors as input vectors for training LSTM. The real-time collected time-series operating condition feature vectors are input into the trained LSTM and mapped by the Sigmoid function to obtain the fault probability. The fault probability threshold is dynamically updated based on the LSTM output. When the fault probability exceeds the fault probability threshold for a continuous set period, a fault is determined to have occurred.

[0119] The fault type intelligent identification module is used to input statistical feature vectors into ANN and sheath current data into 1D-CNN for fault classification, and to use adaptive weighted voting integration decision on the outputs of the two networks to identify the fault type.

[0120] Example 3 This embodiment provides an intelligent grounding box, such as... Figure 2 As shown, the intelligent grounding box adopts a compartmentalized structure, including a primary grounding compartment and a secondary monitoring compartment. The two compartments are separated by an epoxy resin insulation board with a thickness of not less than 12mm and an insulation resistance of not less than 1000MΩ. The box body is made of stainless steel, with a protection level of IP66, and can operate normally under ambient temperatures of -40℃ to +70℃ and relative humidity of 5% to 95%. The primary grounding compartment includes a grounding terminal system and an intelligent vacuum switch matrix, while the secondary monitoring compartment includes a multi-modal sensor array, an intelligent control unit, and a software system. The grounding box structure can be constructed according to actual application scenarios. The following is a preferred grounding box structure in this embodiment.

[0121] The grounding terminal system is equipped with 6-9 grounding terminals, corresponding to single-circuit or dual-circuit cables. Single-core cables are grounded through the grounding terminals, and the grounding terminals are connected by cross-interconnected copper busbars. It is also equipped with an intelligent vacuum switch matrix with a switching capacity of AC1500V / 500A, an action time of less than 20ms, and supports fast switching of various grounding topologies such as direct grounding, protective grounding, and cross-interconnection.

[0122] The multi-modal sensor array comprises: a sheath current sensor, 6-9 open type Rogowski coil sensors, accuracy level 0.2, measurement range 0-500A, frequency band range 50Hz-100kHz, linearity better than 0.5%, temperature coefficient less than 50ppm / ℃; a high-frequency current sensor, frequency response range 10kHz-10MHz, sensitivity 300mV / A, used for capturing high-frequency transient high-frequency traveling wave current signals at the moment of failure, rise time less than 100ns, the high-frequency transient high-frequency traveling wave current refers to a high-frequency transient current component generated in the cable at the moment of failure, frequency range 10kHz-10MHz, used for fault feature identification; cable core current sensors / through-type current transformers 3, variable ratio 2000:5 or 3000:5, accuracy level 0.5S, rated load 5VA; voltage sensors / resistance-capacitance voltage dividers 6, measurement range 0-80kV, accuracy level 0.5, frequency response DC-10kHz, input impedance greater than 100M; temperature and humidity sensors use sensor chips 2, temperature measurement range -40℃~+125℃, accuracy ±0.2℃, humidity measurement range 0-100%RH, accuracy ±1.5%RH; three-axis vibration sensor: MEMS acceleration sensor 1, range ±16g, frequency response 0.5Hz~3kHz, resolution 0.1mg, used for detecting mechanical vibration of the cable and grounding system; partial discharge sensor: ultra-high frequency (UHF) sensor 2, detection band 300MHz~1.5GHz, sensitivity -80dBm, used for detecting partial discharge signals of the cable insulation.

[0123] The intelligent control unit comprises: a main controller: four-core ARM Cortex-A53, main frequency 1.8GHz, integrated NPU neural network acceleration unit, computing power up to 2.3TOPS; a data acquisition module: 16-bit synchronous sampling ADC, sampling rate 100kSPS / channel, 16-channel parallel acquisition, signal-to-noise ratio ≥90dB; a GPS / Beidou time service module: supporting GPS L1 / L2 and Beidou B1 / B2 dual frequency, positioning accuracy ≤2.5 m, time service accuracy ≤30ns, providing 1PPS second pulse output; a data storage module: 32GB eMMC local storage, 128GB SD card expansion, supporting circular storage and event triggered storage; a communication module: supporting 4G / 5G full network, downlink rate ≥100Mbps, gigabit Ethernet, RS485 with isolation, optical fiber and other communication modes; a power management module: wide voltage input AC85V~265V, output power 50W, configured with super capacitor as backup power supply, which can continuously work for 15 minutes.

[0124] The software system adopts a layered design, including: a data acquisition layer: real-time data acquisition, time synchronization, data caching; a preprocessing layer: digital filtering, feature extraction, data compression; an algorithm layer: a power cable grounding fault adaptive diagnosis system, the system being the system of the embodiment two of the application; an application layer: a man-machine interface, data visualization, report generation, remote maintenance; a communication layer: data transmission, protocol conversion, encryption authentication.

[0125] Embodiment four The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method in the embodiment one of the application.

[0126] In the embodiments of the present application, "at least one" refers to one or more, and "multiple" refers to two or more. "And / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the cases of A alone, A and B together, and B alone. Wherein A and B can be singular or plural. The character " / " generally represents that the associated objects before and after it are in an "or" relationship. "At least one of the following" and the like means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, wherein a, b, c can be single or multiple.

[0127] Those skilled in the art can realize that the units and algorithm steps described in the embodiments disclosed in the present application can be realized by electronic hardware, computer software and combination of electronic hardware and computer software. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0128] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0129] In several embodiments provided in the present application, any function, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0130] The above description is only some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, based on the content of the specification and drawings of the present application, are also included in the patent protection scope of the present application.

Claims

1. An adaptive diagnostic method for grounding faults in power transmission cables, characterized in that, Includes the following steps: Multimodal data including sheath current, three-phase conductor current, three-phase voltage, temperature and humidity, vibration, and partial discharge are collected and preprocessed to extract time-domain features, frequency-domain features, time-frequency-domain features, and environmental and state features to form statistical feature vectors. OVMD adaptive decomposition of sheath current data is used to calculate segmented TKEO energy operators, and early fault warning is achieved by judging the ratio of the maximum value of TKEO derivatives of adjacent segments. The feature vectors of the time-series operating conditions are collected and used as the input vectors of the LSTM for training. The real-time collected feature vectors of the time-series operating conditions are input into the trained LSTM and the fault probability is obtained by mapping through the Sigmoid function. The fault probability threshold is dynamically updated based on the LSTM output. When the fault probability exceeds the fault probability threshold for a continuous set period, the fault is determined to have occurred. The statistical feature vectors are input into the ANN and the sheath current data are input into the 1D-CNN for fault classification. The outputs of the two networks are then integrated using an adaptive weighted voting method to identify the fault type.

2. The adaptive diagnosis method for grounding faults in power transmission cables according to claim 1, characterized in that, Triggered by GPS / BeiDou timing signals, all multimodal data are synchronously collected through sensors. Preprocessing of the collected multimodal data specifically includes: Denoising: Wavelet thresholding is used for denoising, with the mother wavelet using db4 and a decomposition layer of 5. Soft thresholding is used for processing. Filtering: Notch filter is used to remove power frequency interference. Normalization: Min-Max normalization is used to normalize the multimodal data to the [0,1] interval. The statistical feature vector is composed of time-domain features, frequency-domain features, time-frequency-domain features, and environmental and state features extracted from the preprocessed multimodal data. Preliminary time-domain features, including RMS value, peak value, peak factor, waveform factor, impulse factor, margin factor, skewness, and kurtosis, were extracted from the multimodal data. Based on feature importance analysis and fault sensitivity ranking, eight types of time-domain features were selected from all the preliminary time-domain features and included in the feature vector. The time-domain signal of the multimodal data is converted into a frequency-domain signal through a fast Fourier transform, and frequency-domain features including fundamental amplitude, harmonic content, and total harmonic distortion are extracted from the frequency-domain signal. The multimodal data is decomposed into 5 levels using db4 wavelets, and a preset number of frequency bands with the largest energy proportion in the wavelet packet energy distribution are selected as time-frequency domain features. Extract environmental and state feature vectors, including the effective values ​​of the three-phase conductor current and the effective values ​​of the sheath current.

3. The adaptive diagnosis method for grounding faults in power transmission cables according to claim 1, characterized in that, The method of using OVMD adaptive decomposition of sheath current data, calculating segmented TKEO energy operators, and achieving early fault warning through the ratio of the maximum values ​​of TKEO derivatives of adjacent segments specifically involves: OVMD adaptive decomposition of sheath current data is used to obtain multiple IMFs. The pulse factor of each IMF and the cosine similarity of the original sheath current data are calculated. The two IMFs with the largest pulse factor and the highest similarity are selected for reconstruction to obtain the reconstructed signal. Calculate the TKEO energy operator for the reconstructed signal: in, For TKEO energy operators; For the sampling point index of the discrete time series, Values ​​are 1, 2, ... , This represents the total number of sampling points for the protective layer current signal. Indicates the reconstructed signal at the th The instantaneous amplitude at each sampling point, measured in amperes; Indicates the reconstructed signal at the th The instantaneous amplitude of each sampling point; Indicates the reconstructed signal at the th The instantaneous amplitude of each sampling point; The calculated TKEO energy operators are arranged into a TKEO energy operator sequence. The TKEO energy operator sequence is divided into time segments, each segment corresponding to half a power frequency cycle. The maximum value of the TKEO derivative of each segment is calculated. The ratio of the maximum TKEO derivative values ​​of adjacent segments is calculated. When the ratio of the maximum TKEO derivative values ​​of adjacent segments is greater than a set threshold, it is determined to be an early fault. Based on the ratio, a graded early warning is implemented for the early fault according to a preset threshold range.

4. The adaptive diagnosis method for grounding faults in transmission cables according to claim 1, characterized in that, The acquisition of the time-series operating condition feature vectors is specifically as follows: Calculate the sheath current mutation: In the formula, Indicates the current sampling time; Indicates the power frequency period; Indicates the first Road cover current The effective value at any given time; Indicates the first Road cover current The effective value at any given time; Indicates the first The periodic variation of the current in the road cover layer; The three-phase conductor current is decomposed into symmetrical positive-sequence, negative-sequence, and zero-sequence components using the symmetrical component method. Calculate the unbalance of the three-phase conductor current; The time-series operating condition feature vector is composed of normalized parameters such as sheath current mutation, positive sequence, negative sequence and zero sequence components of three-phase conductor current, unbalance of three-phase conductor current, effective value of three-phase voltage, sheath voltage to ground, temperature, humidity, vibration amplitude, partial discharge intensity and power factor.

5. The adaptive diagnostic method for grounding faults in power transmission cables according to claim 1, characterized in that, The fault probability is obtained by mapping using the Sigmoid function as follows: In the formula, Indicates that LSTM is in The raw output value at any given time reflects the degree of deviation between the current time and the historical normal pattern. When normal, it is close to 0, and when abnormal, it deviates significantly. express The probability of failure at time t, with a value range of [0,1]. This indicates that the system is operating normally. Indicates a high probability of failure; The specific steps for dynamically updating the fault probability threshold are as follows: in, for The fault probability threshold at any given time; This represents the average historical failure probability. for The corresponding weighting factor; This represents the standard deviation of historical failure probabilities. for The corresponding weighting factor; for The corresponding weighting factor.

6. The adaptive diagnosis method for grounding faults in transmission cables according to claim 1, characterized in that, The method further includes: Preliminary fault location is based on circulating differential. 1D-CNN is used to calculate the membership matrix of each segment. The three criteria of membership matrix, correlation between adjacent segments and historical fault probability are integrated to determine the fault segment and limit the fault to a single smart grounding box segment. Generate a comprehensive diagnostic report that includes fault type, location, severity, and development trend, upload it to the cloud platform, and trigger an automatic work order.

7. The adaptive diagnosis method for grounding faults in power transmission cables according to claim 6, characterized in that, Preliminary fault location is performed based on circulating differential current. A 1D-CNN is used to calculate the membership matrix of each segment. The fault segment is determined by fusing the membership matrix, the correlation between adjacent segments, and the historical fault probability as three criteria, thus confining the fault to a single smart grounding box segment. Based on the differential principle of sheath circulation current, the fault section can be preliminarily determined by comparing the current difference between the two ends of each section. According to the geographical location of the smart grounding box, the sheath current data from the beginning to the end of the line of the smart grounding box are spliced ​​into a one-dimensional sequence, and input into 1D-CNN to obtain the membership matrix of each section. The failure probability of each segment is extracted based on the membership matrix output by 1D-CNN as one criterion. The abrupt change in circulation correlation of each segment is calculated based on the Pearson correlation coefficient as a second criterion. The historical failure probability of each segment is calculated based on historical failure records as a third criterion. The failure probability score of each segment is calculated by weighted summation based on the three criteria. The segment with the highest score is selected as the final failure segment, and the confidence level of the segment is calculated by the ratio of the highest score to the second highest score. in, Indicates the confidence level of the final faulty section; This represents the highest score in the ranking results of the section fault probability scores; This indicates the second highest score in the ranking results of the section fault probability scores.

8. An adaptive diagnostic system for grounding faults in power transmission cables, characterized in that, The system is used to implement the method as described in any one of claims 1 to 7, and includes a multimodal data acquisition and preprocessing module, an early fault warning module, an adaptive threshold-free fault detection module, and a fault type intelligent identification module. The multimodal data acquisition and preprocessing module is used to acquire multimodal data including sheath current, three-phase conductor current, three-phase voltage, temperature and humidity, vibration, and partial discharge, and to preprocess the data to extract time-domain features, frequency-domain features, time-frequency-domain features, and environmental and state features to form statistical feature vectors. The early fault warning module is used to adaptively decompose the sheath current data using OVMD, calculate the segmented TKEO energy operator, and realize early fault warning by judging the ratio of the maximum value of the TKEO derivative of adjacent segments. The adaptive threshold-free fault detection module is used to collect time-series operating condition feature vectors as input vectors for training LSTM. The real-time collected time-series operating condition feature vectors are input into the trained LSTM and mapped by the Sigmoid function to obtain the fault probability. The fault probability threshold is dynamically updated based on the LSTM output. When the fault probability exceeds the fault probability threshold for a continuous set period, a fault is determined to have occurred. The fault type intelligent identification module is used to input statistical feature vectors into ANN and sheath current data into 1D-CNN for fault classification, and to use adaptive weighted voting integration decision on the outputs of the two networks to identify the fault type.

9. An intelligent grounding box, characterized in that, The intelligent grounding box adopts a compartmentalized structure, including a primary grounding compartment and a secondary monitoring compartment, which are separated from each other. The primary grounding compartment includes a grounding terminal system and an intelligent vacuum switch matrix. The secondary monitoring compartment includes a multi-mode sensor array, an intelligent control unit, and an adaptive diagnostic system for power transmission cable grounding faults, wherein the system is the adaptive diagnostic system for power transmission cable grounding faults as described in claim 8. The grounding terminal system is equipped with 6-9 grounding terminals, corresponding to single-circuit or double-circuit cables. Single-core cables are grounded through the grounding terminals, and the grounding terminals are connected by cross-interconnected copper busbars, and an intelligent vacuum switch matrix is ​​configured. The multimodal sensor array includes: a sheath current sensor, a high-frequency current sensor, a cable core current sensor, a voltage sensor, a temperature and humidity sensor, a vibration sensor, and a partial discharge sensor. The intelligent control unit includes: a main controller, a data acquisition module, a GPS / BeiDou timing module, a data storage module, and a communication module.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the adaptive diagnosis method for grounding faults of power transmission cables as described in any one of claims 1 to 7.

Images