A method and system for testing the insulation performance of a distribution box

Abstract

This invention belongs to the field of testing technology, specifically relating to a method and system for testing the insulation performance of a distribution box. The method includes: acquiring multi-dimensional operating data of the distribution box, constructing a multi-dimensional state vector, and using a clustering algorithm to screen stable operating cycles; extracting residual current harmonic characteristics within the stable cycles, and constructing an aging Gaussian mixture probability model for the reference insulation; constructing a tensor product B-spline environmental compensation model by combining temperature and humidity data; using the environmental compensation model to process the residual current characteristics to be measured to obtain an adjusted feature vector, calculating the minimum squared Mahalanobis distance between the adjusted feature vector and the components of the probability model, and determining the insulation performance status. This invention achieves high-precision continuous online detection of the insulation status of power distribution equipment by finely evaluating the reference state and offsetting environmental fluctuation interference.

Description

Technical Field

[0001] This invention relates to the field of testing technology. More specifically, this invention relates to a method and system for testing the insulation performance of a distribution box. Background Technology

[0002] The safe and stable operation of distribution boxes is crucial for ensuring the reliability of power supply for various industrial and residential applications. Under the combined effects of long-term high-voltage, high-current operating loads, the insulation materials inside the distribution box inevitably age, become damp, or are damaged, leading to a gradual decline in insulation performance. This deterioration not only easily causes local electrical faults such as leakage and short circuits, but in severe cases, it can induce arcing fires or even large-scale power outages. Manual inspection is not only costly in terms of manpower and resources, severely impacting power supply continuity, but it also cannot achieve real-time, continuous online monitoring of the insulation condition, failing to detect potential problems in the early stages of insulation deterioration, resulting in a significant delay.

[0003] Currently, existing technologies combine sensor monitoring with data analysis algorithms to preliminarily assess the insulation status online based on real-time collected operating parameters of power distribution equipment. Chinese patent document CN116381424A discloses a method for detecting abnormal insulation performance of shore power boxes based on multi-source data fusion. This method utilizes multiple sensors to collect multi-source one-dimensional time-series signals of internal temperature, humidity, main insulation voltage of the neutral grounding system, and short-circuit electromotive force within the shore power box. The aligned multi-source data is fused to obtain a multivariate two-dimensional recursive graph. This multivariate two-dimensional recursive graph is used to train a VAE-SVM neural network model and determine whether abnormal insulation performance has occurred. Based on the above technical means, abnormal features from multiple sensor dimensions can be extracted, thereby achieving online detection of the insulation performance of electrical equipment.

[0004] However, the aforementioned technical solutions can, to some extent, achieve automatic extraction and anomaly detection of equipment insulation performance. However, existing technologies, when constructing health benchmark models, cannot screen out truly stable operating cycles under complex grid load fluctuations and multi-dimensional state interweaving, leading to contamination and bias in the benchmark data. Health benchmark samples are easily contaminated by abnormal operating conditions. Furthermore, environmental factors such as temperature and humidity can alter the dielectric constant and surface conductivity of insulation materials, causing non-faulty, drastic fluctuations in residual current. Existing technologies lack sufficient strategies to address the susceptibility of residual current characteristics to environmental fluctuations, making it difficult to eliminate environmental interference and improve assessment accuracy. Moreover, existing technologies do not extract and integrate the higher harmonic characteristics of residual current, nor do they scientifically represent the essential deviation between real-time characteristics and the benchmark model from a high-dimensional probability distribution perspective. This makes the detection model extremely insensitive to minor early insulation degradation, and single threshold judgments are insufficient to accurately identify early insulation degradation, frequently resulting in false alarms and missed alarms, failing to meet the high-reliability requirements of power distribution operation and maintenance. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as the reliance on offline testing for insulation performance testing of distribution boxes, the difficulty in achieving continuous online monitoring, the susceptibility of health benchmark samples to contamination by abnormal operating conditions, the susceptibility of residual current characteristics to interference from temperature and humidity fluctuations, the difficulty in accurately identifying early insulation degradation with a single threshold determination, and the high false alarm and false alarm rates, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for detecting the insulation performance of a distribution box, comprising: S1, acquiring operating current, residual current, bus voltage, box temperature, and humidity data within a preset time period; extracting current wavelet features of the operating current, constructing a multi-dimensional state vector by combining bus voltage stability and box temperature change rate, and selecting stable operating cycles representing healthy states through an unsupervised clustering algorithm; S2, extracting residual current within each stable operating cycle representing healthy states, calculating the root mean square value of the residual current and the normalized amplitude of multiple harmonic components as residual current features, and constructing an aging Gaussian mixture probability model representing the baseline insulation state; and based on the stable operating cycles representing healthy states within the box... Temperature and humidity data, along with corresponding residual current characteristics, are used to construct a tensor product B-spline function model for environmental compensation. S3: Extract the residual current feature vector for the current test period. Calculate the environmental compensation vector using the tensor product B-spline function model based on the current temperature and humidity data within the box. Subtract the environmental compensation vector from the residual current feature vector and add the overall mean vector of the aging Gaussian mixture probability model to obtain the adjusted feature vector. Using the covariance matrix of each component in the aging Gaussian mixture probability model, calculate the square of the Mahalanobis distance between the adjusted feature vector and the corresponding mean vector. Select the minimum value as the insulation state deviation index, compare it with the preset chi-square distribution confidence threshold, and determine the insulation performance status of the distribution box.

[0007] This invention acquires multi-dimensional operating data of power distribution equipment, combines high-frequency features and clustering algorithms to remove contaminated samples, extracts harmonics to construct an aging Gaussian model to refine the baseline state assessment, combines a tensor product B-spline model to offset environmental fluctuation interference, and uses the minimum Mahalanobis distance to measure deviation index to determine the state; it realizes automatic extraction and real-time continuous online monitoring of the insulation state of power distribution equipment, effectively eliminates the contamination of the baseline sample by abnormal operating conditions such as complex power grid load fluctuations, eliminates the interference of temperature and humidity environmental fluctuations on residual current characteristics, accurately identifies early minor insulation degradation, and effectively reduces false alarm rate and missed alarm rate.

[0008] Preferably, the construction of the multidimensional state vector and the selection of stable operating cycles representing a healthy state through an unsupervised clustering algorithm includes: performing discrete wavelet decomposition on the operating current, calculating the variance of the detailed coefficient sequence corresponding to a preset high-frequency analysis band as the current wavelet feature; calculating the residual between the bus voltage sequence and a preset standard sine wave sequence, and obtaining the standard deviation of the residual sequence, dividing the standard deviation by the rated voltage reference value to obtain the voltage fluctuation rate, and obtaining the bus voltage stability based on the reciprocal of the voltage fluctuation rate; calculating the temperature difference within the tank within a preset time window, dividing the temperature difference by the duration of the time window to obtain the temperature change rate within the tank; normalizing the current wavelet feature, bus voltage stability, and temperature change rate within the tank and concatenating them into a multidimensional state vector, clustering them using a clustering algorithm and selecting target clusters, and selecting the operating cycles contained in the target clusters as stable operating cycles.

[0009] Preferably, the construction of the aging Gaussian mixture probability model representing the reference insulation state includes: performing a Fourier transform on the residual current waveform during a stable operating cycle representing the healthy state to extract the fundamental amplitude and the amplitudes of the third, fifth, and seventh harmonics; dividing the amplitudes of the third, fifth, and seventh harmonics by the sum of the fundamental amplitude and a preset minimum positive number to obtain the corresponding normalized amplitudes, and combining them with the root mean square value of the residual current to form a four-dimensional reference feature vector set; initializing the mixture weights, mean vector, and covariance matrix of the aging Gaussian mixture probability model, and establishing a log-likelihood function; using the expectation-maximization algorithm to iteratively train the four-dimensional reference feature vector set, stopping the iteration when the change in the log-likelihood function between two adjacent iterations is less than a preset convergence threshold, and outputting the optimized aging Gaussian mixture probability model.

[0010] This invention overcomes the limitation of using a single time-domain feature to characterize the insulation state in a one-sided manner. It can utilize higher harmonic components to mine nonlinear impedance distortion information during early insulation degradation, thereby accurately constructing a baseline distribution model of the healthy state.

[0011] Preferably, the construction of the tensor product B-spline function model for environmental compensation includes: extracting and equally dividing the extreme temperature ranges within the chamber from the historical dataset, setting a first node vector at the boundary, and constructing a one-dimensional B-spline basis function based on temperature; extracting and equally dividing the extreme humidity ranges from the historical dataset, setting a second node vector at the boundary, and constructing a one-dimensional B-spline basis function based on humidity; performing a tensor product operation on the one-dimensional B-spline basis function based on temperature and the one-dimensional B-spline basis function based on humidity to generate a two-dimensional tensor product B-spline basis function matrix; and using the temperature and humidity data within the chamber as input and the corresponding four-dimensional baseline feature vector as the desired output, solving the control vertex weight coefficient matrix to obtain the tensor product B-spline function model.

[0012] This invention avoids fitting oscillations caused by abrupt changes in boundary conditions and fully characterizes the nonlinear coupling relationship between temperature and humidity on insulation performance using tensor product structure, thereby improving the model's feature compensation capability for complex temperature and humidity environments.

[0013] Preferably, the step of calculating the environmental compensation vector and determining the insulation performance status of the distribution box using the tensor product B-spline function model includes: substituting the temperature and humidity data inside the box for the test period into the tensor product B-spline function model to output the environmental compensation vector; multiplying the mean vectors of each Gaussian component in the aging Gaussian mixture probability model by their corresponding mixture weights and summing them to obtain the overall mean vector; subtracting the environmental compensation vector from the residual current test feature vector and adding the overall mean vector to obtain the adjusted feature vector; using the covariance matrix corresponding to each Gaussian component in the aging Gaussian mixture probability model, calculating the squared Mahalanobis distance between the adjusted feature vector and the corresponding mean vector, and selecting the minimum value as the insulation state deviation index; when the insulation state deviation index is greater than the confidence threshold, determining that the insulation performance is in an abnormal degradation state; otherwise, determining it as a healthy and stable state.

[0014] This invention significantly improves the sensitivity of identifying early minor insulation defects under multidimensional probability distributions by offsetting the characteristic baseline drift caused by drastic environmental fluctuations and eliminating the correlation redundancy between features using Mahalanobis distance.

[0015] Preferably, after acquiring the operating current, residual current, bus voltage, box temperature and humidity data within a preset time period of the distribution box, the method further includes: performing band-limited denoising and abnormal spike suppression preprocessing on the acquired raw electrical signal using a Butterworth filter function, retaining the target frequency band components used for feature extraction.

[0016] Preferably, the discrete wavelet decomposition of the operating current includes: specifying a fourth-order Dobesi wavelet basis in the Dobesi wavelet family as the mother wavelet, performing multi-level discrete wavelet decomposition on the operating current sequence, and extracting the first-level detail coefficient sequence corresponding to the highest frequency band as the detail coefficient sequence corresponding to the preset high-frequency analysis band.

[0017] Preferably, the step of performing a Fourier transform on the residual current waveform within a stable operating cycle representing a healthy state to extract the fundamental amplitude, as well as the amplitudes of the third, fifth, and seventh harmonics, includes: when performing a fast Fourier transform on the residual current waveform, suppressing spectral leakage by adding a Hanning window, and after correcting the windowed spectral amplitude with window function coherence gain, extracting the fundamental amplitude and the amplitudes of each harmonic.

[0018] This invention reduces the spectral energy divergence caused by asynchronous sampling, avoids frequency domain feature distortion, and ensures the accuracy and fidelity of high-order harmonic amplitude feature extraction.

[0019] Preferably, after determining that the insulation performance is in an abnormal degradation state when the insulation state deviation index is greater than the confidence threshold, the method further includes: triggering an insulation fault early warning mechanism through hard-wired switch output or network bus, and sending abnormal log data to indicate the abnormal insulation deviation.

[0020] Secondly, the present invention provides a distribution box insulation performance testing system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned distribution box insulation performance testing method is implemented.

[0021] By adopting the above technical solution, a computer program for detecting the insulation performance of a distribution box is generated and stored in a memory, so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and the processor for convenient use.

[0022] The beneficial effects of this invention are as follows: This invention comprehensively extracts the wavelet features of the operating current and constructs a multi-dimensional state vector by combining the bus voltage stability and the rate of temperature change within the enclosure. It then uses an unsupervised clustering algorithm to accurately select healthy and stable operating cycles, eliminating interference from abnormal operating conditions on the baseline data. Furthermore, it not only extracts the root mean square value of the residual current and multiple harmonic features to construct an aging Gaussian mixture probability model for a refined representation of the baseline insulation state, but also establishes a complex nonlinear mapping relationship between the temperature and humidity environment and the residual current characteristics using a tensor product B-spline function model.

[0023] Furthermore, during the detection phase, the extracted features to be tested are adjusted by using an environmental compensation vector to eliminate the influence of temperature and humidity fluctuations on insulation characteristics. In addition, statistical evaluation is performed by combining the minimum squared Mahalanobis distance and the chi-square distribution confidence threshold, which improves the accuracy of the insulation performance status determination of the distribution box and its resistance to environmental interference. It can sensitively identify early insulation deterioration defects and ensure the safe operation of power distribution equipment. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for testing the insulation performance of a distribution box according to the present invention; Figure 2 This is a schematic diagram of the single-cycle spectrum characteristics of the residual current in this invention; Figure 3 This is a schematic diagram of real-time monitoring of insulation state deviation index in this invention; Figure 4 This is a schematic diagram comparing the evaluation effects of different experimental schemes in this invention. Detailed Implementation

[0025] 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 some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a method for testing the insulation performance of a distribution box, referring to... Figure 1 This includes steps S1-S3: S1. Screen the stable operating cycle of the distribution box.

[0027] In an optional embodiment, the operating current, residual current, bus voltage, box temperature and humidity data within a preset time period of the distribution box are acquired, the current wavelet feature of the operating current is extracted, and a multi-dimensional state vector is constructed by combining the bus voltage stability and the box temperature change rate. Stable operating cycles representing a healthy state are then selected by an unsupervised clustering algorithm.

[0028] Specifically, high-frequency through-hole current transformers, residual current transformers, voltage transformers, and temperature and humidity sensors installed at the inlet and outlet terminals of the distribution box collect operating current, residual current, and bus voltage data at a preset electrical sampling frequency, and collect temperature and humidity data inside the box at a preset environmental sampling frequency. The discrete data sequences of operating current, residual current, bus voltage, box temperature, and humidity are continuously acquired over a preset time period. The Butterworth filter function in the SciPy signal processing library is then used to preprocess the acquired raw electrical signals for band-limited denoising and abnormal peak suppression, preserving the target frequency band components for wavelet high-frequency feature extraction.

[0029] The PyWavelets wavelet analysis library employs a multi-level discrete wavelet decomposition function, specifying a fourth-order Dobesh wavelet basis from the Dobesh wavelet family as the mother wavelet. A three-level discrete wavelet decomposition is performed on the denoised operating current sequence to extract detail coefficient sequences corresponding to the preset high-frequency analysis bands. When using conventional binary discrete wavelet decomposition, the highest frequency band corresponds to the first-level detail coefficient sequence; the variance of the first-level detail coefficient sequence is calculated to obtain the current wavelet characteristic. For the acquired bus voltage data sequence, the residual between the bus voltage sequence and the preset standard sine wave sequence is calculated, and the standard deviation of the residual sequence is obtained. The standard deviation is divided by the rated bus voltage reference value to calculate the voltage fluctuation rate. The reciprocal of the voltage fluctuation rate plus a preset minimum positive number is used as the bus voltage stability. For the acquired in-box temperature data sequence, the in-box temperature difference value including the start and end times of a preset time window of the current operating cycle is obtained. The temperature difference value is divided by the preset time window duration to obtain the in-box temperature change rate. After normalizing the current wavelet features, bus voltage stability, and the rate of temperature change inside the tank, the calculated current wavelet features, bus voltage stability, and the rate of temperature change inside the tank are spliced ​​together according to the feature dimensions to construct a three-dimensional multidimensional state vector.

[0030] The density-based spatial clustering algorithm DBSCAN from the Scikit-learn machine learning library was used, with a multidimensional state vector as the input dataset. The neighborhood radius parameter was set to 0.5, and the minimum neighborhood sample number parameter for the core point was set to 10 for clustering operations. A comprehensive evaluation was performed based on the sample density, sample number, and the current wavelet feature value, bus voltage stability, and absolute value of the temperature change rate within the tank corresponding to each cluster center. Clusters meeting the criteria were marked as operating mode. These criteria included a current wavelet feature value less than a preset current fluctuation threshold, a bus voltage stability greater than a preset stability threshold, and an absolute value of the temperature change rate within the tank less than a preset temperature rise rate threshold. All time segments corresponding to the clusters marked as operating mode were extracted as stable operating cycles representing the healthy state.

[0031] In some embodiments, current wavelet features of the operating current are extracted, and a multi-dimensional state vector is constructed by combining the bus voltage stability and the rate of temperature change within the enclosure. An unsupervised clustering algorithm is then used to select stable operating cycles representing a healthy state. Specifically, the operating current undergoes five-level discrete wavelet decomposition to extract the detail coefficient sequence corresponding to a preset high-frequency analysis band. The variance of the wavelet coefficients within one operating cycle is calculated as the current wavelet features. Within each operating cycle, the bus voltage sequence is continuously sampled, and the residual between the bus voltage sequence and a preset standard sine wave sequence is calculated. The standard deviation of the residual sequence is then obtained, and the voltage fluctuation rate is obtained by dividing the standard deviation by the rated voltage reference value. The bus voltage stability is obtained by adding the reciprocal of a preset minimum positive number to the voltage fluctuation rate. The temperature difference within the enclosure, including the start and end times of a preset time window for each operating cycle, is obtained. The temperature difference is divided by the preset time window duration to obtain the rate of temperature change within the enclosure. The current wavelet features, bus voltage stability, and rate of temperature change within the enclosure are then respectively... After normalization, the data are concatenated into a multi-dimensional state vector, where the rate of change of the temperature inside the chamber is the environmental thermal trend within the time window of the operating cycle. Normalization is used to eliminate the differences in the dimensions and numerical magnitudes of different feature quantities, balancing the weights of each feature in the clustering calculation. The K-Means clustering algorithm is used to cluster the multi-dimensional state vector of historical data. Clusters are selected based on the bus voltage stability, current wavelet feature quantity, and absolute value of the rate of change of the temperature inside the chamber corresponding to the cluster center. Target clusters that meet the conditions are selected, namely, the bus voltage stability is greater than a preset stability threshold, the current wavelet feature quantity is less than a preset current fluctuation threshold, and the absolute value of the rate of change of the temperature inside the chamber is less than a preset temperature rise rate threshold. The operating cycle contained in the target cluster is taken as the stable operating cycle representing the healthy state.

[0032] Taking a 50Hz power distribution box system as an example, the operating cycle is set to 0.02s. Within a preset time window, a current transformer with a sampling frequency of 10kHz is used to collect the operating current. Using Daubechies4 as the wavelet mother function, a five-level discrete wavelet decomposition is performed on the operating current sequence, decomposing the signal into multiple detail layers and approximation layers. The first-level detail coefficient sequence corresponding to the frequency range of 2.5kHz to 5kHz is extracted as the wavelet coefficient sequence for the highest frequency analysis band. The length of the wavelet coefficient sequence is determined by the number of original sampling points, the wavelet decomposition method, and the boundary extension method. The variance of the wavelet coefficients within a single operating cycle is calculated to obtain the current wavelet characteristic quantity reflecting high-frequency transient fluctuations. Under certain experimental conditions, the value of the wavelet characteristic quantity under normal healthy conditions can be between 10 and 10. -4 A² to 10 - Within the range of ³A², the bus voltage sequence is continuously sampled using a voltage transformer within the same operating cycle. The residual between the bus voltage sequence and the preset standard sine wave sequence is calculated, and the standard deviation of the residual sequence is obtained and divided by the rated bus voltage reference value to obtain the voltage fluctuation rate. The reciprocal of the voltage fluctuation rate plus a preset minimum positive number is used as the bus voltage stability parameter. The stability under healthy conditions is usually greater than 100.

[0033] When acquiring environmental data, a preset time window of 10 minutes (600 seconds) was set. A PT100 temperature sensor recorded the absolute temperature inside the chamber at the start and end of the preset time window with an accuracy of 0.1℃. The temperature difference was obtained by subtracting the two values ​​and dividing by 600 seconds to obtain the temperature change rate inside the chamber in °C / s, ranging from -0.05℃ / s to 0.05℃ / s. The current wavelet characteristic, bus voltage stability, and temperature change rate inside the chamber were subjected to Min-Max normalization, scaled to the real number range of 0 to 1, and concatenated to form a three-dimensional multidimensional state vector set. The K-Means algorithm was used to cluster the historical state vector data (e.g., 30 days), with K set to 4, an upper limit of 500 iterations, and a tolerance threshold of 10. -5 After clustering, the coordinates of the center points of the four clusters are traversed, and the optimal cluster is extracted based on the high dimension of bus voltage stability, low dimension of current wavelet feature quantity, and low dimension of absolute value of temperature change rate within the tank. The operating cycle covered by the optimal cluster is taken as the stable operating cycle representing the healthy state.

[0034] S2. Construct a benchmark and environmental compensation model.

[0035] In an optional embodiment, the residual current during each stable operating cycle representing a healthy state is extracted, the root mean square value of the residual current and the normalized amplitudes of the third, fifth and seventh harmonic components are calculated, an aging Gaussian mixture probability model representing the baseline insulation state is constructed, and a tensor product B-spline function model for environmental compensation is constructed based on the box temperature and humidity data and corresponding residual current characteristics during each stable operating cycle representing a healthy state.

[0036] Specifically, for the residual current time series data within each selected stable operating cycle, the root mean square (RMS) algorithm from the NumPy library is used to calculate the RMS value of the residual current. The time-domain residual current signal is converted to the frequency domain, and the amplitudes corresponding to the 50Hz fundamental frequency and the 150Hz, 250Hz, and 350Hz frequencies are extracted. When using window function processing, the extracted spectral amplitudes are corrected using the corresponding window function amplitude. The amplitudes of the third, fifth, and seventh harmonic components are divided by the sum of the fundamental amplitude and a preset minimum positive number, respectively, to calculate the normalized amplitudes of the third, fifth, and seventh harmonic components. The RMS value of the residual current is combined sequentially with the three normalized amplitudes to form a four-dimensional reference feature vector. The four-dimensional residual current feature vector is then normalized or standardized, and the corresponding scaling parameters are saved.

[0037] The Gaussian Mixture model from the Scikit-learn machine learning library is used, with the number of Gaussian mixture components set to 3. The processed four-dimensional residual current feature vectors from all stable operating cycles are used as the input set. An Expectation-Maximization (EM) algorithm is employed for iterative fitting until the model converges, outputting the weight coefficients, mean vector, and covariance matrix of each of the three Gaussian components. This completes the construction of an aging Gaussian mixture probability model representing the normal fluctuation range of the baseline insulation state. Based on this, the corresponding temperature and humidity data within the enclosure during each stable operating cycle are extracted as a two-dimensional input sample point set. The corresponding four-dimensional residual current feature vector is used as the target dependent variable. Two-dimensional spline fitting functions from the SciPy interpolation library or a least-squares tensor product B-spline fitting algorithm are used to approximate the temperature and humidity dimensions using 3rd-order B-spline basis functions. The tensor product algorithm is used to multiply the temperature and humidity basis functions to obtain the basis matrix. The least-squares algorithm is then used to solve for the control point weight coefficient matrix, thus completing the construction of a tensor product B-spline function model for environmental compensation. When the temperature and humidity data form a regular grid, a two-dimensional regular grid spline fitting function is used; when the temperature and humidity data are discrete scattered samples, the least squares tensor product B-spline fitting method is used to solve the control point weight coefficient matrix.

[0038] In some embodiments, the residual current within each stable operating cycle representing a healthy state is extracted, and the root mean square value of the residual current and the normalized amplitudes of the third, fifth, and seventh harmonic components are calculated to construct an aging Gaussian mixture probability model representing the reference insulation state. Specifically, a fast Fourier transform is performed on the residual current waveform within the stable operating cycle representing a healthy state to extract the fundamental amplitude and the amplitudes of the third, fifth, and seventh harmonics; the amplitudes of the third, fifth, and seventh harmonics are divided by the sum of the fundamental amplitude and a preset minimum positive number to obtain the normalized amplitudes of the corresponding third, fifth, and seventh harmonic components, and these are combined with the root mean square value of the residual current to form a four-dimensional reference feature vector set; the four-dimensional reference feature vector set is then normalized or... Standardize the model and save the scaling parameters; initialize the mixing weights, mean vector, and covariance matrix of the aged Gaussian mixture probability model, and establish the log-likelihood function; use the expectation-maximization algorithm to iteratively train the four-dimensional baseline feature vector set, calculate the posterior probability of each feature vector belonging to each Gaussian component in the expectation step, and update the mixing weights, mean vector, and covariance matrix in the maximization step; stop the iteration when the change in the log-likelihood function between two adjacent iterations is less than the preset convergence threshold, and output the optimized aged Gaussian mixture probability model.

[0039] The residual current waveform data collected by the zero-sequence current transformer during a stable operating cycle, representing a healthy state, is acquired at a sampling rate of 12.8kHz to satisfy the Nyquist sampling theorem and provide high anti-aliasing performance. A Fast Fourier Transform is performed on 256 sampling points for a single cycle. A Hanning window is added as needed to suppress spectral leakage, and the windowed spectral amplitude is corrected using a window function coherent gain. From this, the fundamental amplitude at 50Hz, as well as the third harmonic amplitude at 150Hz, the fifth harmonic amplitude at 250Hz, and the seventh harmonic amplitude at 350Hz, are extracted.

[0040] Reference Figure 2 The fundamental frequency amplitude reaches its highest value at 50Hz, dominating the spectrum, while the amplitudes of the third, fifth, and seventh harmonic components at 150Hz, 250Hz, and 350Hz are significantly lower than that of the fundamental frequency. This reflects that under healthy conditions, the residual current is mainly composed of the fundamental frequency component, with relatively small high-frequency harmonic components, thus proving the rationality of constructing a reference characteristic by extracting the aforementioned specific frequency components.

[0041] The normalized amplitudes are obtained by dividing the amplitudes of the third, fifth, and seventh harmonics by the sum of the fundamental amplitude and a preset minimum positive number, respectively. Simultaneously, the root mean square (RMS) value of the residual current within a single cycle is calculated as the RMS amplitude. A single four-dimensional baseline feature vector is constructed by strictly concatenating the RMS amplitude and the normalized amplitudes of the three harmonics in strict order. The feature vectors from 100,000 healthy cycles are aggregated to form a large-scale training feature set. This training feature set is then normalized or standardized to reduce the impact of features with different dimensions on model training.

[0042] When constructing the aging Gaussian mixture probability model, the number of Gaussian components, K, is preset to 3, to represent multiple normal statistical sub-distributions that may exist in the residual current characteristics under healthy insulation conditions. The GMM parameters are initialized using the K-Means++ algorithm, i.e., each component is assigned an initial mixing weight of 1 / 3. The cluster centers of K-Means are used as the initial values ​​for the mean vector, and the within-cluster sample covariance is used as the initial value for the covariance matrix to construct the global log-likelihood function. The expectation-maximization algorithm is then initiated for iterative model fitting: in the E-step, Bayes' theorem is used to calculate the posterior probability that each sample in the current four-dimensional baseline feature vector set belongs to the above three Gaussian components; in the M-step, the mixing weights, 4×1 dimensional mean vector, and 4×4 dimensional covariance matrix of the above three components are updated weighted according to the calculated posterior probabilities. The maximum number of iterations in the EM iteration process is set to 1000, and the increment of the log-likelihood function value is monitored in real time. When the absolute value of the difference between two adjacent iterations of the log-likelihood function is less than the preset convergence threshold of 10, the model is considered complete. -6 When the model reaches the preset convergence state, the iteration is automatically stopped and the aging Gaussian mixture probability model parameters reflecting the normal fluctuation boundary of the baseline insulation state are solidified and output.

[0043] In some embodiments, a tensor product B-spline function model for environmental compensation is constructed based on the chamber temperature and humidity data and corresponding residual current characteristics within stable operating cycles representing healthy states. Specifically, extreme temperature ranges are extracted from the historical dataset, and these ranges are divided into several equal temperature intervals. A first node vector is set at the boundary of each temperature interval to construct a temperature-based one-dimensional cubic B-spline basis function. Extreme humidity ranges are extracted from the historical dataset, and these ranges are divided into several equal humidity intervals. A second node vector is set at the boundary of each humidity interval to construct a humidity-based one-dimensional cubic B-spline basis function. The temperature-based and humidity-based one-dimensional cubic B-spline basis functions are then subjected to a tensor product operation to generate a two-dimensional tensor product B-spline basis function matrix. Using the chamber temperature and humidity data within stable operating cycles representing healthy states as input, and the corresponding four-dimensional reference feature vectors after the same normalization or standardization processing as the expected output, the control vertex weight coefficient matrix is ​​solved using the least squares method to obtain a tensor product B-spline function model that maps the chamber temperature and humidity to the residual current feature reference values.

[0044] A comprehensive 12-month historical health operation dataset was scanned to extract the extreme temperature range within the distribution box's microenvironment. This range was then uniformly divided into several temperature intervals with a step size of 10℃. Based on B-spline mathematical theory, repeating nodes were placed at both ends of the intervals to ensure boundary stability. Internal nodes were distributed equidistantly along the interval boundaries, generating the first node vector. Then, a 3rd-order, 4th-degree cubic B-spline basis function for temperature was generated using the DeBoor-Cox recursive formula. Similarly, the extreme relative humidity range within the box was extracted, and the range was uniformly divided into several humidity intervals with a step size of 10%RH. Repeating nodes were also placed at both ends to form the second node vector, thus constructing a 3rd-order, 4th-degree cubic B-spline basis function for humidity.

[0045] To achieve nonlinear compensation modeling of the insulation characteristics caused by the multidimensional coupling of temperature and humidity, Kronecker tensor product operations are performed on the one-dimensional basis functions in the two directions to generate a two-dimensional tensor product B-spline basis function matrix. In one implementation, 11 cubic B-spline basis functions are set in the temperature direction and 10 cubic B-spline basis functions are set in the humidity direction. After tensor product, 11×10, or 110, two-dimensional basis elements are generated. During model training, hundreds of thousands of two-dimensional temperature and humidity coordinate pairs representing stable operating cycles under healthy conditions are extracted as input parameters and substituted into the basis function matrix to generate a large-scale observation matrix; simultaneously, the four-dimensional baseline feature vectors corresponding to each cycle obtained in the previous steps are used as the desired output target values. Using 10... -5A multivariate linear least squares method with Tikhonov regularization terms is used to prevent overfitting and matrix ill-conditioned problems, solving for a control vertex weight coefficient matrix corresponding to the number of two-dimensional basis elements in one step; in the above implementation with 110 two-dimensional basis elements, the control vertex weight coefficient matrix is ​​110 rows and 4 columns. The two-dimensional tensor product B-spline basis function matrix is ​​multiplied by the control vertex weight coefficient matrix by an inner product to establish a tensor product B-spline function model that maps from real-time temperature and humidity point fitting to the four-dimensional residual current characteristic benchmark estimate.

[0046] S3. Determine the insulation performance status of the distribution box.

[0047] In an optional embodiment, the residual current feature vector is extracted for the current test period. An environmental compensation vector is calculated based on the current temperature and humidity data inside the box using a tensor product B-spline function model. The residual current feature vector is subtracted from the environmental compensation vector, and then the overall mean vector of the aging Gaussian mixture probability model is added to obtain the adjusted feature vector. The square of the Mahalanobis distance between the adjusted feature vector and the corresponding mean vector is calculated using the covariance matrix of each component in the aging Gaussian mixture probability model. The minimum value is selected as the insulation state deviation index and compared with a preset chi-square distribution confidence threshold to determine the insulation performance status of the distribution box.

[0048] Specifically, for the residual current signal acquired in real time for the current test period, the root mean square value of the residual current and the normalized amplitudes of the third, fifth, and seventh harmonics are calculated using the same root mean square algorithm, fast Fourier transform (FFT) function processing flow, harmonic normalization processing flow, and scaling transformation flow as in the training phase. These are combined to form a four-dimensional residual current test feature vector, and the residual current test feature vector is processed at the same scale using the normalization or standardization parameters saved in the training phase. Real-time temperature and humidity data inside the chamber for the current test period are read and input into the trained tensor product B-spline function model. The evaluation interface of this spline model is called to calculate the four-dimensional output vector corresponding to the current temperature and humidity boundary conditions, which serves as the environmental compensation vector. The environmental compensation vector and the residual current test feature vector are in the same feature scale space. Subtract the value of the corresponding dimension of the environmental compensation vector from the value of each dimension of the residual current to be measured feature vector. At the same time, extract the mean vector of the three Gaussian components in the aging Gaussian mixture probability model. Multiply the vector by their respective weight coefficients and add them together to obtain the total expectation, thus obtaining the overall mean vector of the probability model. The overall mean vector and the residual current to be measured feature vector are in the same feature scale space. Use array addition instructions to add the residual vector after subtraction to the overall mean vector to obtain the adjusted feature vector after reducing the interference of environmental temperature and humidity.

[0049] The Mahalanobis function from the SciPy spatial distance library is called to calculate the distance between the adjusted feature vector and the mean vector of the three Gaussian components in the aging Gaussian mixture probability model. When calling the Mahalanobis function, the inverse matrix parameter of the covariance matrix corresponding to each Gaussian component after adding a preset regularization term is passed in. The three Mahalanobis distance values ​​returned by the calculation are squared respectively, and the minimum value is extracted. This minimum value is determined as the insulation state deviation index.

[0050] Call the chi2 function of the chi-square distribution continuous random variable in the SciPy statistical distribution library, set the degrees of freedom parameter of the chi-square distribution to 4 according to the dimension of the feature vector, and set the preset confidence parameter, such as 95% or 99%. Call the percentage point ppf function to calculate the theoretical critical value as the preset confidence threshold of the chi-square distribution, or calibrate the confidence threshold in combination with the validation sample.

[0051] If the calculated insulation status deviation index value is less than or equal to the chi-square distribution signal threshold, the current insulation performance status of the distribution box is determined to be normal and output as a signal. If the insulation status deviation index value is greater than the chi-square distribution signal threshold, the current insulation performance status of the distribution box is determined to have deteriorated abnormally, and the system interface is called to trigger hardware alarm actions and send abnormal log data according to the early warning strategy.

[0052] In some embodiments, the residual current feature vector to be measured is extracted for the current test period. The environmental compensation vector is calculated based on the current temperature and humidity data inside the box using the tensor product B-spline function model. The residual current feature vector is subtracted from the environmental compensation vector and then the overall mean vector of the aging Gaussian mixture probability model is added to obtain the adjusted feature vector. The square of the Mahalanobis distance between the adjusted feature vector and the corresponding mean vector is calculated using the covariance matrix of each component in the aging Gaussian mixture probability model. The minimum value is selected as the insulation state deviation index and compared with the preset chi-square distribution confidence threshold to determine the insulation performance status of the distribution box.

[0053] Specifically, the current temperature and humidity data inside the chamber during the test period are collected and substituted into the tensor product B-spline function model. The model outputs a baseline residual current feature vector predicted by the model under the current temperature and humidity conditions inside the chamber. This baseline residual current feature vector is used as the environmental compensation vector, where the environmental compensation vector and the baseline feature vector from the training phase are in the same feature scale space. The mean vectors of each Gaussian component in the aging Gaussian mixture probability model are multiplied by their corresponding mixture weights and summed to obtain the overall mean vector. The environmental compensation vector is subtracted from the residual current test feature vector and then added to the overall mean vector to obtain the adjusted feature vector after reducing the influence of the temperature and humidity environment inside the chamber. In the Gaussian mixture probability model, the covariance matrix corresponding to each Gaussian component is calculated, and the squared Mahalanobis distance between the adjusted eigenvector and the corresponding mean vector is calculated. The smallest squared Mahalanobis distance is selected as the insulation state deviation index. In the chi-square distribution table where the degrees of freedom are equal to the dimension of the residual current eigenvector to be measured, the critical value corresponding to the preset significance level is found as the confidence threshold, or the confidence threshold is calibrated based on the validation samples. When the insulation state deviation index is greater than the confidence threshold, the insulation performance of the distribution box is determined to be in an abnormal degradation state, triggering an insulation fault warning. When the insulation state deviation index is less than or equal to the confidence threshold, the insulation performance of the distribution box is determined to be in a healthy and stable state.

[0054] During real-time online monitoring of the distribution box, the residual current signal of the test period is extracted in real time with a test period of 0.02s. The temperature and relative humidity sensor data inside the box, which are most recently acquired or processed by moving average, are read and substituted into the trained and converged two-dimensional tensor product B-spline function model. By evaluating the 110 basis functions at the current temperature and humidity point and performing a linear combination with the control vertex weight matrix, the baseline predicted value of the four-dimensional features under the coupled temperature and humidity state is calculated, forming a 4×1 dimensional environmental compensation vector. The previously optimized aging Gaussian mixture probability model is called, and the model's internal K is extracted as the 4×1 dimensional mean vector corresponding to each of the three components and the assigned scalar mixture weights. The three mean vectors are multiplied by their corresponding mixture weights and then vector-summed to synthesize the overall mean vector representing the global desired health state. The residual current four-dimensional feature vector to be measured is obtained by using the actual calculation of the residual current during the test period. After processing it to the feature scale space consistent with the environmental compensation vector and the overall mean vector according to the scale transformation parameters saved in the training stage, the environmental compensation vector is subtracted one by one according to the corresponding element position to reduce the influence of feature drift caused by temperature and humidity changes. Then, the overall mean vector is added to the whole and translated to obtain the four-dimensional adjusted feature vector reflecting the change of insulation state.

[0055] The squared Mahalanobis distance between the four-dimensional adjusted eigenvector and the three components of the aging Gaussian mixture probability model is calculated. During the calculation, a Mahalanobis transformation is performed on the inverse matrix of the 4×4 covariance matrix corresponding to each Gaussian component after adding a preset regularization term. This reduces the differences in dimensions between features and potential multicollinearity interference. Three distance results are compared, and the one with the smallest value is selected as the insulation state deviation index. A significance level of 0.01 is set. Since the dimension of the eigenvector to be tested is constant at 4, a standard chi-square distribution table with 4 degrees of freedom is consulted, and a critical value of 13.277 is extracted and used as the chi-square distribution confidence threshold. The critical value corresponds to a 99% confidence level. In other implementations, the corresponding threshold can also be determined based on a preset confidence level or validation samples. When the minimum squared Mahalanobis distance calculated in real time is greater than 13.277, it indicates that the adjusted insulation characteristics have exceeded the baseline tolerance distribution boundary of GMM in the multidimensional probability space. The insulation performance of the distribution box is judged to be in an abnormal degradation risk state, and the insulation fault early warning mechanism is triggered through hard-wired switch output or network bus. Conversely, if the distance index does not exceed 13.277, it is confirmed that the distribution box maintains a healthy and stable insulation state, and no intervention is taken.

[0056] Reference Figure 3 During the test period from 0 to 80, the insulation condition deviation from the index curve fluctuated, but the values ​​were all below the chi-square distribution information threshold level (dashed line), reflecting that the insulation performance of the distribution box was in a healthy and stable state at this time. In the test period after 80, the index curve jumped significantly upward, with values ​​exceeding the chi-square distribution information threshold level (dashed line), indicating that the insulation characteristics deviated abnormally at this time. This proves that combining the squared Mahalanobis distance with the chi-square distribution information threshold can effectively identify the insulation degradation state.

[0057] The experiment selected a 50Hz power distribution box system as the test object, with the ambient temperature fluctuating between -20℃ and 80℃, and the relative humidity continuously varying between 10%RH and 100%RH. The hardware sampling frequency was set to 10kHz for acquiring the operating current and 12.8kHz for residual current waveform detection. Three comparison groups were constructed: the first group consisted of a complete scheme including a tensor product B-spline environmental compensation module, an aging Gaussian mixture probability model, and wavelet and harmonic feature extraction; the second group was an uncompensated control group without tensor product B-spline environmental compensation calculations; and the third group was a basic feature control group that removed high-frequency wavelet features and high-order harmonic extraction, retaining only the root mean square value of the residual current. The dataset used for model testing included 100,000 training samples representing stable operating cycles in a healthy state, and 5,000 test samples with complex operating conditions involving weak insulation leakage and abnormal impedance changes.

[0058] After iterative training using the same expectation-maximization algorithm and validation on the test set, the first complete solution group demonstrated excellent overall fault detection performance, achieving an insulation condition assessment accuracy of 98.5%, a false alarm rate of only 1.2%, and a false negative rate of 1.5%. The second uncompensated control group, due to interference from drastic fluctuations in temperature and humidity within the enclosure, saw its minimum Mahalanobis distance deviation frequently exceed the preset chi-square distribution confidence threshold, resulting in an overall accuracy drop to 89.2% and a false alarm rate increase to 9.8%. For the third basic feature control group, with its feature dimension reduced to a single dimension, the chi-square distribution freedom for Mahalanobis distance determination was simultaneously reduced to 1, and an independent critical value of 6.635 corresponding to a 99% confidence level was used as the abnormal confidence threshold. Despite using a threshold adapted to its dimension, the third basic feature control group, due to its single feature dimension and lack of detection capability for high-frequency transient fluctuations, still failed to reliably identify a large number of early, weak insulation degradation samples, achieving an accuracy of only 84.6% and a false negative rate of 18.2%.

[0059] Comparing the results of the first and second groups, it can be seen that using the tensor product B-spline function model to output the environmental compensation vector and reconstruct and adjust the eigenvector can effectively reduce the nonlinear drift of eigenvalues ​​caused by thermal expansion and contraction and water vapor adsorption, resulting in a decrease in the false alarm rate of 8.6 percentage points. Furthermore, the data differences between the first and third groups confirm that performing multi-level discrete wavelet decomposition on the operating current to extract high-frequency wavelet coefficients and performing fast Fourier transform on the residual current to extract the normalized amplitude of higher harmonics can fully characterize latent early faults such as capacitive coupling offset, reducing the false alarm rate by 16.7 percentage points.

[0060] Reference Figure 4 The data includes a complete solution group, an uncompensated group, and a basic feature group. As can be seen from the differences in the height of the bars in the graph, the complete solution group has the highest accuracy rate and maintains the lowest false alarm and false negative rates. The false alarm rate of the uncompensated group is significantly higher than that of the complete solution group, while the false negative rate of the basic feature group is significantly higher than that of the complete solution group. This demonstrates that using a combination of environmental compensation and multi-feature extraction mechanisms can improve the accuracy of insulation detection in distribution boxes and reduce false alarm and false negative rates.

[0061] This invention also discloses a distribution box insulation performance testing system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a distribution box insulation performance testing method according to the present invention.

[0062] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0063] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for testing the insulation performance of a distribution box, characterized in that, include: S1. Obtain data on operating current, residual current, bus voltage, box temperature, and humidity within a preset time period for the distribution box; Extract the wavelet features of the operating current, and construct a multi-dimensional state vector by combining the bus voltage stability and the temperature change rate inside the box. Use an unsupervised clustering algorithm to select stable operating cycles that represent the healthy state. S2. Extract the residual current in each stable operating cycle that represents the healthy state, calculate the root mean square value of the residual current and the normalized amplitude of multiple harmonic components as residual current features, and construct an aging Gaussian mixture probability model that represents the reference insulation state. Based on the box temperature and humidity data and corresponding residual current characteristics during stable operating cycles representing various health states, a tensor product B-spline function model for environmental compensation is constructed. S3: Extract the residual current feature vector for the current test cycle. Calculate the environmental compensation vector using the tensor product B-spline function model based on the current box temperature and humidity data. Subtract the environmental compensation vector from the residual current feature vector and add the overall mean vector of the aging Gaussian mixture probability model to obtain the adjusted feature vector. Calculate the square of the Mahalanobis distance between the adjusted feature vector and the corresponding mean vector using the covariance matrix of each component in the aging Gaussian mixture probability model. Select the minimum value as the insulation state deviation index and compare it with the preset chi-square distribution confidence threshold to determine the insulation performance status of the distribution box.

2. The method for testing the insulation performance of a distribution box according to claim 1, characterized in that, The process of constructing a multidimensional state vector and selecting stable operating cycles representing a healthy state using an unsupervised clustering algorithm includes: performing discrete wavelet decomposition on the operating current, calculating the variance of the detailed coefficient sequence corresponding to a preset high-frequency analysis band as a current wavelet feature; calculating the residual between the bus voltage sequence and a preset standard sine wave sequence, and obtaining the standard deviation of the residual sequence, dividing the standard deviation by the rated voltage reference value to obtain the voltage fluctuation rate, and obtaining the bus voltage stability based on the reciprocal of the voltage fluctuation rate; calculating the temperature difference within the tank within a preset time window, dividing the temperature difference by the duration of the time window to obtain the temperature change rate within the tank; normalizing the current wavelet feature, bus voltage stability, and temperature change rate within the tank and concatenating them into a multidimensional state vector, clustering them using a clustering algorithm and selecting target clusters, and selecting the operating cycles contained in the target clusters as stable operating cycles.

3. The method for testing the insulation performance of a distribution box according to claim 1, characterized in that, The construction of the aging Gaussian mixture probability model representing the baseline insulation state includes: performing a Fourier transform on the residual current waveform during a stable operating cycle representing the healthy state to extract the fundamental amplitude and the amplitudes of the third, fifth, and seventh harmonics; dividing the amplitudes of the third, fifth, and seventh harmonics by the sum of the fundamental amplitude and a preset minimum positive number to obtain the corresponding normalized amplitudes, and combining them with the root mean square value of the residual current to form a four-dimensional baseline feature vector set; initializing the mixture weights, mean vector, and covariance matrix of the aging Gaussian mixture probability model, and establishing a log-likelihood function; using the expectation-maximization algorithm to iteratively train the four-dimensional baseline feature vector set, stopping the iteration when the change in the log-likelihood function between two adjacent iterations is less than a preset convergence threshold, and outputting the optimized aging Gaussian mixture probability model.

4. The method for testing the insulation performance of a distribution box according to claim 1, characterized in that, The construction of the tensor product B-spline function model for environmental compensation includes: extracting and equally dividing the extreme temperature ranges within the chamber from the historical dataset, setting first node vectors at the boundaries, and constructing a one-dimensional B-spline basis function based on temperature; extracting and equally dividing the extreme humidity ranges from the historical dataset, setting second node vectors at the boundaries, and constructing a one-dimensional B-spline basis function based on humidity; performing a tensor product operation on the one-dimensional B-spline basis function based on temperature and the one-dimensional B-spline basis function based on humidity to generate a two-dimensional tensor product B-spline basis function matrix; using the temperature and humidity data within the chamber as input and the corresponding four-dimensional baseline feature vectors as the desired output, solving the control vertex weight coefficient matrix to obtain the tensor product B-spline function model.

5. The method for testing the insulation performance of a distribution box according to claim 1, characterized in that, The method of calculating the environmental compensation vector and determining the insulation performance status of the distribution box using the tensor product B-spline function model includes: substituting the box's internal temperature and humidity data for the test period into the tensor product B-spline function model to output the environmental compensation vector; multiplying the mean vectors of each Gaussian component in the aging Gaussian mixture probability model by their corresponding mixture weights and summing them to obtain the overall mean vector; subtracting the environmental compensation vector from the residual current test feature vector and adding it to the overall mean vector to obtain the adjusted feature vector; using the covariance matrix corresponding to each Gaussian component in the aging Gaussian mixture probability model, calculating the squared Mahalanobis distance between the adjusted feature vector and the corresponding mean vector, and selecting the minimum value as the insulation state deviation index; when the insulation state deviation index is greater than the confidence threshold, the insulation performance is determined to be in an abnormal degradation state; otherwise, it is determined to be in a healthy and stable state.

6. The method for testing the insulation performance of a distribution box according to claim 1, characterized in that, After acquiring the operating current, residual current, bus voltage, box temperature and humidity data within a preset time period of the distribution box, the process further includes: using a Butterworth filter function to perform band-limited denoising and abnormal spike suppression preprocessing on the acquired raw electrical signal, retaining the target frequency band components used for feature extraction.

7. The method for testing the insulation performance of a distribution box according to claim 2, characterized in that, The discrete wavelet decomposition of the operating current includes: specifying a fourth-order Dobesi wavelet basis in the Dobesi wavelet family as the mother wavelet, performing multi-level discrete wavelet decomposition on the operating current sequence, and extracting the first-level detail coefficient sequence corresponding to the highest frequency band as the detail coefficient sequence corresponding to the preset high-frequency analysis band.

8. The method for testing the insulation performance of a distribution box according to claim 3, characterized in that, The step of performing a Fourier transform on the residual current waveform during a stable operating cycle representing a healthy state to extract the fundamental amplitude, as well as the amplitudes of the third, fifth, and seventh harmonics, includes: when performing a fast Fourier transform on the residual current waveform, suppressing spectral leakage by adding a Hanning window, and then extracting the fundamental amplitude and the amplitudes of each harmonic after correcting the windowed spectral amplitude with a window function coherence gain.

9. A method for testing the insulation performance of a distribution box according to claim 5, characterized in that, When the insulation state deviation index is greater than the confidence threshold, after determining that the insulation performance is in an abnormal degradation state, the method further includes: triggering an insulation fault early warning mechanism through hard-wired switch output or network bus, and sending abnormal log data to indicate abnormal insulation deviation.

10. A distribution box insulation performance testing system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a method for testing the insulation performance of a distribution box according to any one of claims 1-9.

Images