Test method for detecting power frequency withstand voltage of insulation tool

Abstract

The invention relates to the technical field of insulation tool detection, and provides an insulation tool detection power frequency withstand voltage test method, which realizes dynamic evaluation and risk prediction of insulation performance by applying preset high voltage to an insulation tool and collecting electrical signals in real time in combination with a multi-modal data processing and intelligent analysis algorithm. The method specifically comprises the following steps: acquiring voltage, current and related environmental parameters of an insulation tool, and performing feature extraction and anomaly identification on a signal by using an adaptive fuzzy prediction algorithm; performing clustering analysis on the abnormal waveform through a Gaussian mixture model, and calculating an insulation risk probability of the tool; and outputting a comprehensive risk assessment and health status report in combination with historical data and a real-time measurement result. According to the method, the insulation performance of the insulation tool can be continuously monitored and accurately judged without shutdown, and the high-voltage operation safety and the equipment management efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of testing technology for insulating tools and equipment, and in particular to a method for testing power frequency withstand voltage of insulating tools and equipment. Background Technology

[0002] Insulating tools are crucial safety equipment for high-voltage electrical work, and their insulation performance directly affects the personal safety of workers and the stable operation of power equipment. Therefore, it is necessary to assess the insulation status of insulating tools.

[0003] Currently, the insulation performance of insulating tools is mainly tested using preventative testing methods, which involve conducting offline withstand voltage tests at fixed intervals. This method typically involves applying power frequency or DC high voltage and measuring electrical parameters such as leakage current and insulation resistance, then judging whether the insulation of the tool is up to standard based on empirical thresholds. While this approach has played a role in long-term engineering practice, the limitations of existing testing methods are becoming increasingly apparent with the continuous expansion of power grids, the widespread application of new composite insulation materials, and the increasing demands for lean operation and maintenance.

[0004] First, traditional withstand voltage tests rely on a limited number of static electrical parameters as criteria. However, insulating materials often exhibit weak, nonlinear, and non-stationary electrical characteristics in the early stages of aging, such as multi-mode partial discharge signals or complex harmonic components. These characteristics are key early indicators of insulation defects (such as air gaps, cracks, and interface degradation), but they are difficult to capture using traditional power frequency withstand voltage tests, resulting in a low detection rate and a high risk of missed detections for latent defects. Furthermore, relying on fixed thresholds for acceptance testing is ill-suited to the increasingly complex insulation structures. When dealing with multi-section insulating rods, composite insulators, and other similar tools, the problems of misjudgment and insufficient adaptability become more pronounced, and further localization and analysis of potential defects are impossible.

[0005] Secondly, existing testing methods primarily rely on manual offline testing, which is characterized by long testing cycles, large equipment size, high operator skill requirements, and low overall efficiency. Furthermore, the manual recording and judgment process is easily influenced by personnel experience and operating habits, making it difficult to guarantee the integrity and traceability of test data. In addition, this method cannot continuously and non-destructively monitor the condition of large quantities of in-use tools and equipment, forcing maintenance strategies to rely mainly on periodic inspections. This makes it difficult to detect dynamic degradation trends in insulation performance in a timely manner, increasing safety risks.

[0006] In recent years, some studies have attempted to introduce signal processing and machine learning techniques to improve the ability to identify insulation conditions. However, the feature extraction of such methods is still limited, and the models are difficult to effectively cope with the multi-source and complex signal environment in the field. Most of them only achieve simple classification of abnormal signals, and there is still a lack of effective technical means for the location, severity and collaborative analysis of defects inside the insulation structure and multiple types of defects, which makes it difficult to meet the needs of practical applications.

[0007] Therefore, it is necessary to provide a power frequency withstand voltage test method for testing insulating tools to solve the problems existing in the prior art. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the prior art by providing a power frequency withstand voltage test method for testing insulating tools, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for testing the power frequency withstand voltage of insulating tools includes the following steps: S1. Based on the historical power frequency withstand voltage data, environmental parameters, and insulation loss factor of the insulating tool under test, the starting voltage, voltage ramp rate, and target withstand voltage value of the withstand voltage test are set using an adaptive fuzzy prediction algorithm. S2. During the boost and withstand voltage tests, a high-precision sensor array with multiple channels is used to collect the current and voltage signals of each channel in real time. The sampling gain of each channel is dynamically adjusted based on the active identification-gain matching algorithm to enhance the perception of weak abnormal signals. S3. When leakage current or abnormal signal is detected, multi-scale wavelet transform is performed on the abnormal interval waveform and the amplitude, statistical features and energy at each scale are extracted. The abnormal waveform is clustered using a self-learning deep clustering algorithm, and the abnormal waveform identified by clustering is combined with the sensor position information of the corresponding acquisition channel to achieve abnormal location. S4. The withstand voltage process adopts a multi-cycle progressive approach. After each cycle, the historical and current cycle features are integrated, and Bayesian discrimination is used to calculate the posterior probability of failure. The discrimination threshold is dynamically corrected based on gradient reduction calibration to optimize sensitivity and false alarm rate. S5. After all testing and analysis are completed, the test results are standardized, weighted, and graded based on the comprehensive fuzzy reasoning scoring model to generate safe usage suggestions and subsequent maintenance or re-inspection plans.

[0010] Preferably, the adaptive fuzzy prediction algorithm in S1 uses the input vector The inputs are: D, where D is the insulation loss factor obtained through standard power frequency loss measurement; T is the ambient temperature; H is the ambient humidity; and R is the insulation dielectric resistance, aligned with historical data by timestamps. By establishing several fuzzy subsets for each input variable, and through fuzzy inference, the initial voltage is obtained by weighted average or maximum membership. Boost rate and target withstand voltage V max .

[0011] Preferably, the fuzzy subset is represented by a three- or four-segment membership function, and the membership function is in the form of a trapezoid or a triangle.

[0012] Preferably, in S2, the active identification-gain matching algorithm first calculates the instantaneous amplitude of each channel. ; in, and The first Instantaneous current and voltage of the channel, Channel number, For time; And estimate the rate of change: ; in Then Compared with historical average and standard deviation In comparison, when The channel gain will be adjusted accordingly. Increase a dynamic weight .

[0013] Preferably, the active identification-gain matching algorithm also incorporates real-time signal-to-noise ratio: ; in The current channel noise amplitude is estimated using sliding window variance estimation or spectral noise estimation; when Below the preset threshold Adjusted proportionally If after adjustment If the set limit is exceeded, a rollback will be triggered.

[0014] Preferably, in S3, the wavelet basis functions are Daubechies or Symlet, which decompose the waveform of the abnormal interval into n scales to obtain the wavelet coefficients at each scale. , where j represents the scale and k is the time-domain sampling point; Then calculate the energy at each scale. ; Forming energy eigenvectors Simultaneously, extract the peak amplitude of the original waveform. And statistical characteristics, ultimately the data feature vector of each abnormal waveform segment is obtained. Input into the self-learning deep clustering module.

[0015] Preferably, the self-learning deep clustering module uses a Gaussian mixture model for parameter estimation, and the number of clusters K is determined by the Bayesian information criterion, the Akaike information criterion, or an adaptive split-merge strategy. The clustering results are output in probabilistic form. ,in It is the probability density function of a Gaussian distribution.

[0016] Preferably, the clustering labels are associated with the spatial coordinate information of the sensor and back-labeled to the structure of the tool being tested, thereby realizing the spatial mapping and positioning of suspected abnormal areas.

[0017] Preferably, in S4, the first Feature vectors are collected over each withstand voltage cycle. Calculate posterior probability based on Bayesian discriminant method ; when Current periodic warning threshold The system provides timely warnings of potential failures; to achieve threshold adaptation, a discriminative loss function L is defined, and based on the discriminative error... ; in For actual feedback tags; Update the threshold using gradient descent: ; Where L is the preset discriminant loss function and γ is the learning rate.

[0018] Preferably, in S5, the comprehensive fuzzy reasoning scoring model standardizes, weights, and maps the detected indicators to risk levels, which include "safe", "warning", "critical" and "high risk", and automatically generates corresponding handling suggestions based on the level.

[0019] The present invention discloses a method for testing the power frequency withstand voltage of insulating tools, which has the following beneficial effects.

[0020] This invention enables real-time acquisition and analysis of electrical signals during the use of insulating tools, without requiring shutdown or disassembly of the equipment, thus achieving continuous and dynamic monitoring of insulation performance. Through real-time assessment and risk prediction of the insulation performance of insulating tools, potential faults can be detected promptly, significantly reducing personal and equipment safety risks during high-voltage operations.

[0021] By combining multimodal data processing, fuzzy prediction algorithms, and cluster analysis, this invention can accurately identify insulation anomalies, provide scientific risk assessments and health status reports, and improve the accuracy of judgments.

[0022] Health status assessments can be conducted based on real-time monitoring results and historical data to guide maintenance and replacement decisions, optimize management processes, and reduce operation and maintenance costs. Attached Figure Description

[0023] Figure 1 This is a flowchart of the power frequency withstand voltage test method for testing insulating tools according to the present invention.

[0024] Figure 2 This is a schematic diagram of the overall layout for the power frequency discharge characteristic test of the present invention.

[0025] Figure 3 This is a schematic diagram of the test specimen for the power frequency discharge characteristics test of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0028] As disclosed in the background section, existing withstand voltage testing methods for insulating tools mostly rely on fixed withstand voltage curves or manually set parameters. This makes it difficult to simultaneously consider both the historical condition of the tool and changes in the field environment. Consequently, the sensitivity is low in detecting early-stage weak discharges, spike leaks, and latent material damage, and the methods are highly dependent on manual intervention and have long testing cycles. Furthermore, existing methods have poor adaptability to tools made of different materials and structures, easily leading to misjudgments or missed detections, thus affecting operational safety.

[0029] To address the aforementioned problems, this invention proposes a method for testing the power frequency withstand voltage of insulating tools, comprising the following steps: S1. Based on the historical power frequency withstand voltage data, environmental parameters, and insulation loss factor of the insulating tool under test, the starting voltage, voltage ramp rate, and target withstand voltage value of the withstand voltage test are set using an adaptive fuzzy prediction algorithm. S2. During the boost and withstand voltage tests, a high-precision sensor array with multiple channels is used to collect the current and voltage signals of each channel in real time. The sampling gain of each channel is dynamically adjusted based on the active identification-gain matching algorithm to enhance the perception of weak abnormal signals. S3. When leakage current or abnormal signal is detected, multi-scale wavelet transform is performed on the abnormal interval waveform and the amplitude, statistical features and energy at each scale are extracted. The abnormal waveform is clustered using a self-learning deep clustering algorithm and the clustering results are mapped with the sensor position information to achieve abnormal location. S4. The withstand voltage process adopts a multi-cycle progressive approach. After each cycle, the historical and current cycle features are integrated, and Bayesian discrimination is used to calculate the posterior probability of failure. The discrimination threshold is dynamically corrected based on gradient reduction calibration to optimize sensitivity and false alarm rate. S5. After all testing and analysis are completed, the test results are standardized, weighted, and graded based on the comprehensive fuzzy reasoning scoring model to generate safe usage suggestions and subsequent maintenance or re-inspection plans.

[0030] The technical solution provided by this invention addresses the shortcomings of traditional detection methods in identifying early-stage minute defects and nonlinear anomalies. It proposes to improve the detection sensitivity of weak discharge behavior and latent material damage in complex insulation degradation mechanisms through multi-source signal fusion and deep feature mining. This enables accurate identification and localization of early faults. Furthermore, it addresses the limitations of existing detection methods in adapting to the diversity of tool structures and materials.

[0031] This invention, through adaptive algorithms and multiple risk criteria, can automatically adjust assessments to meet the personalized inspection needs of tools under different types and operating conditions. Furthermore, addressing the problems of low testing efficiency, high reliance on manual labor, and high testing and management costs in existing technologies, this invention designs an automated integrated process to achieve an intelligent closed loop from data acquisition and anomaly analysis to risk assessment and report output, significantly shortening the testing cycle, reducing manual operations, and lowering maintenance expenses.

[0032] The above plan will be explained in detail below.

[0033] Please refer to Figure 1 A method for testing the power frequency withstand voltage of insulating tools includes the following steps: S1. Based on the historical power frequency withstand voltage data, environmental parameters, and insulation loss factor of the insulating tool under test, the starting voltage, voltage ramp rate, and target withstand voltage value of the withstand voltage test are set using an adaptive fuzzy prediction algorithm. Preferably, in this embodiment, the adaptive fuzzy prediction algorithm in S1 uses the input vector The inputs are: D, where D is the insulation loss factor obtained through standard power frequency loss measurement; T is the ambient temperature; H is the ambient humidity; and R is the insulation dielectric resistance, aligned with historical data by timestamps. By establishing several fuzzy subsets for each input variable, and through fuzzy inference, the initial voltage is obtained by weighted average or maximum membership. Boost rate and target withstand voltage V max Taking into account multiple factors such as the insulation loss factor of tools and equipment and environmental disturbances, the optimal withstand voltage curve is automatically generated by using multi-factor fuzzy control reasoning, which fundamentally avoids the safety risks and errors caused by previous manual subjective setting or fixed curve schemes.

[0034] Preferably, in this embodiment, the fuzzy subset is represented by a three-segment or four-segment membership function, and the membership function is in the form of a trapezoid or a triangle.

[0035] Specifically, in this embodiment, historical power frequency withstand voltage test data of the insulating tool under test are first collected, and current environmental parameters, including temperature T, humidity H, and the resistance R of the insulating medium, are acquired in real time. All of these data will be used as input to the adaptive fuzzy prediction algorithm.

[0036] The algorithm uses the insulation loss factor (denoted as D), ambient temperature T, ambient humidity H, and insulation resistance R are the main input variables. T, H, and R are recorded in units of ℃, %, and MΩ, respectively, and are aligned with historical data by timestamp. By establishing a multi-factor fuzzy control inference system, the starting voltage required for the withstand voltage test is determined. Boost rate It performs joint prediction and dynamic setting with the target withstand pressure value Vmax.

[0037] In the specific implementation process, the input vector is first defined. Using fuzzy membership functions ( (where X is any factor in X), and each input subset is divided into 3 fuzzy subsets, such as "low", "medium", and "high", using a trapezoidal function as the membership function. An example of the interval is as follows: Each fuzzy subset uses a four-parameter trapezoidal function as its membership function to ensure appropriate overlap between adjacent subset intervals, thereby guaranteeing full coverage and smooth transition of the input space. The parameters [a, b, c, d] of the trapezoidal membership function are defined as follows, taking insulation loss factor D, temperature T, humidity H, and insulation resistance R as factors: D (insulation loss factor): Low: [0, 0, 0.01, 0.015]; In the middle: [0.01, 0.015, 0.02, 0.03]; High: [0.02, 0.03, 0.05, 0.05]; T (Temperature, unit: °C): Low: [-10, -10, 5, 18]; Middle: [10, 20, 30, 40]; Height: [35, 45, 60, 60]; H (humidity, unit: %) Low: [0, 0, 30, 50]; Chinese: [35, 50, 70, 85]; Height: [75, 90, 100, 100]; R (insulation resistance, unit: MΩ): Low: [0, 0, 20, 80]; Chinese: [50, 150, 300, 500]; High: [400, 800, 5000, 5000].

[0038] Note: The parameter order is [a, b, c, d], which corresponds to the four inflection points of the trapezoidal membership function.

[0039] When b = c, it degenerates into a triangle membership function; when a = b, the left side is the perpendicular side; when c = d, the right side is the perpendicular side.

[0040] Then, a fuzzy rule base is established based on historical pressure resistance data and expert experience. For example, "If D is high and H is high, then Vmax is low and..." Low".

[0041] Since each of the four input variables is divided into three fuzzy subsets, a maximum of three fuzzy subsets can be formed. 4 =81 rules. The following are 20 examples of the main typical rules; the complete rule base contains 81 rules.

[0042] The remaining fuzzy rules are automatically generated based on all possible combinations of each risk indicator, following the principle of condition coverage: that is, for all key indicator ranges (low, medium, high) and their interactions, the system pre-determines "if indicator A is a certain range, indicator B is a certain range, ... then the risk level is corresponding" through enumeration combinations, ensuring that any actual detection result can be matched with the corresponding inference path in the rule base. Parameters are adjusted using expert experience and historical data to achieve full coverage and rationality of the rules.

[0043] An example of a fuzzy rule base is as follows: Serial Number Insulation loss factor D Temperature T Humidity H Insulation resistance R <![CDATA[Initial voltage V0]]> boost rate α Target withstand voltage Vmax 1 high any high any Low Low Low 2 high high any Low Low Very low Low 3 Low Low Low high high high high 4 Low middle Low high high high high 5 middle middle middle middle middle middle middle 6 high Low Low high middle middle middle 7 any any high Low Low Low Very low 8 Low any any high high high high 9 middle high high middle Low Low Low 10 high middle middle Low Low Very low Low 11 middle Low Low middle middle high high 12 middle high Low Low Low middle middle 13 high high high high Very low Very low Very low 14 Low high high Low Low Low Low 15 Low Low middle high high high high 16 middle Low high high middle middle middle 17 high Low middle middle middle Low Low 18 high middle high high Low Very low Low 19 Low middle high high middle Low Low 20 middle middle high Low Low Low Very low During real-time operation, the degree of condition satisfaction is calculated through membership degree and the corresponding fuzzy rules are activated. Then, the outputs of all activated rules are normalized and integrated using either a weighted average method or a maximum membership degree method. Finally, the optimal parameter settings for the withstand voltage test are given based on the inference results, including the starting voltage. Boost rate The outputs of the target withstand voltage value Vmax are expressed by the following general formulas: ; In the formula , , This is a nonlinear mapping function generated by a fuzzy inference system. Through the aforementioned fuzzy logic inference and dynamic parameter setting, the mismatch problem caused by the traditional fixed curve method is effectively avoided, improving the personalization and safety of the pressure test.

[0044] S2. During the boost and withstand voltage tests, a high-precision sensor array with multiple channels is used to collect the current and voltage signals of each channel in real time. The sampling gain of each channel is dynamically adjusted based on the active identification-gain matching algorithm to enhance the perception of weak abnormal signals. Preferably, in this embodiment, during the boost and withstand voltage testing phases, a multi-channel high-precision sensor array is used to collect the current at each measuring point in real time. With voltage In the S2 signal, the active identification-gain matching algorithm first calculates the instantaneous amplitude of each channel. in, and The first Instantaneous current and voltage of the channel, Channel number, Set the time; then set the signal gain for each acquisition channel. ; And estimate the rate of change: ; in The sampling time is from 1 μs to 1 ms, depending on the system sampling rate; then... Compared with historical average and standard deviation In comparison, when hour( The control factor will affect the channel gain. Increase a dynamic weight .

[0045] The dynamic weight k is derived from the gain adjustment strategy. This weight factor k is set to a constant greater than 1, and its specific value is determined based on system requirements, practical experience, or experiments. It is usually set and dynamically adjusted during system initialization or adaptive algorithms. If the focus of anomaly detection is to quickly capture small-amplitude anomalous signals, a higher k value can be selected.

[0046] For the rate of change Comparison with historical averages This represents the rate of change of the current amplitude (i.e., the amplitude increment per unit time), while and These represent the mean and standard deviation of the rate of change of the channel over a past period, respectively. Since all β values ​​are expressed using... The calculations are complete, and the units are identical (amplitude units / time, amperes / second, or volts / second). Therefore, they are comparable in both numerical and physical sense. This is achieved by weighting the standard deviation with the historical mean (…). This can effectively determine whether the current rate of change is an anomalous mutation: if the current rate of change is... If the fluctuations significantly exceed the normal range of the past, it is determined that there may be an anomaly.

[0047] Historical average in this embodiment and standard deviation Using a sliding window W for estimation, W = 0.5–2s.

[0048] Preferably, in this embodiment, the active identification-gain matching algorithm also incorporates real-time signal-to-noise ratio: ; in To obtain the instantaneous noise power based on the noise amplitude from a sliding window, the sliding window variance or spectral estimation (FFT) is used. and take ;when Below the preset threshold Adjusted proportionally If after adjustment If the set limit is exceeded, a rollback will be triggered.

[0049] Specifically, in this embodiment, to prevent division by zero, the real-time signal-to-noise ratio is defined as: ; in, .

[0050] S3. When leakage current or abnormal signal is detected, multi-scale wavelet transform is performed on the abnormal interval waveform and the amplitude, statistical features and energy at each scale are extracted. The abnormal waveform is clustered using a self-learning deep clustering algorithm, and the abnormal waveform identified by clustering is combined with the sensor position information of the corresponding acquisition channel to achieve abnormal location.

[0051] Leakage current or abnormal signals originate from S2. During the boost and withstand voltage tests, the active identification-gain matching algorithm is used to analyze the rate of change of current and voltage signals collected from each channel. Combined with historical mean and standard deviation as well as real-time signal-to-noise ratio, anomaly detection is performed. When the rate of change of a certain channel signal is detected to be significantly beyond the normal fluctuation range or the signal-to-noise ratio is lower than the preset threshold, it is determined that there is leakage current or abnormal signal in that interval, and it is marked as an abnormal interval for S3 to carry out subsequent feature extraction and cluster analysis.

[0052] Preferably, in this embodiment, the wavelet basis function in S3 is either Daubechies or Symlet, which decomposes the waveform of the abnormal interval into n scales to obtain the wavelet coefficients at each scale. , where j represents the scale and k is the time-domain sampling point; Then calculate the energy at each scale. ; Forming energy eigenvectors Simultaneously, extract the peak amplitude of the original waveform. And statistical characteristics, ultimately the data feature vector of each abnormal waveform segment is obtained. Input into the self-learning deep clustering module.

[0053] Preferably, in this embodiment, the self-learning deep clustering module uses a Gaussian mixture model for parameter estimation, and the number of clusters K is determined by the Bayesian information criterion, the Akaike information criterion, or an adaptive split-merge strategy. The clustering results are output in probabilistic form. ,in It is the probability density function of a Gaussian distribution.

[0054] Historical labeled samples or semi-supervised outlier samples are used as the initial training set for the model to form a preliminary category distribution and parameter estimates. During the online detection phase, new samples collected in real time are periodically merged with the current model in a mini-batch manner. The model parameters are quickly updated using the incremental EM (Expectation-Maximization) algorithm to ensure that the clustering model can adaptively reflect the constantly changing signal characteristics in the field.

[0055] Furthermore, to address the diversity of abnormal patterns in actual working conditions, a threshold-based splitting and merging strategy is introduced: when the internal sample variance of a certain class continues to increase or the cluster confidence drops below a set threshold, the splitting of that class is automatically triggered. Similarly, for highly similar and convergent classes, the model structure is optimized through merging operations to improve clustering accuracy and efficiency. Regarding model complexity selection, the optimal number of clusters is automatically determined using either BIC (Bayesian Information Criterion) or AIC (Akaike Information Criterion). The cluster structure is periodically evaluated and updated after each batch of new samples is collected. The typical window size is 500 to 2000 abnormal waveform samples, adjusted according to the amount of on-site data and computing resources. The update cycle is set to once every 10 to 30 minutes to balance computational load and adaptive capability.

[0056] As a typical self-learning implementation, new batches of abnormal samples can be periodically (every 30 minutes or every 1000 samples) integrated into the existing clustering model, and the parameters can be updated using a mini-batch EM algorithm. Simultaneously, the category confidence or cluster distribution entropy of each cluster is monitored. When it is found that some new samples have low probability of clustering with existing categories (the highest probability is below 0.5), a manual annotation feedback mechanism is triggered. These low-confidence samples are then pushed to experts for manual review before being added to the training set, achieving human-machine collaborative model self-evolution. By combining incremental learning with artificial intelligence, the robustness, universality, and adaptability of discharge anomaly detection can be significantly improved.

[0057] In this embodiment, the Gaussian Mixture Model (GMM) is initially trained using manually labeled abnormal / normal samples, and the Expectation-Maximization (EM) algorithm is used to estimate the parameters. .

[0058] Mini-batch EM is used, and parameter updates are triggered whenever the amount of new data reaches the updated sample size; if BIC indicates model degradation, split / merge is performed and manual review is notified.

[0059] Preferably, the clustering labels are associated with the spatial coordinate information of the sensor and back-labeled to the structure of the tool being tested, thereby realizing the spatial mapping and positioning of suspected abnormal areas.

[0060] In this embodiment, the system is assumed to have a total of M sensors, the first... The spatial installation coordinates of each sensor are represented by a three-dimensional vector. When the nth abnormal waveform segment is detected, the sensor from which it originates is recorded, denoted as the nth sensor. A sensor, the installation coordinates of which are... Simultaneously, extract the feature vector of the abnormal segment. ; The self-learning deep clustering module divides all abnormal waveforms into K clusters (each cluster represents a typical discharge / leakage mode); for the Kth cluster, the system calculates the estimated spatial location of the suspected abnormal part represented by that cluster. The formula is as follows: ; in, is the estimated spatial location of the Kth anomaly cluster; is the number of the anomaly waveform segment; It is the set of all abnormal waveform segments belonging to the Kth cluster; For the first The coordinates of the sensor; Let be the posterior probability that the m-th abnormal waveform belongs to the K-th cluster. The value is directly output by the Gaussian mixture model or deep clustering, ranging from 0 to 1, with a larger value indicating a stronger resemblance to that class. To allow operations and maintenance personnel to intuitively assess the reliability of the location, the location reliability of the cluster is also calculated. The formula is as follows: ; in This represents the total number of abnormal samples used for positioning during this pressure resistance test; The closer the value is to 1, the more reliable the positioning; generally The positioning is considered valid when the value is ≥ 0.6.

[0061] Using the above methods, the clustering process not only relies on amplitude information, but also integrates deep feature dimensions such as multi-scale energy, which enables effective identification of complex discharge or leakage modes and precise positioning of specific abnormal areas within the insulation structure, greatly improving the universality of detection and the anomaly detection rate under variable field conditions.

[0062] S4. The withstand voltage process adopts a multi-cycle progressive approach. After each cycle, the historical and current cycle features are integrated, and Bayesian discrimination is used to calculate the posterior probability of failure. The discrimination threshold is dynamically corrected based on gradient reduction calibration to optimize sensitivity and false alarm rate. Preferably, in this embodiment, in S4, the first... Feature vectors are collected over each withstand voltage cycle. Calculate posterior probability based on Bayesian discriminant method ; when Current periodic warning threshold The system provides timely warnings of potential failures; to achieve threshold adaptation, a discriminative loss function L is defined, and based on the discriminative error... ; in, For actual feedback labels, usually Indicates security / failure; Update the threshold using gradient descent: ; Where L is the preset discriminant loss function, γ is the learning rate, and if the on-site label is not available, a delayed label or a pseudo-label is used.

[0063] To ensure the prior probability of Bayesian discrimination and Prior information that accurately reflects the normal and failure states of insulating tools is generally recommended to be initialized according to the statistical frequency of historical samples. For example, if the proportion of "safe" states in historical test data is q, then it can be set to... , If historical frequency information is insufficient, it can be temporarily used. As an unbiased default value, it is dynamically updated as data accumulates. Discriminant loss function. The specific expression can be expressed using the binary cross-entropy loss, i.e. Its criterion threshold gradient Based on the relationship between posterior probability and threshold in practical applications, analytical or numerical approximations are provided. Learning rate. The value is set to 0.05, and dynamically fine-tuned based on the actual convergence speed and false alarm rate to ensure that the threshold adjustment can fully respond to new feedback information without causing system instability due to excessive adjustment. Through the above settings and parameter selection, adaptive and progressive optimization of the withstand voltage test threshold can be achieved, effectively improving the accuracy of early identification of insulation risks.

[0064] In this embodiment, the pressure test is conducted using a multi-cycle progressive pressure increase method, with each pressure increase-pressure withstand cycle... (l represents the cycle number) After the cycle ends, the system will collect the detection features of this cycle. This includes the measured average leakage current. Peak Partial discharge energy Frequency of abnormal triggers And so on, while integrating all feature sequences from historical cycles. This serves as input for subsequent evaluations.

[0065] First, a Bayesian discriminant method is used to statistically model the feature vectors of each period, for example, defining the feature probability density under the two categories of normal and abnormal insulation. and ,in and These represent the "safe" and "failed" states of insulation, respectively. The posterior probability for each cycle can be obtained using Bayes' theorem, allowing for real-time safety margin determination. ; If the posterior probability is greater than the empirical safety threshold ( If the current cycle failure warning line is set, it indicates a significant risk of failure for the insulating tools and equipment.

[0066] To further optimize and adapt the discrimination process, the comprehensive evaluation algorithm introduces a gradient reduction calibration strategy, which is based on the evaluation error in each cycle. (in For actual feedback labels, usually (Indicating safety / failure) The criterion threshold is dynamically adjusted using a gradient decrease method. , For learning rate, To determine the loss function; S5. After all testing and analysis are completed, the test results are standardized, weighted, and graded based on the comprehensive fuzzy reasoning scoring model to generate safe usage suggestions and subsequent maintenance or re-inspection plans.

[0067] As a preferred embodiment, in this embodiment, the comprehensive fuzzy inference scoring model in S5 standardizes, weights, and maps the detected indicators to risk levels, which include "safe", "warning", "critical" and "high risk", and automatically generates corresponding handling suggestions based on the level.

[0068] After all detection and analysis processes are completed, the system automatically summarizes the data and judgment results generated from each step, including signal acquisition, anomaly identification, discharge pattern clustering, and periodic assessment, to construct a complete and time-series-clear dataset. Based on this dataset, the system adopts a comprehensive fuzzy inference scoring model. First, it standardizes and assigns multi-dimensional weights to the various detected indicators (such as withstand voltage cycle damage parameters, discharge risk classification inferred from clustering, and failure probability determined by Bayesian discrimination), mapping the original output to a comprehensive risk score with the overall safety of the tool as the core.

[0069] First, all core indicators, such as withstand voltage cycle damage parameters, discharge risk classification, and failure probability, are standardized and mapped to a standard score range of 0–1. Each standardized indicator is then assigned a different weight based on its importance to safety, and a final weighted average is calculated to obtain a total risk score.

[0070] The comprehensive risk score uses a 0-100 scale for easy interpretation and comparison. To facilitate decision-making and operational implementation by maintenance personnel, the system divides the risk score into four levels: 0-25 points correspond to "Safe," indicating that the current equipment is in good condition and can be used normally, with a recommendation for scheduled re-inspection; 26-50 points are "Warning," indicating potential risks, with recommendations for increased daily monitoring or re-inspection in the short term; 51-75 points are classified as "Critical," indicating that the equipment may be in a sub-healthy state or on the verge of failure, requiring restricted use and priority for maintenance; 76-100 points are deemed "High Risk," requiring immediate shutdown and specialized maintenance or replacement. Different levels automatically match corresponding maintenance and re-inspection recommendations.

[0071] See the table below for specific scoring and level examples: Comprehensive risk score range Risk level Recommended measures 0–25 Safety Normal use, regular check-ups 25–50 Warning Strengthen monitoring and conduct short-term re-inspections. 50–75 critical Restricted use, priority maintenance 75–100 High risk Immediate shutdown and special maintenance required. This scoring and grading method not only achieves unified quantification and intuitive expression of detection results across different dimensions, but also allows operations and maintenance personnel to understand the current security status, major risks, and corresponding management recommendations at a glance, thereby facilitating scientific decision-making and rapid response.

[0072] The model defines levels such as "safe," "warning," "critical," and "high-risk" within specific numerical ranges by setting fuzzy membership functions. It also integrates the correlations between abnormal indicators, recommending immediate shutdown and specialized maintenance for "high-risk" reports, and enhanced monitoring and short-term re-inspections for "warning" reports. Finally, all analyses and recommendations are formatted and compiled into a standardized, traceable, and interpretable report. This report details the core data and anomaly findings from each round of withstand voltage testing. This standardized output significantly improves information access and decision-making transparency for maintenance personnel, providing a strong basis for scientifically assessing equipment status and optimizing maintenance and management plans.

[0073] Furthermore, to verify the effectiveness and safety of the proposed dynamic adaptive fuzzy control method for power frequency withstand voltage of insulating tools, this embodiment conducts simulation experiments to analyze its adaptability under different historical data of tools and variable environmental parameters, as well as the traceability and interpretability of the output report.

[0074] 1. Simulation software: MATLAB / Simulink, or Python (Scikit-fuzzy, Numpy, Matplotlib, etc.).

[0075] 2. Historical power frequency withstand voltage dataset: Includes test voltages, loss factor D, environmental records, etc. of multiple batches of insulating tools over the years.

[0076] 3. Real-time environmental parameter acquisition module (simulation): generates changing T (temperature), H (humidity), and R (insulation resistance).

[0077] 4. Fuzzy control system modeling tools: such as MATLAB FuzzyLogicToolbox or a custom Python system.

[0078] Simulation experiment implementation steps: (1) Preparation of historical data and environmental parameters Data collection and analysis: Collect historical power frequency withstand voltage test data of typical insulating tools, including starting voltage, voltage rise rate, target withstand voltage, and loss factor D.

[0079] Environmental parameter generation: According to experimental requirements, set the variation range of temperature T, humidity H, and insulation resistance R, and input the simulation system using a generating function or measured data.

[0080] (2) Modeling of adaptive fuzzy control system Input variable definitions: D, T, H, R.

[0081] Each variable is divided into multiple levels of membership functions (e.g., "low", "medium", "high"). Assuming the main inputs are insulation loss factor D, ambient temperature T, ambient humidity H, and insulation resistance R, each input variable is divided into three intervals: "low," "medium," and "high" using a fuzzy membership function, with a trapezoidal function as the membership function, as follows: Each fuzzy subset uses a four-parameter trapezoidal function as its membership function to ensure appropriate overlap between adjacent subset intervals, thereby guaranteeing full coverage and smooth transition of the input space. The parameters [a, b, c, d] of the trapezoidal membership function are defined as follows, taking insulation loss factor D, temperature T, humidity H, and insulation resistance R as factors: D (insulation loss factor): Low: [0, 0, 0.01, 0.015]; In the middle: [0.01, 0.015, 0.02, 0.03]; High: [0.02, 0.03, 0.05, 0.05]; T (Temperature, unit: °C): Low: [-10, -10, 5, 18]; Middle: [10, 20, 30, 40]; Height: [35, 45, 60, 60]; H (humidity, unit: %) Low: [0, 0, 30, 50]; Chinese: [35, 50, 70, 85]; Height: [75, 90, 100, 100]; R (insulation resistance, unit: MΩ): Low: [0, 0, 20, 80]; Chinese: [50, 150, 300, 500]; High: [400, 800, 5000, 5000].

[0082] Note: The parameter order is [a, b, c, d], which corresponds to the four inflection points of the trapezoidal membership function.

[0083] When b = c, it degenerates into a triangle membership function; when a = b, the left side is the perpendicular side; when c = d, the right side is the perpendicular side.

[0084] Establishing a fuzzy rule base: Since each of the four input variables is divided into three fuzzy subsets, a maximum of three fuzzy subsets can be formed. 4 =81 rules. The following are 20 examples of the main typical rules; the complete rule base contains 81 rules.

[0085] By combining historical data and domain knowledge, the rule base is established as follows: Serial Number Insulation loss factor D Temperature T Humidity H Insulation resistance R <![CDATA[Initial voltage V0]]> boost rate α Target withstand voltage Vmax 1 high any high any Low Low Low 2 high high any Low Low Very low Low 3 Low Low Low high high high high 4 Low middle Low high high high high 5 middle middle middle middle middle middle middle 6 high Low Low high middle middle middle 7 any any high Low Low Low Very low 8 Low any any high high high high 9 middle high high middle Low Low Low 10 high middle middle Low Low Very low Low 11 middle Low Low middle middle high high 12 middle high Low Low Low middle middle 13 high high high high Very low Very low Very low 14 Low high high Low Low Low Low 15 Low Low middle high high high high 16 middle Low high high middle middle middle 17 high Low middle middle middle Low Low 18 high middle high high Low Very low Low 19 Low middle high high middle Low Low 20 middle middle high Low Low Low Very low Inference mechanism implementation: For all rules, calculate the membership degree under the current input and activate the corresponding rule. The output is then fused using either a weighted average method or a maximum membership degree method.

[0086] (3) Adaptive calculation and simulation of pressure test parameters Simulation process: 1. Read a set of historical data and current environment parameters.

[0087] 2. Input the fuzzy inference system and output the optimal starting voltage, boost rate, and target withstand voltage (Vmax).

[0088] 3. Generate and plot custom withstand voltage boost curves (non-linear, personalized).

[0089] 4. Apply this curve to the object under test to simulate the voltage boosting process and monitor the insulation response of the tool in real time (such as leakage current, breakdown time, etc.).

[0090] (4) Anomaly detection and classification Abnormal indicator extraction: Analyze abnormal leakage current, sudden changes in insulation medium, etc. during the simulation process.

[0091] Classification recommendations: If high-risk indicators are detected, automatic recommendation is given to shut down the system for maintenance; if a warning is issued, it is recommended to increase routine inspections or conduct a re-inspection within a short period of time.

[0092] (5) Output standardized, traceable reports Collect the following information and output a report: Pressure resistance curve, pressure increase rate variation, actual pressure limit Anomalies detected in each round of testing and their corresponding environmental parameters System-recommended maintenance / re-inspection suggestions Full process parameters and data traceability details Results Analysis and Validation Based on the above analysis results, a revision suggestion is proposed for the power frequency test voltage of insulated tools for live working. The rationality of the revised test voltage value is further verified by power frequency discharge characteristic test, power frequency withstand voltage verification test, electrical performance test of insulated tool materials, and live working safety test.

[0093] The following power frequency discharge characteristic test aims to study the discharge characteristics of different types of insulating tools (including natural fiber insulating ropes, synthetic fiber insulating ropes, and glass fiber reinforced epoxy insulating rods) under typical power frequency voltage conditions, obtain critical withstand voltage data for different effective lengths, and analyze the power frequency discharge saturation characteristics at high voltage levels. This will provide basic data and theoretical support for rationally determining the safety margin of power frequency withstand voltage tests for insulating tools and optimizing test parameter settings.

[0094] To further analyze the safety margin range of the power frequency test voltage for insulating tools, it is necessary to determine the critical withstand voltage level of the insulating tools at the corresponding length. Power frequency discharge tests were conducted on insulating ropes and insulating rods.

[0095] Uniform voltage ramp-up discharge tests were conducted on ultra-high voltage (UHV) live-line working natural fiber insulated ropes (silk ropes), synthetic fiber insulated ropes (polyester ropes), and insulated rods (with metal ends) under power frequency voltage. Discharge voltage values ​​were recorded for different effective insulation lengths (5m, 6m, 7m; average value was taken after three uniform voltage ramp-up discharges). An improved test electrode using an 8-split simulated conductor was selected as the high-voltage electrode. The conductor height above ground was 20m, and the distance between the grounding end of the test specimen and the ground was greater than 1m. The overall test setup is as follows: Figure 2 As shown, the sample structure is as follows Figure 3 As shown.

[0096] The insulating ropes are moisture-proof silk rope and moisture-proof polyester rope, and the insulating rods are glass fiber reinforced epoxy foam filled rods, each 2m long, connected by 0.1m metal joints. All properties of the sample meet the requirements of the relevant standards, and the tests were passed.

[0097] Analysis of test results: Considering the saturation of power frequency discharge characteristics over long distances, the test data of power frequency discharge characteristics of insulating ropes and insulating rods were fitted with a logarithmic function and compared with the 330~1000kV preventive test voltage curve. Since the power frequency discharge characteristics of insulating materials tend to saturate over long distances, and the increase in test distance of insulating tools is less than the increase in voltage level, the range of safe margin values ​​for test voltage of insulating tools under high voltage levels is reduced.

[0098] According to the discharge curve fitting results, the power frequency discharge voltages of natural fiber insulating rope, synthetic fiber insulating rope and insulating rod with uniform voltage rise are 1298.5kV, 1312.6kV and 1171.9kV respectively under a test length of 6.3m. The current preventive test voltage (1150kV) is only 1.8% lower than the discharge voltage of insulating rod (1171.9kV). Considering the influence of test environment, discharge randomness and other factors, the test sample is very likely to flashover in the withstand voltage test.

[0099] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Substitutions may include replacements of some structures, devices, or method steps, or may be complete technical solutions. Equivalent substitutions or modifications made to the technical solutions and inventive concepts of the present invention should all be covered within the scope of protection of the present invention.

Claims

1. An insulation tool detection power frequency withstand voltage test method, characterized by, Comprising the following steps: S1, based on the historical power frequency withstand voltage data of the measured insulating tool, environmental parameters and insulation loss factor, the starting voltage, voltage rising rate and target withstand voltage value of the withstand voltage test are set by using an adaptive fuzzy prediction algorithm; S2, during the voltage rising and withstand voltage test process, a high-precision sensor array is arranged in multiple channels synchronously, real-time acquisition of current and voltage signals of each channel is performed, and the sampling gain of each channel is dynamically adjusted based on an active identification-gain matching algorithm to enhance the perception of weak abnormal signals; S3, when the leakage current or abnormal signal is detected, multi-scale wavelet transform is performed on the abnormal interval waveform, the amplitude, statistical characteristics and energy of each scale are extracted, self-learning deep clustering algorithm is used for clustering of the abnormal waveform, the abnormal waveform distinguished by clustering is combined with the sensor position information of the corresponding acquisition channel, and abnormal positioning is realized; S4, the withstand voltage process adopts a multi-cycle progressive mode, the historical and current cycle characteristics are fused after each cycle ends, the failure posterior probability is calculated by using Bayesian discriminant, and the discriminant threshold is dynamically modified based on gradient reduction calibration to optimize the sensitivity and false alarm rate; S5, after all the detection and analysis are completed, the detection results are standardized, weighted and graded based on a comprehensive fuzzy reasoning scoring model, and a safe use suggestion and a subsequent maintenance or re-inspection scheme are generated.

2. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 1, wherein The adaptive fuzzy prediction algorithm in S1 takes as input the vector D is the insulation loss factor obtained by standard power frequency loss measurement method; T is the ambient temperature, H is the ambient humidity, R is the insulation medium resistance, and is aligned with historical data by timestamp; by establishing several fuzzy subsets for each input variable, the starting voltage , the voltage rise rate and the target withstand voltage value V max are obtained by weighted average or maximum membership degree after fuzzy reasoning.

3. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 2, wherein The fuzzy subset is represented by a three-section or four-section membership function, and the membership function is in the form of trapezoidal or triangular.

4. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 1, wherein The S2 active recognition-gain matching algorithm first calculates the instantaneous amplitude of each channel ； wherein, with are the first the instantaneous current and voltage of the channel, is the channel number, is the time; The change rate is estimated: ； wherein ; then the is compared to a historical mean and standard deviation and the channel gain is increased by a dynamic weight when .

5. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 4, wherein The active identification-gain matching algorithm also introduces real-time signal-to-noise ratio: ； wherein is the current channel noise amplitude estimate, obtained using a sliding window variance estimate or spectral noise estimate; when is below a preset threshold is scaled up is scaled up, if is above a set upper limit, a fallback is triggered.

6. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 1, wherein The wavelet base function in the S3 adopts Daubechies or Symlet, and waveforms of an abnormal interval are decomposed to n scales to obtain wavelet coefficients on each scale wherein j represents a scale, and k is a time domain sampling point. Then the energy of each scale is calculated ; Forming energy feature vector ; while extracting amplitude peak value of original waveform and statistical feature, finally input data feature vector of each segment of abnormal waveform to self-learning deep clustering module. ​ 7. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 6, wherein The self-learning deep clustering module adopts a Gaussian mixture model for parameter estimation, the cluster number K is determined through a Bayesian information criterion, Akaike information criterion or adaptive splitting and merging strategy, and the clustering result is output in a probability form wherein is a Gaussian distribution probability density function.

8. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 7, wherein The clustering label is associated with the spatial coordinate information of the sensor and is marked back to the measured tool structure, so as to realize spatial mapping and positioning of the suspected abnormal area.

9. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 1, wherein The characteristic vector is collected in the S4 in the first voltage cycle The posterior probability is calculated based on the Bayesian discrimination ； When > current cycle alarm threshold prompt failure risk; to achieve threshold adaptation, define a discriminant loss function L, and based on the discriminant error ； wherein is the actual feedback tag; The threshold value is updated by using gradient descent method: ； Wherein L is a preset discriminant loss function, and γ is a learning rate.

10. The power frequency withstand voltage test method for detecting an insulated tool according to Claim 1, wherein The comprehensive fuzzy reasoning scoring model in S5 standardizes, weights and maps each type of index detected to a risk level, the risk level includes "safe", "warning", "critical" and "high risk", and corresponding disposal suggestions are automatically generated according to the level.

Images