High and low voltage switch cabinet insulation monitoring method and system

Abstract

The invention discloses an insulation monitoring method and system for a high-low voltage switch cabinet, and belongs to the technical field of electrical monitoring and insulation diagnosis, and the method comprises the steps of continuous monitoring, trigger type active measurement, feature extraction, intelligent diagnosis, uncertainty adaptive measurement and the like. According to the invention, the graph neural network and the deep learning model are adopted, and deep fusion and intelligent diagnosis can be carried out on multi-dimensional parameters. By comparing the confidence coefficient with a threshold value, the system can dynamically adjust measurement parameters and update the model by using a final judgment result to realize closed-loop adaptive optimization. According to the method, a high-confidence diagnosis conclusion can be obtained, accurate fault positioning is carried out, and the reliability and efficiency of insulation monitoring of the high-low voltage switch cabinet are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electrical monitoring and insulation diagnosis, in particular to a high-low voltage switch cabinet insulation monitoring method and system. BACKGROUND

[0002] The high-low voltage switch cabinet is the core equipment of power distribution, control and protection in the power system, and the health status of its internal insulation structure is directly related to the safe and stable operation of the entire power grid. Insulation monitoring is a key technology to ensure the reliable operation of the switch cabinet, which aims to detect potential defects such as insulation aging, dampness, contamination and partial discharge in advance through real-time or periodic detection of insulation performance related parameters, and prevent sudden insulation breakdown accidents. In the prior art, the insulation monitoring methods for high-low voltage switch cabinets mainly include offline insulation resistance test, dielectric loss angle measurement, and online leakage current monitoring, partial discharge detection, etc. In terms of online monitoring, high-frequency current sensors or ultrasonic sensors are usually used to capture the pulse signals generated by partial discharge, or the insulation state is judged by measuring the leakage current flowing through the insulation of the equipment. Some methods will combine temperature, humidity and other environmental parameters for compensation, or use a single threshold alarm logic for fault judgment.

[0003] However, the above prior art solutions have some inherent technical defects. Methods that simply rely on steady-state parameters such as leakage current are not sensitive enough to early and weak insulation degradation phenomena, and often cannot be detected until the fault has developed to a serious level. The monitoring method that relies only on partial discharge pulse signals is easily disturbed by the complex electromagnetic environment on site, resulting in low signal-to-noise ratio, false alarms or missed alarms. In addition, most of the existing technologies independently analyze or simply superimpose the collected different types of data, and lack effective means for deep fusion and correlation analysis of multi-source information, so it is difficult to achieve accurate differentiation of fault types and accurate positioning of fault locations. SUMMARY

[0004] To solve the above problems, the present application provides a high-low voltage switch cabinet insulation monitoring method and system, which adopts a combination of passive monitoring and active detection, deep fusion of multi-modal data, and a closed-loop adaptive control strategy driven by confidence, which can realize early and accurate early warning of insulation faults, accurate identification of fault types and reliable verification of diagnosis conclusions.

[0005] The above object can be achieved by the following scheme: The application discloses an insulation monitoring method for a high-low voltage switch cabinet, which comprises the following steps: collecting a pulse current signal of the high-low voltage switch cabinet in real time; determining whether an insulation fault initial signal exists according to an amplitude, a frequency or a repetition rate of the pulse current signal; if yes, actively injecting a set of alternating current test signals with specific frequencies into the high-low voltage switch cabinet, and simultaneously measuring the pulse current signal, a leakage current and an insulation impedance to obtain transient parameters and steady-state parameters; if no, continuing to collect the pulse current signal in real time; extracting features of the transient parameters and the steady-state parameters, learning a correlation between the features by using a graph neural network, and generating a comprehensive insulation health vector of the high-low voltage switch cabinet; inputting the comprehensive insulation health vector into a pre-trained deep learning diagnosis model for intelligent diagnosis to obtain a diagnosis result and a confidence of an insulation state; comparing the confidence with a preset threshold value, if the confidence is lower than the preset threshold value, dynamically adjusting frequencies and amplitudes of the alternating current test signals according to the diagnosis result, re-executing the active measurement and the determination, if the confidence is higher than the preset threshold value, determining a fault type and an accurate position, quantifying an insulation degradation level, and obtaining a final diagnosis conclusion.

[0006] Optionally, the method further comprises: acquiring historical insulation state data and corresponding environmental noise data of the high-low voltage switch cabinet, and the deep learning diagnosis model is obtained by training the historical insulation state data and the corresponding environmental noise data.

[0007] Optionally, the method further comprises: taking the final diagnosis conclusion as feedback data to update and optimize the deep learning diagnosis model.

[0008] Optionally, the insulation fault initial signal is determined according to that the amplitude, the frequency or the repetition rate of the pulse current signal exceeds a normal operation range within a continuous time.

[0009] Optionally, the step of actively injecting the set of alternating current test signals with specific frequencies into the high-low voltage switch cabinet and simultaneously measuring the pulse current signal, the leakage current and the insulation impedance comprises the following steps: dynamically selecting a set of alternating current test signals with a frequency range in a non-power frequency band according to the insulation fault initial signal; injecting the alternating current test signals into the high-low voltage switch cabinet through a coupling capacitor, measuring the pulse current signal to obtain the transient parameters by using a Rogowski coil, and measuring the leakage current and the insulation impedance to obtain the steady-state parameters by using a Hall sensor.

[0010] Optionally, the feature extraction is performed on the transient parameters and the steady-state parameters, and the correlation between the features is learned by using the graph neural network, including: time domain feature and frequency domain feature extraction is performed on the transient parameters to obtain first features, wherein the first features include pulse amplitude, pulse repetition rate, pulse width, energy spectrum density, and energy distribution; feature extraction is performed on the steady-state parameters to obtain second features, wherein the second features include leakage current phase angle and insulation impedance amplitude; the first features and the second features are taken as feature nodes of the graph neural network, and the nonlinear correlation between the feature nodes is learned by using the graph neural network to generate the comprehensive insulation health vector of the high-low voltage switch cabinet.

[0011] Optionally, the deep learning diagnosis model includes: the deep learning diagnosis model is a multi-task long short-term memory network model, and multi-objective loss function training is performed according to historical insulation state data and corresponding environmental noise data; the multi-objective loss function takes the comprehensive insulation health vector as input and outputs the diagnosis result of the insulation state, the confidence of the diagnosis result, and the signal integrity evaluation score for evaluating the reliability of the diagnosis result; during the training process, the nonlinear correlation between the environmental noise parameters and the signal integrity evaluation score can be learned and established, and the confidence is dynamically corrected during diagnosis.

[0012] Optionally, if the confidence is higher than a preset threshold, then according to the diagnosis result, the fault type and the accurate position are determined, the insulation degradation level is quantified, and the final diagnosis conclusion is obtained, including: the feature node with the highest contribution degree to the diagnosis of the insulation abnormality in the deep learning diagnosis model is analyzed to determine the fault type of the insulation abnormality; the original sensor position associated with the feature node with the highest contribution degree is traced to determine the accurate position of the insulation abnormality; and the insulation degradation level is quantified according to the contribution degree of the feature node to the diagnosis result.

[0013] Based on the same inventive concept, the application also provides a high-low voltage switch cabinet insulation monitoring system, the system comprises: a monitoring module for real-time passive collection of pulse current signals of the high-low voltage switch cabinet, and determining whether there is an insulation fault initial signal according to the amplitude, frequency or repetition rate of the pulse current signals; an active measurement and data processing module for actively injecting a set of specific frequency alternating current test signals into the high-low voltage switch cabinet when there is an insulation fault initial signal, simultaneously measuring the pulse current signals, leakage current and insulation impedance, and performing feature extraction on the measurement data to generate a comprehensive insulation health vector of the high-low voltage switch cabinet; a diagnostic analysis module for inputting the comprehensive insulation health vector into a pre-trained deep learning diagnostic model for intelligent diagnosis to obtain a diagnostic result and a confidence level of the insulation state; a self-adaptive control and conclusion output module for determining whether the confidence level is lower than a preset threshold, if the confidence level is lower than the preset threshold, dynamically adjusting the frequency and amplitude of the alternating current test signals according to the diagnostic result, and controlling the active measurement and data processing module to re-execute the measurement, if the confidence level is higher than the preset threshold, determining the fault type and accurate position, quantifying the insulation degradation level, and obtaining a final diagnostic conclusion; a model optimization module for updating and optimizing the deep learning diagnostic model by taking the final diagnostic conclusion as feedback data.

[0014] Compared with the prior art, the application has the following advantages: 1. The application combines the low-power passive monitoring, non-interference characteristics with the high signal-to-noise ratio, high-precision characteristics of active detection through intelligent trigger type active detection mechanism, and only starts deep measurement when potential fault signs are found, which not only ensures the sensitive capture of early weak fault signals, but also effectively avoids the interference and unnecessary energy consumption of continuous active injection signals on the power grid, realizing the unity of precision and economy of monitoring; 2. The application uses a graph neural network to deeply fuse the synchronous acquired transient and steady-state multi-modal data, which can mine complex nonlinear correlations hidden between different physical quantities that are difficult to find by traditional methods, thereby generating a comprehensive vector that can fully represent the health state of the equipment. This deep information fusion from the data correlation level greatly improves the dimension and depth of fault diagnosis, making it possible to accurately identify the fault type and accurately quantify the degradation level; Other features and advantages of the application will be set forth in the following description of the application, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application can be achieved and obtained by the structures indicated in the specification, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 is a flowchart of a high-low voltage switch cabinet insulation monitoring method according to an embodiment of the present application.

[0017] Figure 2 is a pulse current repetition rate timing diagram according to an embodiment of the present application.

[0018] Figure 3 is a feature parameter correlation heat map according to an embodiment of the present application.

[0019] Figure 4 is a structural diagram of a high-low voltage switch cabinet insulation monitoring system according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] With reference to Figure 1 , one embodiment of the present application proposes a high-low voltage switch cabinet insulation monitoring method, which adopts a combination of passive monitoring and active detection, multi-modal data deep fusion and closed-loop adaptive control strategy driven by confidence, can realize early and accurate early warning of insulation failure, accurate fault location, and obtain high-confidence diagnostic conclusion, and significantly improves the reliability and efficiency of high-low voltage switch cabinet insulation diagnosis.

[0022] The method of the embodiment specifically includes: real-time acquisition of pulse current signals of high-low voltage switch cabinets; Specifically, the sensor continuously acquires pulse current signals on the busbar and cable inside the switch cabinet. This process is a normalized passive monitoring, which aims to capture weak pulse signals generated by early insulation degradation.

[0023] According to the amplitude, frequency or repetition rate of the pulse current signal, it is judged whether there is an insulation fault initial signal. If yes, a set of specific frequency AC test signals is actively injected into the high-low voltage switch cabinet, and the pulse current signal, leakage current and insulation impedance are measured at the same time to obtain transient parameters and steady-state parameters. If no, real-time acquisition continues; Specifically, the characteristics of the collected pulse current signal are analyzed. When these characteristics exceed the set normal range, the active detection mode is started. In this mode, a set of test signals is generated to inject into the switch cabinet, and at the same time, transient parameters such as pulse current and steady-state parameters such as leakage current and insulation impedance are measured to obtain a multi-dimensional data set. If the pulse current signal characteristics are normal, the passive acquisition state is maintained.

[0024] The transient parameters and steady-state parameters are extracted, the correlation between the features is learned using a graph neural network, and a comprehensive insulation health vector of the high-low voltage switch cabinet is generated; Specifically, key features are extracted from transient and steady-state parameters. These features are input into a graph neural network model, and the model learns the mutual influence and complex correlation between these features to generate a comprehensive vector that can fully reflect the insulation health status of the switch cabinet.

[0025] The comprehensive insulation health vector is input into a pre-trained deep learning diagnosis model for intelligent diagnosis to obtain the diagnosis result and confidence of the insulation state; Specifically, the generated comprehensive insulation health vector is input into a deep learning model that has been pre-trained using a large amount of historical data. The model intelligently analyzes the vector to obtain a diagnosis result about the insulation state of the switch cabinet, and at the same time gives a confidence score indicating the reliability of the diagnosis result.

[0026] The confidence is compared with a preset threshold. If the confidence is lower than the preset threshold, the frequency and amplitude of the AC test signal are dynamically adjusted according to the diagnosis result, the active measurement is re-executed, and the judgment is performed. If the confidence is higher than the preset threshold, the fault type and accurate position are determined according to the diagnosis result, the insulation degradation level is quantified, and the final diagnosis conclusion is obtained.

[0027] Specifically, the diagnosed confidence score is compared with a set threshold. If the confidence is low, it is considered that the current data is insufficient to make a reliable judgment, and the test signal parameters will be dynamically adjusted according to the preliminary diagnosis result to re-perform active measurement to obtain clearer insulation parameters. The general rule of dynamic adjustment is generally: when the preliminary diagnosis result tends to "insulation damp" or "contamination", the amplitude of the low-frequency component in the alternating test signal is increased to improve the sensitivity of the measurement of the capacitive change of the insulator. When the preliminary diagnosis result tends to "partial discharge" or "poor contact", the amplitude of the high-frequency component in the alternating test signal is increased to improve the ability to capture transient pulse signals. If the confidence is high, the fault type is determined according to the diagnosis result, the accurate position of the fault is determined using the corresponding algorithm, and the risk level of insulation degradation is quantified, and finally a reliable diagnosis conclusion is obtained.

[0028] The passive monitoring and active detection mode is adopted, the multi-dimensional parameters are deeply fused by using the graph neural network, the comprehensive insulation health vector is formed, the measurement strategy can be dynamically adjusted according to the diagnosis confidence, the more distinguishable insulation parameters can be autonomously obtained, and finally the diagnosis conclusion with high reliability is obtained, and the reliability and efficiency of the insulation monitoring of the high-low voltage switch cabinet are significantly improved.

[0029] Optionally, the method further comprises: acquiring historical insulation state data and corresponding environmental noise data of the high-low voltage switch cabinet, and the deep learning diagnosis model is obtained by training the historical insulation state data and the corresponding environmental noise data.

[0030] Specifically, a database for storing switch cabinet operation data is first established, and past insulation state diagnosis results, measurement parameters and operation condition information are continuously recorded. Environmental noise sensors are deployed inside the switch cabinet to continuously collect noise data generated by device operation, switch opening and closing, etc. The collected historical insulation state data and environmental noise data are used as a training set to train the deep learning model, so that the deep learning model can learn and identify the pattern of insulation failure from a large amount of data.

[0031] Optionally, the method further comprises: updating and optimizing the deep learning diagnosis model by taking the final diagnosis conclusion as feedback data.

[0032] Specifically, after the final diagnosis conclusion is obtained, this conclusion together with the measurement parameters collected at that time will be used as new training samples. These new training samples are added to the training data set of the model for retraining and fine-tuning of the deep learning model, so that the deep learning model can continuously learn new insulation failure patterns. Through this continuous feedback cycle, the model can adapt to changes in the device operating environment, thereby maintaining the accuracy and reliability of the diagnosis.

[0033] Optionally, the basis for determining the insulation fault initial signal is that the amplitude, frequency or repetition rate of the pulse current signal exceeds the normal operating range in the duration.

[0034] Specifically, a normal operating range baseline based on historical data or preset values is established. The baseline is determined by calculating the mean and standard deviation from the pulse current signals collected under normal operating conditions of the equipment. The amplitude, frequency and repetition rate of the pulse current signal are analyzed in real time. If any of these characteristics continuously exceeds the normal operating range defined by the baseline, a timer will be started. Only when the time of exceeding the normal range reaches or exceeds the set duration, the event will be judged as an insulation fault initial signal, in order to avoid misjudgment caused by transient interference. Figure 2 As shown, the change of pulse current repetition rate over time is shown, and when it continuously exceeds the preset threshold, it is judged that there is an insulation fault initial signal.

[0035] Optionally, the active injection of a set of specific frequency AC test signals into the high and low voltage switch cabinet, and the simultaneous measurement of the pulse current signal and the leakage current and insulation impedance include: According to the insulation fault initial signal, a set of AC test signals with frequency range in the non-power frequency band is dynamically selected; Specifically, after obtaining the insulation fault initial signal through passive monitoring, preliminary diagnosis is performed. Even if the confidence of the diagnosis is not high, the potential fault type can be identified, such as partial discharge or insulation damp. In order to detect more accurately, a set of AC test signals with frequency range in the non-power frequency band is dynamically selected. Selecting non-power frequency signals can effectively avoid aliasing interference with the main power grid power frequency signals.

[0036] The AC test signal is injected into the high and low voltage switch cabinet through the coupling capacitor, and the transient parameters are measured by using the Rogowski coil to measure the pulse current signal, and the steady-state parameters are measured by using the Hall sensor to measure the leakage current and insulation impedance.

[0037] Specifically, the selected AC test signal is safely injected into the switch cabinet through the high-voltage coupling capacitor to generate a measurable response. During the measurement, the Rogowski coil is used as a non-contact high-frequency sensor to accurately capture the transient pulse current signals generated by insulation abnormalities such as partial discharge. These signals are transient parameters that characterize the characteristics of the fault moment. At the same time, the Hall sensor is used to measure the weak leakage current flowing through the insulation under the action of the AC test signal. By analyzing the phase relationship between the leakage current and the injected voltage, the insulation impedance can be further calculated, which constitutes the steady-state parameters of the long-term health status of the insulation. These two measurement methods are carried out in parallel, which can obtain comprehensive dynamic and static information at the initial stage of the fault, providing a comprehensive data basis for the subsequent graph neural network deep fusion.

[0038] Optionally, the feature extraction is performed on the transient parameters and the steady-state parameters, and the correlation between the features is learned by using a graph neural network, including: The time-domain features and frequency-domain features of the transient parameters are extracted to obtain first features, wherein the first features include pulse amplitude, pulse repetition rate, pulse width, and energy spectrum density and energy distribution. Specifically, the feature extraction of the transient parameters is a multi-angle analysis process, which is divided into time domain and frequency domain. The time domain features are directly obtained from the signal waveform, including pulse amplitude , pulse width and pulse repetition rate . The pulse repetition rate is an important indicator for measuring the activity of the fault, and its calculation formula is: , wherein represents the number of pulses monitored within a certain time length . The frequency domain features are used to reveal the internal frequency composition of the pulse signal, and the core is the energy spectrum density , which is obtained by Fourier transform of the autocorrelation function of the time domain pulse signal, and reflects the energy distribution of the signal at different frequencies. The formula can be expressed as: , wherein is the autocorrelation function, is the frequency.

[0039] The feature extraction is performed on the steady-state parameters to obtain second features, wherein the second features include leakage current phase angle and insulation impedance amplitude. Specifically, the feature extraction of the steady-state parameters aims to evaluate the overall insulation health of the switch cabinet. This process relies on the injected AC test signal, which accurately measures the leakage current and the amplitude and phase of the injected voltage . The insulation impedance is a direct indicator of the performance of the insulator, and its calculation formula is: , The amplitude of the insulation impedance characterizes the resistance performance of the insulator, and the phase angle of the leakage current reflects the capacitance characteristics of the insulator, both of which are key steady-state features for diagnosing insulation aging.

[0040] The first features and the second features are used as feature nodes of the graph neural network, and the non-linear correlation between the feature nodes is learned by using the graph neural network to generate a comprehensive insulation health vector of the high and low voltage switch cabinet.

[0041] Specifically, the extracted time-domain and frequency-domain features and steady-state parameter features are taken as feature nodes of the graph neural network. The graph neural network is designed to learn and model the complex nonlinear correlations between these nodes. The correlation of the graph neural network is not pre-set, but is learned through a training process in a data-driven manner, so as to capture the collaborative change patterns between multiple parameters. The network aggregates the information of each feature node and its neighbor nodes through multi-layer graph convolution operations, to generate a comprehensive insulation health vector that can comprehensively represent the overall health condition of the insulation system .This process can be represented by the following formula: wherein, is a nonlinear activation function in the network, is an input matrix containing all features, represents the weight parameters learned by the model in training. As shown in Figure 3 , a correlation heat map between different insulation feature parameters is shown, and the darker the color, the stronger the correlation.

[0042] Optionally, the deep learning diagnosis model comprises: The deep learning diagnosis model is a multi-task long short-term memory network model, which is trained according to historical insulation state data and corresponding environmental noise data through a multi-objective loss function; Specifically, the deep learning diagnosis model adopts a multi-task learning model composed of two layers of long short-term memory networks and three fully connected output layers. Each long short-term memory network layer contains 64 hidden units, which is responsible for processing the time sequence health vector and capturing the dynamic changes of the insulation state. The three fully connected output layers are respectively used for predicting the insulation state, the confidence and the risk level. The model is trained using a large amount of historical insulation state data and corresponding environmental noise data, and the training target is to minimize the prediction errors of multiple tasks through a multi-objective loss function . The loss function is the weighted sum of the losses of each task, and the formula is as follows: wherein, is the loss of the insulation state diagnosis result, is the loss of the confidence prediction, is the loss of the signal integrity evaluation score, is the weight coefficient of each loss. In the training process, the Adam optimizer is adopted, the initial learning rate is set to 0.001, and the early stopping mechanism is set to avoid overfitting.​​​​

[0043] The multi-target loss function takes the comprehensive insulation health vector as input, and outputs the diagnosis result of the insulation state, the confidence of the diagnosis result, and a signal integrity evaluation score for evaluating the reliability of the diagnosis result. Specifically, the trained multi-task long short-term memory network model takes the comprehensive insulation health vector as input, and can simultaneously perform prediction of multiple tasks. It not only outputs the diagnosis result of the insulation state and the confidence of the diagnosis result, but also outputs a signal integrity evaluation score for evaluating the reliability of the diagnosis result . The nonlinear mapping function is an output sublayer of a deep learning model, and the score is obtained by the model according to the environmental noise parameter in the input vector, and the formula is represented as: , wherein, is the environmental noise parameter, which represents the quality of the signal.

[0044] During the training process, the nonlinear correlation between the environmental noise parameter and the signal integrity evaluation score can be learned and established, and the confidence is dynamically corrected during diagnosis.

[0045] Specifically, during the training process, the model can learn and establish the nonlinear correlation between the environmental noise parameter and the signal integrity evaluation score , so as to understand the influence of noise on the accuracy of diagnosis. During actual diagnosis, the confidence of the initial prediction will be dynamically corrected . The corrected confidence can be represented by the formula: , wherein, the correction factor is a monotonic increasing function of the signal integrity evaluation score . When is lower, is smaller, thereby reducing the final confidence, and being more cautious in the face of low-quality signals to avoid misjudgment.

[0046] Optionally, if the confidence is higher than a preset threshold, the fault type and the accurate position are determined according to the diagnosis result, the insulation degradation level is quantified, and the final diagnosis conclusion is obtained, including: The fault type of the insulation abnormality is determined by analyzing the feature node with the highest contribution to the diagnosis of the insulation abnormality in the deep learning diagnosis model. Specifically, after the deep learning diagnosis model obtains the diagnosis result of insulation abnormality, contribution degree analysis is performed on the feature nodes that have the greatest impact on the insulation abnormality diagnosis result. For example, when the insulation is damp, the contribution degree of the feature node related to the leakage current phase angle will be significantly improved. According to the feature node with the highest contribution degree and the associated mode, the specific fault type of the insulation abnormality is determined, such as partial discharge, insulation damp, or cable aging, etc.

[0047] The original sensor position associated with the feature node with the highest contribution degree is traced back to determine the accurate position of the insulation abnormality. Specifically, after the fault type is determined, the accurate position thereof needs to be determined. This process uses the pulse current signals collected by the multiple sensor arrays deployed inside the switch cabinet to calculate the electrical signal time difference of the signals received by the sensors at different positions , and applies a triangular positioning algorithm to solve it. For example, when three sensors S1, S2, and S3 are used, the discharge point coordinates can be determined according to the signal time difference by solving the following equation set: , wherein, is the propagation speed of electromagnetic waves in air, , , is the electrical signal time delay measured by the three sensors, , , is the coordinates of the three sensors.

[0048] According to the contribution degree of the feature node to the diagnosis result, the insulation deterioration level is quantified.

[0049] Specifically, the quantification of the insulation deterioration level is based on the contribution degree of the feature node to the diagnosis result. The comprehensive contribution degree of the feature node group with the highest contribution degree is calculated , and is mapped to a preset insulation deterioration level . The nonlinear mapping function is an output sublayer of the deep learning model, which maps to a specific risk level through learning, and the formula is: , wherein, is the nonlinear mapping function. In this way, the interpretive output of the model is directly converted into an operable risk level.

[0050] Based on the same inventive concept, as shown in Figure 4 , the present application also provides a high-low voltage switch cabinet insulation monitoring system, which comprises: a monitoring module for collecting pulse current signals of the high-low voltage switch cabinet in real time, and determining whether there is an initial insulation fault signal according to the amplitude, frequency or repetition rate of the pulse current signals; an active measurement and data processing module for actively injecting a set of alternating test signals of specific frequencies into the high-low voltage switch cabinet when there is an initial insulation fault signal, simultaneously measuring pulse current signals, leakage currents and insulation impedance, and performing feature extraction on the measurement data to generate a comprehensive insulation health vector of the high-low voltage switch cabinet; a diagnostic analysis module for inputting the comprehensive insulation health vector into a pre-trained deep learning diagnostic model for intelligent diagnosis to obtain a diagnostic result and a confidence level of the insulation state; a self-adaptive control and conclusion output module for determining whether the confidence level is lower than a preset threshold; if the confidence level is lower than the preset threshold, dynamically adjusting the frequency and amplitude of the alternating test signals according to the diagnostic result, and controlling the active measurement and data processing module to re-perform the measurement, and if the confidence level is higher than the preset threshold, determining a fault type and an accurate position, quantifying an insulation degradation level, and obtaining a final diagnostic conclusion according to the diagnostic result; a model optimization module for updating and optimizing the deep learning diagnostic model by taking the final diagnostic conclusion as feedback data.

[0051] Embodiment 1: The present application adopts a high-low voltage switch cabinet insulation monitoring method and system, and applies it in a certain high-voltage substation. The substation undertakes the task of power supply for the core area, and its health status is directly related to the reliability of the substation. The present application combines the process test and the predictive test of the insulation monitoring method, which can effectively prevent potential insulation degradation and ensure the stability of the power grid when insulation failure occurs.

[0052] The substation selects 10 switch cabinets as the test group, deploys the insulation monitoring system of the present application, and sets 10 switch cabinets of the same type and working condition as the control group, which uses the traditional periodic inspection and offline test method. The test cycle is 6 months, and the operation data of the switch cabinets in the test group and the control group are continuously collected and analyzed during the period to verify the effectiveness of the present application.

[0053] When the monitoring module of the present application monitors an A-phase incoming line cabinet in real time, it is found that the repetition rate of the pulse current signal of the cabinet continuously exceeds the preset time, which is set to 5 minutes, and is always higher than the baseline of the normal operating range, indicating that there is an initial insulation fault signal, while the traditional inspection of the control group does not find any abnormalities.

[0054] The active measurement and data processing module of the A-phase incoming line cabinet is activated immediately. According to the initially discovered signal, a set of AC test signals in the non-power frequency band ranging from 50 kHz to 500 kHz is dynamically selected and injected. The transient pulse current signal, leakage current and insulation impedance are obtained through the coupling capacitor injection switch cabinet, the Rogowski coil and Hall sensor synchronous measurement. Then, all parameters are extracted and feature domain extraction is performed to obtain the comprehensive health vector reflecting the insulation abnormality degree, energy intensity and multi-parameter space-time correlation. The insulation health vector is input into the deep learning model. After the model executes the judgment, the preliminary diagnosis result of "insulation damp" is found.

[0055] The confidence of the diagnosis result is 65%. Since the confidence is lower than the preset confidence threshold of 85%, the adaptive control module immediately makes the preliminary diagnosis result of "continue detection", dynamically adjusts the AC test signal, enhances the low-frequency component, and instructs the active measurement module to re-execute a round of targeted active detection. After the second round of measurement and diagnosis, the diagnosis result is still "insulation damp", but the confidence is significantly improved to 92%.

[0056] The final diagnosis conclusion with high confidence is obtained: by analyzing the feature nodes with the highest contribution to the insulation in the model, it is found that the "leakage current phase angle" and "insulation energy spectrum density" have the highest contribution, so the fault type is determined to be insulation damp. By tracing the original sensor associated with the feature node with the highest contribution, the fault location is accurately determined to be "cable terminal head accessory", and according to the contribution of the feature node, the insulation degradation level is quantified as "slight".

[0057] The final conclusion is feedback data, which is used by the model module for subsequent retraining of the deep learning model. According to this final conclusion, the operation and maintenance personnel arranged a timely drying and reinforcement treatment of the cable terminal head in the short power-off window the next day, avoiding further development of the fault.

[0058] Table 1 Comparison table of early fault monitoring Table 2 Data of adaptive diagnosis process Table 3 Comparison table of final diagnosis conclusion and traditional method From the data in Table 1, Table 2 and Table 3 above, it can be seen that the high-low voltage switch cabinet insulation monitoring method and system proposed by the present application exhibits significant technical advantages in actual application. Table 1 clearly shows that the present application can capture signs of weak insulation failure in real time, achieving precise early warning of failure, with a confidence level much higher than traditional methods. The data in Table 2 details the unique confidence-driven self-adaptive closed-loop diagnostic process of the present application. Through intelligent response to low-confidence results, the test strategy is dynamically adjusted and iteratively verified, successfully converting uncertain preliminary judgments into high-confidence final results, greatly improving the accuracy and reliability of diagnosis. The comparative results in Table 3 highlight the ultimate value of the present application. The present application can detect faults in advance and provide detailed diagnostic reports including fault type, precise location and quantitative degradation level, achieving a leap from "detecting abnormalities" to "precise positioning and quantification", providing strong technical support for condition-based maintenance and predictive maintenance, and effectively ensuring the safe and stable operation of critical power equipment.

[0059] It should be noted that the electrical connection between the above-mentioned units does not necessarily represent the direct connection of the line, and the indirect connection mode can also be applied to the embodiments of the present application as long as the purpose of the present application is achieved. The above-mentioned is only an exemplary embodiment of the present application, and cannot limit the scope of the present application.

[0060] That is, any equivalent changes and modifications made according to the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the specification and practice of the true disclosure. The present application is intended to cover any variations, uses or adaptive changes of the present application that follow the general principles of the present application and include common knowledge or conventional technical means in the art not disclosed by the present application.

Claims

1. A method for monitoring insulation of high and low voltage switchgear, characterized in that: The method comprises: Real-time collection of pulse current signals of high and low voltage switchgear; Based on the amplitude, frequency or repetition rate of the pulse current signal, determine whether there is an initial signal of insulation fault. If yes, actively inject a set of AC test signals of specific frequencies into the high and low voltage switchgear, and simultaneously measure the pulse current signal as well as the leakage current and insulation impedance to obtain transient parameters and steady-state parameters. If no, continue real-time acquisition; Extracting features of the transient parameters and steady-state parameters, using a graph neural network to learn the correlation between the features, and generating a comprehensive insulation health vector of the high and low voltage switchgear; Inputting the comprehensive insulation health vector into a pre-trained deep learning diagnostic model for intelligent diagnosis to obtain a diagnosis result and confidence level of the insulation status; The confidence level is compared with a preset threshold. If the confidence level is lower than the preset threshold, the frequency and amplitude of the AC test signal are dynamically adjusted based on the diagnosis result, and the active measurement is re-executed and a judgment is made. If the confidence level is higher than the preset threshold, the fault type and precise location are determined based on the diagnosis result, the insulation degradation level is quantified, and a final diagnosis conclusion is drawn.

2. A method for monitoring insulation of high and low voltage switchgear according to claim 1, characterized in that: The method also includes: obtaining historical insulation status data of high and low voltage switchgear and corresponding environmental noise data, and the deep learning diagnostic model is trained based on the historical insulation status data and the corresponding environmental noise data.

3. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 1, characterized in that: The method also includes: using the final diagnostic conclusion as feedback data to update and optimize the deep learning diagnostic model.

4. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 1, characterized in that: The basis for determining the initial signal of the insulation fault is that the amplitude, frequency or repetition rate of the pulse current signal exceeds the normal operating range within the duration.

5. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 1, characterized in that: The method of actively injecting a set of AC test signals of specific frequencies into the high and low voltage switchgear and simultaneously measuring the pulse current signal, leakage current, and insulation impedance includes: Dynamically selecting a set of AC test signals with a frequency range in a non-power frequency band according to the insulation fault initial signal; The AC test signal is injected into the high and low voltage switch cabinet through the coupling capacitor, the pulse current signal is measured using the Rogowski coil to obtain transient parameters, and the leakage current and insulation resistance are measured using the Hall sensor to obtain steady-state parameters.

6. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 1, characterized in that: Extracting features of the transient parameters and steady-state parameters and using a graph neural network to learn the correlation between the features includes: Extracting time domain features and frequency domain features of the transient parameters to obtain first features, wherein the first features include pulse amplitude, pulse repetition rate and pulse width, energy spectrum density and energy distribution; Performing feature extraction on the steady-state parameter to obtain a second feature, wherein the second feature includes a leakage current phase angle and an insulation resistance amplitude; The first and second features are used as feature nodes of the graph neural network. The nonlinear correlation between the feature nodes is learned through the graph neural network to generate a comprehensive insulation health vector of the high and low voltage switchgear.

7. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 2, characterized in that: The deep learning diagnostic model includes: The deep learning diagnostic model is a multi-task long short-term memory network model that performs multi-objective loss function training based on historical insulation state data and corresponding environmental noise data; The multi-objective loss function takes the comprehensive insulation health vector as input and outputs the diagnosis result of the insulation status, the confidence level of the diagnosis result, and the signal integrity assessment score for evaluating the reliability of the diagnosis result; During the training process, it is possible to learn and establish a nonlinear correlation between environmental noise parameters and signal integrity assessment scores, and dynamically correct the confidence level during diagnosis.

8. The method for monitoring insulation of a high and low voltage switch cabinet according to claim 1, characterized in that: If the confidence level is higher than a preset threshold, the fault type and precise location are determined based on the diagnosis results, the insulation degradation level is quantified, and the final diagnosis conclusion is obtained, including: Determining the fault type of the insulation abnormality by analyzing the feature nodes with the highest contribution to the diagnosis of the insulation abnormality in the deep learning diagnosis model; Tracing back the original sensor position associated with the characteristic node with the highest contribution, and determining the precise position of the insulation anomaly; The insulation degradation level is quantified based on the contribution of the characteristic nodes to the diagnosis results.

9. A high and low voltage switchgear insulation monitoring system, characterized in that: The system is used in a high- and low-voltage switchgear insulation monitoring method according to any one of claims 1 to 8, and the system includes: The monitoring module is used to passively collect the pulse current signals of the high and low voltage switchgear in real time, and determine whether there is an initial signal of insulation fault based on the amplitude, frequency or repetition rate of the pulse current signals; An active measurement and data processing module is used to actively inject a set of AC test signals of specific frequencies into the high and low voltage switchgear when an initial insulation fault signal is present, while simultaneously measuring the pulse current signal, leakage current, and insulation impedance. Feature extraction is then performed on the measurement data to generate a comprehensive insulation health vector for the high and low voltage switchgear. A diagnostic analysis module, configured to input the comprehensive insulation health vector into a pre-trained deep learning diagnostic model for intelligent diagnosis to obtain a diagnostic result and confidence level of the insulation status; An adaptive control and conclusion output module is used to determine whether the confidence level is lower than a preset threshold. If the confidence level is lower than the preset threshold, the frequency and amplitude of the AC test signal are dynamically adjusted based on the diagnosis result, and the active measurement and data processing module is controlled to re-execute the measurement. If the confidence level is higher than the preset threshold, the fault type and precise location are determined based on the diagnosis result, the insulation degradation level is quantified, and a final diagnosis conclusion is drawn. The model optimization module is used to update and optimize the deep learning diagnosis model using the final diagnosis conclusion as feedback data.

Images