Non-invasive detection method and system for high-voltage circuit breakers based on electromagnetic signal analysis

Abstract

This application relates to a non-invasive detection method and system for high-voltage circuit breakers based on electromagnetic signal analysis, belonging to the field of high-voltage circuit breaker detection technology. The method includes: acquiring multiple raw signals based on a broadband microstrip antenna array; extracting arc duration, peak energy, and energy decay rate to construct an initial feature vector; correcting a preset distance metric based on the parameter distribution of the energy decay rate to calculate the current feature discrimination; generating an optimized feature vector by iteratively adjusting the spatial sensing weights of the broadband microstrip antenna array using a weight optimization algorithm, with the goal of maximizing the current feature discrimination; inputting the optimized feature vector into a multi-state classifier, which outputs contact and mechanical health indicators respectively, and performing health status assessment and early warning based on threshold fusion. This invention solves the problem that traditional circuit breaker condition detection methods mainly rely on offline preset thresholds and fixed analysis models, adopting a passive response mode, resulting in diagnosis lagging behind the actual changes inside the equipment.

Description

Technical Field

[0001] This application relates to the field of high-voltage circuit breaker testing, and in particular to a non-invasive testing method and system for high-voltage circuit breakers based on electromagnetic signal analysis. Background Technology

[0002] High-voltage circuit breakers are core equipment for the safe and stable operation of power systems. Faults such as contact degradation and mechanical defects directly threaten the reliability of power grid supply. Accurate detection and early warning of equipment status have become a key technical requirement in the field of power operation and maintenance.

[0003] Currently, traditional circuit breaker condition detection methods mainly rely on offline preset thresholds and fixed analysis models, adopting a passive response mode. This not only fails to dynamically adapt to real-time changes in equipment operation, but also makes it difficult to capture the early evolution trend of internal faults. This not only causes diagnostic results to lag behind the actual equipment status, but also reduces the pertinence and timeliness of operation and maintenance decisions, and increases the risk of power grid faults. Summary of the Invention

[0004] This application provides a non-intrusive detection method and system for high-voltage circuit breakers based on electromagnetic signal analysis, which improves the problems of traditional circuit breaker condition detection relying on offline preset thresholds and fixed analysis models, and adopting a passive response mode, which leads to diagnosis lagging behind the actual changes inside the equipment.

[0005] This application discloses the following technical solution:

[0006] In a first aspect, this application provides a non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis, the method comprising:

[0007] Based on the capture of electromagnetic pulse signals of contact arc during the opening process of the target circuit breaker using a broadband microstrip antenna array, multiple raw signals are obtained;

[0008] The multiple raw signals are preprocessed, and the arc duration, peak energy and energy decay rate of each signal are extracted to construct an initial feature vector;

[0009] Based on the parameter distribution of the energy decay rate, the preset distance metric is corrected, and the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector is calculated as the current feature discrimination.

[0010] With the goal of maximizing the current feature discriminative power, the spatial sensing weights of the broadband microstrip antenna array are iteratively adjusted using a gradient-guided weight optimization algorithm to generate an optimized feature vector.

[0011] The optimized feature vector is input into a pre-trained multi-state classifier, which maps the output contact health index and mechanical health index, and then uses adaptive threshold fusion to make judgments, thereby realizing health status assessment and early warning.

[0012] Secondly, this application provides a non-invasive detection system for high-voltage circuit breakers based on electromagnetic signal analysis, the system comprising:

[0013] The electromagnetic signal acquisition module is used to capture the contact arc electromagnetic pulse signal during the opening process of the target circuit breaker based on a broadband microstrip antenna array, and to acquire multiple raw signals.

[0014] The feature extraction and construction module is used to preprocess the multiple original signals and extract the arc duration, peak energy and energy decay rate of each signal to construct an initial feature vector.

[0015] The feature discrimination calculation module is used to correct the preset distance metric based on the parameter distribution of the energy decay rate, and calculate the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector as the current feature discrimination.

[0016] The weight optimization generation module is used to generate an optimized feature vector by iteratively adjusting the spatial sensing weights of the broadband microstrip antenna array through a gradient-guided weight optimization algorithm with the goal of maximizing the current feature discrimination.

[0017] The health assessment and early warning module is used to input the optimized feature vector into a pre-trained multi-state classifier, map the output contact health index and mechanical health index, and make judgments based on adaptive threshold fusion to realize health status assessment and early warning.

[0018] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0019] This application proposes a non-invasive detection method and system for high-voltage circuit breakers based on electromagnetic signal analysis. By capturing electromagnetic signals step by step, extracting and constructing features, calculating feature discrimination, optimizing antenna weights, model classification evaluation, and adaptive threshold early warning, it achieves accurate detection of contact degradation and mechanical defects of high-voltage circuit breakers and dynamic early warning of their health status. First, based on a broadband microstrip antenna array surrounding the target circuit breaker, the electromagnetic pulse signal of the contact arc during the opening process is simultaneously captured, acquiring multiple raw signals. Then, the raw signals undergo preprocessing such as baseline calibration and bandpass filtering to extract the arc duration, peak energy, and energy decay rate. After outlier correction, an initial feature vector is constructed. Subsequently, the distance metric is corrected based on the parameter distribution of the energy decay rate, and the statistical distance between two feature subsets—contact degradation and mechanical defects—is calculated as the current feature discrimination. With the goal of maximizing feature discrimination, the antenna spatial sensing weights are iteratively adjusted using a gradient-guided weight optimization algorithm to generate an optimized feature vector. The optimized feature vector is input into a pre-trained dual-branch multi-state classifier, mapping and outputting two types of health indicators. Finally, by constructing a threshold particle set, fusing adjacent thresholds, and dynamically adjusting adaptive thresholds based on health trend prediction, health status assessment and targeted early warning are achieved.

[0020] The technical solution of this application solves the problems of traditional detection methods, such as reliance on offline fixed thresholds, diagnostic lag, low accuracy of single signal recognition, and susceptibility to interference. It avoids false and missed fault detection and improves the accuracy, timeliness and reliability of high-voltage circuit breaker status detection. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating the non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis provided in this application embodiment;

[0023] Figure 2 This is a schematic diagram of the structure of a non-invasive detection system for high-voltage circuit breakers based on electromagnetic signal analysis, provided in an embodiment of this application.

[0024] The components represented by each number in the attached diagram are explained below:

[0025] Electromagnetic signal acquisition module 01, feature extraction and construction module 02, feature discrimination calculation module 03, weight optimization and generation module 04, and health assessment and early warning module 05. Detailed Implementation

[0026] This application provides a non-intrusive detection method and system for high-voltage circuit breakers based on electromagnetic signal analysis, which solves the technical problem that existing traditional circuit breaker condition detection methods mainly rely on offline preset thresholds and fixed analysis models, adopt a passive response mode, and cause the diagnosis to lag behind the actual changes inside the equipment.

[0027] Example 1, as shown in the appendix Figure 1 As shown, this application provides a non-intrusive detection method for high-voltage circuit breakers based on electromagnetic signal analysis. The method includes the following steps:

[0028] S110: Based on a broadband microstrip antenna array, capture the contact arc electromagnetic pulse signal during the opening process of the target circuit breaker and acquire multiple raw signals;

[0029] In this embodiment of the application, in order to capture key characteristic signals of the high-voltage circuit breaker's operating status and provide data support for fault diagnosis, it is necessary to achieve non-intrusive signal acquisition through a dedicated antenna array and ensure the integrity and synchronization of the signals in order to build a high-quality raw data foundation.

[0030] First, a broadband microstrip antenna array is set up based on the deployment method of surrounding the target circuit breaker. The array uses a rectangular microstrip antenna, which is suitable for the field monitoring needs of the circuit breaker due to its advantages of low cost, light weight and mature design.

[0031] Specifically, during antenna deployment, it is essential to ensure that all antennas are located in the far-field region. The installation distance is determined based on the correspondence between the antenna's maximum size and the wavelength of its center frequency. For example, when the antenna's maximum size is 0.3 meters and the wavelength corresponding to its center frequency is 1 meter, the antenna installation distance is set to 3 meters. This satisfies the requirements for spatial field distribution stability while also adapting to the space constraints at the substation site. Simultaneously, a foam plastic base provides mechanical support for the antenna; its dielectric constant is close to that of air, which reduces interference with signal acquisition.

[0032] Furthermore, signal acquisition for broadband microstrip antenna arrays needs to focus on the electromagnetic pulse signals of the contact arc during circuit breaker opening operations. These signals contain crucial information about the contact state and mechanical actions. During acquisition, it is essential to ensure the synchronization of multiple antennas and set a uniform acquisition frequency. For example, a sampling rate of 2.5 GS / s can be used to ensure that the signals captured by different antennas remain consistent in the time dimension.

[0033] For example, when monitoring a minimum oil circuit breaker with a rated voltage of 15kV, four broadband microstrip antennas are deployed around its switch rod, each installed at a distance of 3 meters. During the process of the circuit breaker performing a tripping operation and interrupting the 300A continuous current, the electromagnetic pulse signal radiated by the electric arc is captured synchronously, and finally four raw signals are obtained. Each signal contains five million data points, which completely records the signal change trajectory of the electric arc from ignition to extinction during the tripping process.

[0034] S120: Preprocess the multiple original signals and extract the arc duration, peak energy and energy decay rate of each signal to construct an initial feature vector;

[0035] In this embodiment of the application, in scenarios where the original signal is susceptible to environmental interference, contains noise, and has an irregular data form, in order to purify the effective information in the signal and extract the key features that can reflect the circuit breaker status, it is necessary to carry out systematic preprocessing and feature extraction on multiple original signals, while eliminating the influence of outliers in the data, so as to ensure the accuracy of subsequent feature differentiation and status assessment.

[0036] Specifically, baseline calibration and bandpass filtering are first performed on each original signal. The calibration corrects the signal reference deviation, and the filtering removes unwanted noise caused by environmental interference, thereby obtaining a calibrated signal with higher purity.

[0037] Furthermore, for each calibrated signal, continuous signal segments exceeding a preset energy threshold are identified. The arc duration is determined by calculating the time difference between the start point of the first signal segment and the end point of the last signal segment. The peak energy is obtained by calculating the energy integral value within all continuous signal segments. At the same time, the reciprocal of the time it takes for the energy to decay from the global maximum value to the preset proportion is calculated as the energy decay rate.

[0038] Furthermore, the arc duration, peak energy, and energy decay rate extracted from all paths are summarized to form corresponding sequence data. For the peak energy sequence and the energy decay rate sequence, their respective quartiles, upper quartiles, and interquartile ranges are calculated. Based on these statistics and preset multiplication coefficients, the statistical upper and lower limits of each parameter sequence are determined, and measurement points exceeding these limits are marked as outliers. For the marked outliers, the moving median of their adjacent normal measurements in their respective parameter sequences is used for replacement and correction, generating a corrected parameter sequence.

[0039] Finally, based on the corrected peak energy sequence, the corrected energy decay rate sequence, and the arc duration sequence, all effective feature data are integrated in a preset order to construct an initial feature vector.

[0040] This step transforms the original electromagnetic pulse signal into a structured feature vector through preprocessing, feature extraction, and anomaly correction of the original signal. This provides a standardized analytical object for subsequent correction of distance metrics based on energy decay rate parameter distribution and calculation of feature discrimination.

[0041] Step S120 in the method provided in this application embodiment includes:

[0042] Baseline calibration and bandpass filtering are performed on each raw signal to obtain the calibrated signal;

[0043] Identify continuous signal segments exceeding a preset energy threshold in each calibrated signal, and calculate the time difference between the start point of the first signal segment and the end point of the last signal segment as the arc duration;

[0044] Calculate the energy integral value of each calibrated signal over all continuous signal segments, and use it as the peak energy.

[0045] Calculate the reciprocal of the time it takes for the energy of each calibrated signal to decay from the global maximum value to a preset ratio, and use this as the energy decay rate.

[0046] The arc duration, peak energy, and energy decay rate extracted from all paths are summarized to form arc duration sequence, peak energy sequence, and energy decay rate sequence, respectively.

[0047] For the peak energy sequence and the energy decay rate sequence, calculate their respective quartiles, upper quartiles, and interquartile ranges;

[0048] Based on the quartiles, upper quartiles, and preset multiplication coefficients, calculate the upper and lower statistical limits of each parameter sequence;

[0049] In the peak energy sequence and energy decay rate sequence, measurement points whose values ​​exceed the corresponding statistical upper limit or fall below the corresponding statistical lower limit are marked as outliers.

[0050] For the outlier values, the moving median of the adjacent normal measurements of the corresponding parameter sequence is used for replacement and correction to generate a corrected parameter sequence;

[0051] Based on the corrected peak energy sequence, the corrected energy decay rate sequence, and the arc duration sequence, the initial feature vector is constructed in a preset order.

[0052] In this embodiment of the application, in order to remove interference noise and abnormal data in the original signal and build a standardized analysis basis, it is necessary to preprocess, extract features and correct data on multiple original signals to ensure the accuracy of subsequent feature discrimination calculation and weight optimization.

[0053] Specifically, baseline calibration and bandpass filtering are first performed on each raw signal. Baseline calibration aims to correct for reference offsets caused by sensor drift, environmental electromagnetic interference, and other factors during signal acquisition, ensuring a consistent measurement reference for signals acquired from different paths and at different times. Bandpass filtering, based on the inherent frequency characteristics of the contact arc electromagnetic pulse signal during circuit breaker opening, sets a filtering frequency band to retain effective signal components while filtering out irrelevant interference such as grid noise and equipment operating noise.

[0054] For example, taking into account the spectral characteristics of circuit breaker arc radiation signals, which are mostly concentrated in the 300MHz-1000MHz range, the passband range of the bandpass filter is set to 300MHz-1000MHz. This effectively filters out low-frequency power grid interference and high-frequency environmental noise outside this frequency band, thereby obtaining a calibrated signal that better reflects the true state of the arc.

[0055] Furthermore, after obtaining the calibrated signal, key feature parameters are extracted for each signal.

[0056] Specifically, firstly, continuous signal segments exceeding a preset energy threshold are identified. These continuous signal segments directly correspond to the discharge process of the contact arc when the circuit breaker trips, and their start and end times are directly related to the contact state and arc-extinguishing performance of the contacts. By locating the start point of the first continuous signal segment and the end point of the last continuous signal segment, the time difference between the two can be calculated to obtain the arc duration. This parameter can intuitively reflect the duration of the arc discharge; the more severe the contact degradation, the longer the arc duration tends to be.

[0057] The preset energy threshold is determined by combining the rated voltage, current and typical energy range of the circuit breaker and the opening arc, while also referring to the normal signal energy distribution in historical monitoring data.

[0058] Furthermore, the energy integral value of each calibrated signal over all continuous signal segments is calculated as the peak energy. This parameter reflects the intensity of the arc discharge and is closely related to state parameters such as contact resistance and discharge gap.

[0059] At the same time, the reciprocal of the time it takes for the signal energy to decay from the global maximum value to a preset proportion (e.g., 50%) is calculated as the energy decay rate. This parameter reflects the efficiency of arc extinguishing. Whether the mechanical transmission mechanism is smooth and whether the arc extinguishing device is effective will affect the magnitude of the energy decay rate.

[0060] For example, in a calibrated signal, a continuous signal segment exceeding the energy threshold starts at 0.01 seconds and ends at 0.07 seconds, corresponding to an arc duration of 0.06 seconds. By integrating the energy within this signal segment, the peak energy is calculated to be 150 μJ. The energy decays from its maximum value to 50% in 0.012 seconds, corresponding to an energy decay rate of approximately 83.3 s. -1 .

[0061] Furthermore, after completing the feature extraction of a single signal, the arc duration, peak energy, and energy decay rate of all paths are collected and extracted to form arc duration sequence, peak energy sequence, and energy decay rate sequence, respectively.

[0062] For example, signals are acquired through a 4-channel antenna array. For each signal, one arc duration, one peak energy, and one energy decay rate are extracted. The results are then combined to form an arc duration sequence, a peak energy sequence, and an energy decay rate sequence containing 4 data points.

[0063] Further data anomaly identification and correction operations were carried out on the compiled peak energy sequence and energy decay rate sequence.

[0064] First, calculate the quartiles, upper quartiles, and interquartile ranges for each of the two sequences. Specifically, the quartiles are the three dividing points that divide the sequence data into four equal parts after sorting it from smallest to largest. The upper quartile corresponds to the dividing point where 75% of the data is located. The interquartile range is the difference between the upper quartile and the lower quartile (the dividing point where 25% of the data is located), which is an indicator of the dispersion of the data and is not easily affected by extreme values.

[0065] Furthermore, based on the calculated quartiles, upper quartiles, and preset multiplication factors (determined based on historical monitoring data of circuit breakers and engineering practice experience, usually set to 1.5), the statistical upper and lower limits of each parameter sequence are calculated. The statistical upper limit is the upper quartile plus 1.5 times the interquartile range, and the statistical lower limit is the lower quartile minus 1.5 times the interquartile range. This range effectively defines the distribution interval of normal data; measurement points exceeding this range are likely outliers.

[0066] For example, if the lower quartile of a peak energy sequence is 100 μJ, the upper quartile is 180 μJ, and the interquartile range is 80 μJ, and the upper limit of the statistical calculation is calculated using a multiplier of 1.5, the upper limit of the statistical calculation is 180 + 1.5 × 80 = 300 μJ, and the lower limit of the statistical calculation is 100 - 1.5 × 80 = -20 μJ (since the energy value cannot be negative, the actual lower limit is taken as 0 μJ). The measurement point of 320 μJ in the sequence will be marked as an outlier.

[0067] For identified outliers, the moving median of their nearest normal measurements in the parameter sequence is used for replacement and correction. This moving median correction method effectively reduces the impact of extreme outliers on the overall trend of the sequence while preserving local correlations. Compared to simple mean replacement, it better maintains the consistency of the sequence data.

[0068] For example, in a certain energy decay rate sequence, a certain outlier is 150s. -1 Its three adjacent normal measurements were 90s. -1 85s -1 92s -1 Take the median of these three values, 85s -1 Replace the outlier value to generate a corrected energy decay rate sequence.

[0069] Finally, based on the corrected peak energy sequence and the corrected energy decay rate sequence, combined with the arc duration sequence that does not require outlier processing, all effective feature data are integrated in a preset order (arc duration sequence - peak energy sequence - energy decay rate sequence) to construct an initial feature vector.

[0070] For example, the integrated initial feature vector is [0.06s, 0.058s, 0.062s, 0.059s, 150μJ, 145μJ, 152μJ, 148μJ, 83.3s]. -1 85s -1 81.2s -1 84.5s -1 This vector transforms scattered multi-path, multi-dimensional feature parameters into structured analysis objects, retaining the core information that reflects the circuit breaker status, and ensuring data quality through preprocessing and anomaly correction, laying the foundation for subsequent correction of distance metrics and calculation of feature discrimination based on energy decay rate parameter distribution.

[0071] S130: Based on the parameter distribution of the energy decay rate, correct the preset distance metric, and calculate the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector, as the current feature discrimination.

[0072] In this embodiment of the application, in order to separate the feature information corresponding to contact degradation and mechanical defects and quantify the degree of difference between the two types of fault features, it is necessary to combine the parameter distribution characteristics of energy decay rate to correct the distance metric, and then calculate the statistical distance between the two feature subsets to improve the pertinence of subsequent weight optimization and the reliability of health status assessment.

[0073] Specifically, firstly, based on the fault feature mapping table, the first feature subset corresponding to contact degradation and the second feature subset corresponding to mechanical defects are separated from the initial feature vector, thereby clarifying the feature dimensions corresponding to the two types of faults.

[0074] Furthermore, the statistical variance and skewness of the energy decay rate sequence are calculated. The statistical variance reflects the dispersion of the energy decay rate data, while the skewness reflects the asymmetry of the data distribution. The product of the two can comprehensively characterize the parameter distribution characteristics of the energy decay rate. This product is used as a distribution correction factor to adapt to the signal distribution differences of circuit breakers under different operating conditions.

[0075] Subsequently, the distribution correction factor is multiplied by the basic weight coefficient to generate an adaptive weighting coefficient. This coefficient dynamically adjusts the importance of each feature dimension according to the actual distribution of the energy decay rate. Using this adaptive weighting coefficient, the weights of each dimension in the preset distance metric are adjusted item by item to obtain an adaptive distance metric that better reflects the distribution characteristics of the current data.

[0076] Furthermore, using the adjusted adaptive distance metric, the cluster centers of the first and second feature subsets in the feature space are calculated respectively. The cluster centers can centrally reflect the distribution location of each type of feature. The Euclidean distance between the two cluster centers is calculated to intuitively reflect the degree of separation between the two types of features in space. Then, it is divided by the sum of the maximum feature dispersion within each of the first and second feature subsets to eliminate the influence of the data dispersion within the subsets on the distance quantization. Finally, the calculation result is normalized to the interval [0, 1], and the output is the current feature discrimination score.

[0077] This step defines the boundary between two types of fault features, contact degradation and mechanical defects, by dynamically correcting the distance metric and calculating the feature discrimination. This provides direction for subsequent antenna weight optimization aimed at maximizing feature discrimination, ensuring that the optimized feature vector can be applied more accurately to fault type identification and health status assessment.

[0078] Step S130 in the method provided in this application embodiment includes:

[0079] According to the fault feature mapping table, a first feature subset corresponding to contact degradation and a second feature subset corresponding to mechanical defects are separated from the initial feature vector;

[0080] Calculate the statistical variance and skewness of the energy decay rate sequence, and obtain the distribution correction factor based on their product;

[0081] The distribution correction factor is multiplied by the basic weight coefficient to generate an adaptive weighting coefficient;

[0082] Using the aforementioned adaptive weighting coefficients, the weights of each dimension in the preset distance metric are adjusted item by item to obtain an adaptive distance metric;

[0083] Using the adaptive distance metric, the cluster centers of the first feature subset and the second feature subset in the feature space are calculated respectively;

[0084] Calculate the Euclidean distance between the two cluster centers, divide it by the sum of the maximum feature dispersion within each of the first and second feature subsets, normalize the result to the [0, 1] interval, and output the current feature discrimination score.

[0085] In this embodiment of the application, in order to avoid feature discrimination bias caused by using a fixed distance metric, the distance metric needs to be dynamically corrected in combination with the parameter distribution characteristics of the energy decay rate, and then the feature discrimination degree is obtained through statistical calculation, so as to improve the pertinence of subsequent weight optimization and the reliability of health status assessment.

[0086] Specifically, feature subset separation is first performed based on the fault feature mapping table. This table is constructed based on statistical analysis of electromagnetic signal characteristics from a large number of historical circuit breakers under different fault states, clarifying the correlation weights between each feature parameter and specific fault modes. For example, the correlation weight between the arc duration parameter and contact degradation is higher than that between mechanical defects, and the correlation weight between the energy decay rate parameter and mechanical defects has a specific threshold range.

[0087] By using the fault feature mapping table, feature dimensions directly related to contact degradation can be selected from the initial feature vector to form the first feature subset; at the same time, feature dimensions corresponding to mechanical defects can be separated to form the second feature subset, ensuring that the preliminary division of the two types of fault features has clear physical meaning and data support.

[0088] Furthermore, the statistical variance and skewness of the energy decay rate sequence are calculated. Specifically, the statistical variance reflects the dispersion of the energy decay rate data; the more dispersed the data, the larger the variance. For example, the variance of the energy decay rate sequence of a healthy circuit breaker is relatively small, while the variance of the energy decay rate sequence of a circuit breaker with mechanical jamming will be significantly larger.

[0089] Furthermore, skewness reflects the asymmetry of the data distribution. Positive skewness indicates a significant long tail on the right side of the data, while negative skewness corresponds to a long tail on the left side. The skewness of the energy decay rate sequence shows regular differences under different fault types. Multiplying the statistical variance by the skewness yields a distribution correction factor that comprehensively characterizes the distribution features of the energy decay rate parameter. This distribution correction factor can dynamically adapt to the signal distribution differences of circuit breakers under different operating conditions, providing a basis for subsequent distance measurement correction.

[0090] Furthermore, the distribution correction factor is multiplied by the basic weight coefficients to generate adaptive weighting coefficients. The basic weight coefficients are based on the preset physical importance of the feature parameters. For example, the basic weight of peak energy for fault judgment is higher than that of other auxiliary features. The distribution correction factor dynamically adjusts this weight according to the current signal distribution, so that the final adaptive weighting coefficients can reflect both the inherent importance of the feature itself and adapt to the actual distribution characteristics of the current data.

[0091] Furthermore, the obtained adaptive weighting coefficients are used to adjust the weights of each dimension in the preset distance metric item by item (i.e., each dimension weight is multiplied by its corresponding adaptive weighting coefficient) to obtain an adaptive distance metric. For example, if the initial weight of a certain feature dimension is 0.3 and the corresponding adaptive weighting coefficient is 1.2, then the adjusted weight is 0.36, making the distance metric more consistent with the distribution pattern of the current data and avoiding misjudgment of feature differences caused by fixed weights.

[0092] Furthermore, using the adjusted adaptive distance metric, the cluster centers of the first and second feature subsets in the feature space are calculated respectively. Specifically, the adaptive distance metric, as a spatial measure, uses an iterative optimization algorithm to find the geometric median of all feature vectors within each of the two feature subsets. This geometric median is defined as the position that minimizes the sum of the adaptive distances from all feature vectors within the subset to that point, thus centrally reflecting the core distribution locations of various features.

[0093] Among them, the iterative optimization algorithm can adopt the coordinate descent method, that is, fix the coordinates of other dimensions of the candidate cluster center, and only adjust and optimize the coordinates of the current dimension. By iteratively reducing the sum of the adaptive distances of all feature vectors in the subset to the candidate center in a dimensional loop, the convergent and stable geometric median is finally obtained.

[0094] For example, the first feature subset contains multiple feature vectors reflecting contact degradation, each vector covering multiple dimensions of data such as arc duration and peak energy. When using the coordinate descent method, firstly, the coordinates of other dimensions such as peak energy are fixed, and only the candidate center values ​​of the arc duration dimension are optimized to reduce the sum of distances to the optimal value for the current dimension; then, the arc duration dimension is fixed, and the coordinates of the peak energy dimension are optimized, and this process is repeated iteratively. After multiple rounds of adjustment, the point that minimizes the sum of adaptive distances of all vectors is found, which is the cluster center of the first feature subset.

[0095] Furthermore, the Euclidean distance between the two cluster centers is calculated. This distance reflects the degree of separation between the two types of features in space; a larger distance indicates a more significant difference between the two types of fault features. Subsequently, this distance is divided by the sum of the maximum feature dispersion within each of the first and second feature subsets to eliminate the influence of data dispersion within the subsets on the distance quantization and ensure the objectivity of the calculation results.

[0096] For example, the maximum feature dispersion of the first feature subset is 0.8, the maximum feature dispersion of the second feature subset is 0.6, and the sum of the two is 1.4. If the Euclidean distance of the cluster centers is 1.2, then the calculation result is 1.2 / 1.4≈0.857.

[0097] Finally, the calculation results are normalized to the interval [0, 1], and the output result is the current feature discrimination degree. The closer the value is to 1, the clearer the distinction between the two types of features, contact degradation and mechanical defects. Conversely, the lower the discrimination degree, the more the two types of fault features overlap in the feature space, which may lead to misjudgment of fault type.

[0098] S140: With the goal of maximizing the current feature discrimination, the spatial sensing weights of the broadband microstrip antenna array are iteratively adjusted through a gradient-guided weight optimization algorithm to generate an optimized feature vector;

[0099] In this embodiment of the application, in order to enhance the difference between the two types of features, contact degradation and mechanical defects, and improve the ability of the feature vector to identify the fault state, it is necessary to dynamically adjust the antenna spatial sensing weights through gradient-guided iterative methods to generate optimized feature vectors, so as to ensure the diagnostic accuracy and reliability of the subsequent multi-state classifier.

[0100] Specifically, the initial weights of each antenna in the broadband microstrip antenna array are first used to perform weighted fusion with multiple original signals. The initial weights are preset based on the antenna installation position and radiation characteristics to ensure that the fused signal set can initially cover electromagnetic signal information in various dimensions, thus generating a weighted signal set.

[0101] Furthermore, based on the weighted signal set, the trial feature vector under the current weight is obtained according to the same feature extraction process, and the trial feature discrimination corresponding to the trial feature vector is obtained through the same calculation method, so as to quantify the separation effect of the two types of fault features under the current weight configuration.

[0102] Subsequently, the gradient direction and gradient magnitude of the trial feature discrimination with respect to each antenna weight are calculated. The gradient direction indicates the direction in which adjusting the antenna weights improves the feature discrimination, while the gradient magnitude reflects the sensitivity of the weight changes to the discrimination, providing a quantitative basis for weight updates.

[0103] Furthermore, based on the obtained gradient direction and gradient magnitude, and combined with the preset learning rate, the weights of each antenna are updated in the direction that increases the feature discrimination, so as to enhance the sensitivity of sensing contact degradation and mechanical defect features.

[0104] Furthermore, the iterative process of weighted fusion, feature extraction, discrimination calculation and weight update is repeated. After each iteration, a new weight configuration and corresponding trial feature discrimination are generated until the growth rate of the trial feature discrimination between two adjacent iterations is lower than the preset convergence threshold, indicating that the weight adjustment has reached the optimal effect. The iteration is then stopped and the optimal weight configuration is obtained.

[0105] Finally, the original signals from multiple sources are weighted and fused using the optimal weight configuration. This gives higher weights to antenna signals with high contributions, highlighting key fault characteristics, suppressing irrelevant interference, and generating optimized feature vectors. This step, by dynamically optimizing antenna weights, further amplifies the differences between the two types of fault characteristics, providing more discriminative feature inputs for subsequent health status assessments and effectively improving the accuracy of the detection method.

[0106] Step S140 in the method provided in this application embodiment includes:

[0107] The initial weights of each antenna in the broadband microstrip antenna array are used to perform weighted fusion with the multiple original signals to generate a weighted signal set;

[0108] Based on the weighted signal set, the trial feature vector under the current weight is obtained, and the corresponding trial feature discrimination is calculated.

[0109] Calculate the gradient direction and gradient magnitude of the discriminative power of the trial feature with respect to the weights of each antenna;

[0110] Based on the gradient direction and gradient magnitude, and combined with the preset learning rate, the weights of each antenna are updated in the direction that increases the feature discrimination.

[0111] Repeat the iterative process of weighted fusion, feature extraction, discrimination calculation and weight update until the growth rate of the trial feature discrimination is lower than the preset convergence threshold, and obtain the optimal weight configuration;

[0112] The optimal weight configuration is used to perform final weighted fusion of the multiple original signals to generate the optimized feature vector.

[0113] In this embodiment of the application, in order to enhance the distinguishability between two types of features, namely contact degradation and mechanical defects, it is necessary to maximize the feature discrimination and dynamically adjust the antenna spatial sensing weights through gradient-guided iterative methods to optimize the feature vector quality, so as to solve the problem that the initial weights are difficult to adapt to the signal distribution and key fault features are masked.

[0114] Specifically, the initial weights of each antenna in the broadband microstrip antenna array are first used to perform weighted fusion with the multiple original signals. The initial weights are preset based on the antenna's installation location, radiation characteristics, and the propagation law of the circuit breaker arc signal. For example, antennas closer to areas with strong arc radiation have slightly higher initial weights to ensure that the fused signal set can initially cover the effective information in all dimensions, generating the weighted signal set.

[0115] Furthermore, based on the weighted signal set, following the same preprocessing and feature extraction process as the initial feature vector, baseline calibration, bandpass filtering, feature parameter extraction, and outlier correction are performed to obtain the trial feature vector under the current weight.

[0116] Similarly, using the same feature separation steps, two types of fault feature subsets are separated from the trial feature vector, and the corresponding trial feature discrimination is calculated to quantify the separation effect of the two types of features under the current weight configuration, providing a basis for weight adjustment.

[0117] Furthermore, the gradient direction and magnitude of the trial feature discrimination with respect to each antenna weight are calculated to capture the influence of weight changes on feature discrimination performance, providing a clear quantitative basis for subsequent updates of antenna weights along the optimal direction.

[0118] The method provided in this application embodiment calculates the gradient direction and gradient magnitude of the trial feature discrimination with respect to each antenna weight, including:

[0119] Keeping the weights of the other antennas in the broadband microstrip antenna array unchanged, a small positive perturbation is added to the current weight of the target antenna, and the discrimination after the first perturbation is recalculated;

[0120] Add the same small negative perturbation to the current weights of the target antenna, and recalculate the discrimination after the second perturbation;

[0121] Subtract the second perturbation-based discrimination from the first perturbation-based discrimination, and then divide by twice the small perturbation value. The result is used as an approximate value of the partial derivative of the target antenna weight.

[0122] By iterating through each antenna in the broadband microstrip antenna array and repeatedly performing the perturbation calculation process, a set of partial derivative values ​​for all antenna weights is obtained.

[0123] The vector direction of the set of partial derivative values ​​is taken as the gradient direction, and the absolute value of each partial derivative is taken as the gradient magnitude.

[0124] Specifically, the target antenna to be calculated is first determined, while keeping the weights of the other antennas in the broadband microstrip antenna array unchanged. Only the current weight of the target antenna is perturbed. The perturbation value is chosen as a constant with an extremely small absolute value, such as 1 × 10⁻⁶. -6This value ensures that the discrimination after disturbance produces detectable differences, while preventing the calculation results from deviating from the true gradient characteristics due to excessive disturbance, thus meeting the needs of fine calculation of the electromagnetic signal characteristics of circuit breakers.

[0125] Furthermore, after adding a selected small positive perturbation to the current weights of the target antenna, the complete process of weighted fusion, feature extraction, and discrimination calculation is re-executed to obtain the discrimination after the first perturbation.

[0126] For example, the target antenna currently has a weight of 0.4 (antenna weights typically range from [0, 1], used to characterize the contribution ratio of each antenna signal, ensuring that the signal energy after weighted fusion is within a reasonable range and avoiding numerical overflow), and 1×10 is added. -6 After the positive perturbation, the weight becomes 0.400001. Based on this weight and other antenna weights, the multiple original signals are weighted and fused. Then, the trial feature vector is extracted and the discrimination is calculated according to the established process, which is the discrimination after the first perturbation.

[0127] Furthermore, a small negative perturbation of the same value is added to the current weight of the target antenna, making the target antenna weight 0.399999. The weighted fusion, feature extraction, and discrimination calculation process is repeated to obtain the discrimination score after the second perturbation. By using symmetrical perturbation, the interference of other irrelevant factors on the discrimination score can be effectively canceled, making the calculation results closer to the true gradient change.

[0128] Furthermore, the obtained discrimination after the first perturbation is subtracted from the discrimination after the second perturbation, and then divided by twice the small perturbation value. This calculation process is essentially an approximation of the partial derivative of the trial feature discrimination with respect to the target antenna weight by using the finite difference method.

[0129] For example, the discrimination index after the first perturbation is 0.85, the discrimination index after the second perturbation is 0.84998, and the value of the small perturbation is 1×10. -6 Then the approximate value of the partial derivative is (0.85-0.84998) / (2×1×10). -6 =10, which intuitively reflects the magnitude and direction of the change in the discriminative power of the probe feature when the target antenna weight unit changes.

[0130] Similarly, following the above process, each antenna in the broadband microstrip antenna array is traversed, and perturbation calculations are performed on each antenna as a target antenna to obtain a set of partial derivative values ​​corresponding to all antenna weights. For example, if the array contains 4 antennas, after traversal, 4 partial derivative values ​​are obtained (e.g., 10, 8.5, 6.2, 9.1), forming a complete set of partial derivative values.

[0131] Finally, the vector direction of the partial derivative set is determined as the gradient direction, which indicates the optimal path for adjusting the antenna weights to maximize the discriminative power of the probe features. The absolute value of each partial derivative is used as the gradient magnitude. The larger the gradient magnitude, the more sensitive the corresponding antenna weights are to the discriminative power of the features, and the larger the adjustment step size can be adapted in subsequent weight updates.

[0132] For example, the partial derivative value of a certain antenna is 10, and its gradient magnitude is 10, which is greater than the gradient magnitude of other antennas. This indicates that adjusting the weight of this antenna is more critical to improving feature discrimination.

[0133] This step quantifies the correlation between the antenna weights and the discriminative power of the trial features through rigorous symmetrical perturbation and numerical calculation. This provides a clear basis for subsequent updates of the antenna weights with a preset learning rate, ensuring that the weight optimization process can proceed efficiently in the optimal direction, thereby enhancing the differentiation effect between two types of features: contact degradation and mechanical defects.

[0134] S150: Input the optimized feature vector into a pre-trained multi-state classifier, map the output contact health index and mechanical health index, and perform health status assessment and early warning based on adaptive threshold fusion judgment.

[0135] In this embodiment of the application, in order to distinguish between the health status of contact degradation and mechanical defects, and to achieve early warning and identification of faults, it is necessary to analyze and optimize the feature vector through a pre-trained multi-state classifier, and then combine it with adaptive threshold fusion judgment to improve the accuracy of health status assessment and the timeliness of early warning, so as to provide a reliable decision basis for equipment operation and maintenance.

[0136] Specifically, the process first relies on a pre-trained multi-state classifier to process and optimize the feature vectors, so as to accurately separate and quantify the health status of the contact and mechanical system.

[0137] In the method provided in this application embodiment, the pre-training step of the multi-state classifier includes:

[0138] A large number of electromagnetic signal samples from historical circuit breakers under known contact and mechanical conditions were collected, processed to obtain an optimized feature vector set of the samples, and labeled with the corresponding real contact health label and mechanical health label.

[0139] Based on the fault feature mapping table, the optimized feature vectors of each sample are divided into a contact-related feature subset and a mechanical-related feature subset;

[0140] A network architecture for a multi-state classifier is constructed, wherein the input layer of the network architecture is configured as a dual-branch structure, the first branch channel receives the contact-related feature subset, the second branch channel receives the mechanical-related feature subset, and the output layer of the network architecture is configured with dual output nodes to generate contact health prediction value and mechanical health prediction value respectively.

[0141] The multi-state classifier is trained under supervision using the sample optimization feature vector set and the corresponding health labels until the combined error between the health prediction value output by the model and the true label is lower than a preset accuracy threshold, thus completing the model pre-training.

[0142] Specifically, a large number of electromagnetic signal samples were first collected from historical circuit breakers under known contact and mechanical conditions. The samples needed to cover circuit breakers in different health conditions, including various combinations of scenarios such as low, medium, and high contact degradation, normal mechanical condition, slight jamming, and severe defects. They also needed to cover circuit breakers of different models, rated voltages, and rated currents to ensure the diversity and representativeness of the samples.

[0143] During the data collection process, the actual contact condition and mechanical condition of each circuit breaker are recorded simultaneously. Professional testing equipment is used to determine the degree of contact degradation, wear of mechanical transmission mechanisms, and other real information, which serves as the basis for labeling.

[0144] The collected electromagnetic signal samples are preprocessed, feature extracted, and weighted according to the same processing flow as the actual detection. This involves sequentially performing operations such as baseline calibration, bandpass filtering, extraction of parameters such as arc duration and peak energy, outlier correction, and gradient-guided weight adjustment to finally obtain the optimized feature vector set of the samples.

[0145] Simultaneously, based on the actual test results, corresponding labels for actual contact health and mechanical health are labeled. The labels are represented by values ​​in the range of 0-1, with the closer the value is to 1, the better the health condition. For example, a label of 0.95 is used when there is no contact degradation, and a label of 0.3 is used when there is severe mechanical jamming.

[0146] Furthermore, the optimized feature vectors of each sample are divided according to the fault feature mapping table. Based on the fault feature mapping table, feature dimensions related to contact degradation are selected from the optimized feature vectors of each sample to form a contact-related feature subset; feature dimensions related to mechanical defects are selected to form a mechanical-related feature subset, providing targeted data for the input of the dual-branch network architecture.

[0147] Subsequently, a multi-state classifier network architecture was constructed. Considering the identification requirements of the two types of fault features, a dual-branch parallel network structure was adopted. Specifically, the input layer is divided into two independent branch channels. The first branch channel receives a subset of contact-related features, and the second branch channel receives a subset of mechanical-related features, ensuring that the two types of features can be learned separately and specifically.

[0148] In addition, the intermediate layers of the network adopt an alternating structure of fully connected layers and Batch Norm layers. The fully connected layers are used to explore the non-linear correlations between features, while the Batch Norm layers are used to normalize the feature distribution and accelerate training convergence. At the same time, Dropout layers are set to suppress overfitting.

[0149] For example, the sample optimization feature vector has a dimension of 20, and the contact-related feature subset and the mechanical-related feature subset each have 10 dimensions. An intermediate structure containing 3 fully connected layers can be constructed, with the number of neurons in each layer set to 64, 32 and 16 respectively, and the Dropout probability set to 0.2.

[0150] Meanwhile, the output layer is configured with dual output nodes, which use the Sigmoid activation function to output the contact health prediction value and the mechanical health prediction value in the 0-1 range respectively, so as to achieve synchronous quantitative output of the two types of health status.

[0151] Furthermore, the network architecture is trained under supervised supervision by using a sample-optimized feature vector set and corresponding health labels. Before training, the sample set is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for iterative updates of model parameters, the validation set is used to monitor model performance during training, and the test set is used to finally evaluate the model's generalization ability.

[0152] During training, a subset of sample features is input into the corresponding branch channel. After the model outputs a health prediction value, the mean squared error is calculated as the loss function to quantify the difference between the predicted value and the true label. Then, the network weights and bias parameters are adjusted based on backpropagation using the gradient descent algorithm. During training, the overall error on the validation set also needs to be dynamically monitored. When the overall error no longer decreases after 5 consecutive training rounds and falls below a preset accuracy threshold (e.g., 0.01), training is stopped.

[0153] For example, after a certain round of training, the mean squared error of the contact health prediction on the validation set is 0.008, the mean squared error of the mechanical health prediction is 0.007, and the combined error is 0.0075, which is lower than the preset threshold of 0.01, indicating that the model has a stable prediction ability.

[0154] In addition, if overfitting signs such as an increase in validation set error appear during training, the Dropout probability can be appropriately increased or the learning rate can be decreased, for example, the learning rate can be adjusted from 0.001 to 0.0005. If the prediction error of a certain type of health status is consistently high, such as the prediction error of mechanical health status being consistently above 0.02, sample data related to mechanical defects can be supplemented, such as adding 200 sets of electromagnetic signal samples of different types of mechanical faults, and reintegrating them into the training set for continued training until the overall error meets the requirements.

[0155] Finally, the multi-state classifier obtained through the above pre-training steps can parse and optimize the two types of fault feature information in the feature vector, and stably output the contact health index and mechanical health index, providing reliable model support for subsequent adaptive threshold fusion judgment and health status early warning.

[0156] Furthermore, after pre-training the multi-state classifier, the optimized feature vectors generated in actual detection are input into the pre-trained model. The classifier performs nonlinear transformations and calculations on the two types of features respectively through the learned feature mapping relationship, and finally outputs the contact health index and mechanical health index in the 0-1 interval simultaneously.

[0157] For example, in a certain detection scenario, the output contact health index is 0.82, indicating that the contact degradation is relatively mild; the mechanical health index is 0.45, indicating that there is a significant defect in the mechanical system. The two indicators quantify the health status of the corresponding system.

[0158] Furthermore, based on the output health index, an adaptive threshold fusion judgment is made to achieve dynamic health status assessment and early warning with accurate identification and trend prediction of contact degradation and mechanical defects, thereby avoiding sudden equipment failures in advance.

[0159] The method provided in this application embodiment, which uses adaptive threshold fusion judgment to achieve health status assessment and early warning, includes:

[0160] Based on the initial warning threshold, a threshold particle set is constructed, which includes the current detection cycle threshold, the previous detection cycle threshold, and the historical average threshold.

[0161] Based on the temporal changes of the contact health index and the mechanical health index, the trend of health index changes in the current detection cycle is calculated as the optimization direction for the current threshold particles.

[0162] Extract the current feature discrimination scores from multiple historical detection periods and calculate the average score as the cost for the current detection period;

[0163] The fusion radius of the current threshold particle is calculated based on the cost value, and adjacent threshold particles within the fusion radius are selected from the set of threshold particles.

[0164] Calculate the variance of the adjacent threshold particles and the current threshold particle on each health index sequence as a comprehensive fluctuation measure, obtain the fusion weight of each particle, perform weighted fusion on the adjacent threshold particles, and generate a preliminary optimized threshold.

[0165] When there are false intersections in the reverse extension lines of the health indicator change trends of multiple consecutive detection cycles, the preliminary optimization threshold is dynamically adjusted based on the predicted deterioration time corresponding to the false intersection, and an adaptive threshold for this round of early warning judgment is generated.

[0166] The contact health index and the mechanical health index are compared with their corresponding adaptive thresholds. If either index is lower than the corresponding adaptive threshold, a corresponding warning is triggered.

[0167] Specifically, a threshold particle set is first constructed based on a preset initial warning threshold. The initial warning threshold is determined according to the circuit breaker's rated parameters, operating standards, and historical fault data. For example, the initial warning threshold for contact health is set to 0.7, and the initial warning threshold for mechanical health is set to 0.6.

[0168] In addition, the threshold particle set covers the current detection period threshold, the previous detection period threshold, and the historical average threshold. For example, the current period threshold is 0.68, the previous period threshold is 0.72, and the historical average threshold for 10 periods is 0.71. By aggregating the thresholds from multiple time periods, a reference basis is provided for subsequent threshold optimization.

[0169] Furthermore, based on the temporal changes of contact health indicators and mechanical health indicators, the trend of health indicator changes in the current testing cycle is calculated. Specifically, by analyzing indicator data from multiple consecutive testing cycles, it is determined whether the indicators show a continuous downward trend, a stable fluctuation, or a gradual upward trend.

[0170] For example, if the contact health index drops from 0.85 to 0.82 and then to 0.80 within three consecutive cycles, and the mechanical health index drops from 0.50 to 0.45 and then to 0.42, it is determined that both types of indicators show a continuous deterioration trend. This trend is the optimization direction for the current threshold particle, ensuring that the threshold adjustment can fit the actual changes in the equipment's condition.

[0171] Furthermore, multiple sets of current feature discrimination scores from various historical detection periods are extracted. For example, feature discrimination score data from the past 15 detection periods are selected, and their arithmetic mean is calculated as the cost value for the current detection period. The cost value reflects the average discrimination level between the two types of fault features in historical detections; a higher value indicates better historical feature discrimination performance.

[0172] Furthermore, the fusion radius of the current threshold particles is calculated based on the cost value to screen out neighboring threshold particles that are strongly correlated with the current threshold, thus avoiding interference from irrelevant thresholds.

[0173] Specifically, the fusion radius can be obtained by calculating the ratio of a preset standard fusion radius to a cost value. The standard fusion radius is preset based on historical fault data of the circuit breaker and early warning accuracy requirements, for example, set to 0.1. This avoids irrelevant thresholds being mixed in due to an excessively large radius, or missing effective reference thresholds due to an excessively small radius, ensuring that the subsequently selected adjacent threshold particles can adapt to the current equipment health status assessment requirements. For example, if the standard fusion radius is 0.1 and the cost value is 0.8, then the current fusion radius is 0.125.

[0174] Furthermore, the Euclidean distances between the remaining threshold particles in the threshold particle set and the current threshold particle in the two-dimensional space formed by the contact health index sequence and the mechanical health index sequence are calculated, and each distance is compared with the fusion radius to filter out adjacent threshold particles whose distances are less than or equal to the fusion radius.

[0175] Specifically, the contact health threshold and mechanical health threshold of the current threshold particle are used as two-dimensional coordinate points, and the corresponding thresholds of the other threshold particles are used as comparison coordinate points. The straight-line distance between the two points is calculated using the Euclidean distance formula. This distance reflects the degree of threshold matching between particles.

[0176] For example, if the two-dimensional Euclidean distance between a neighboring threshold particle and the current particle is 0.11, which is less than the fusion radius of 0.125, it is included in the set of neighboring threshold particles to ensure that the selected particles have a strong correlation. Conversely, if the distance between a particle and the current particle is 0.15, which is greater than the fusion radius of 0.125, it is determined that its correlation is weak and is excluded to avoid irrelevant thresholds interfering with the subsequent fusion effect.

[0177] Furthermore, the variance of adjacent threshold particles and the current threshold particle in each health index sequence is calculated and used as a comprehensive fluctuation measure. At the same time, the fusion weight of each particle is obtained so that higher weights are given to threshold particles with small fluctuations and strong correlations, thereby achieving accurate weighted fusion of adjacent threshold particles.

[0178] The method provided in this application embodiment calculates the variance of the adjacent threshold particles and the current threshold particle on each health index sequence as a comprehensive fluctuation measure, and obtains the fusion weight of each particle, including:

[0179] Calculate the first variance of the adjacent threshold particles and the current threshold particles on the contact health index sequence, and the second variance on the mechanical health index sequence, respectively.

[0180] Calculate the average of the first variance and the second variance as a comprehensive volatility measure;

[0181] Divide the value 1 by the comprehensive volatility metric as the volatility compensation factor;

[0182] The Euclidean distance between the current threshold particle and each adjacent threshold particle on the contact health index sequence and the mechanical health index sequence is calculated as the inter-particle distance.

[0183] The product of the reciprocal of the distance between each particle and the fluctuation compensation factor is used as the initial fusion weight of the corresponding adjacent threshold particles.

[0184] The initial fusion weights of all adjacent threshold particles are normalized to obtain the final particle fusion weights.

[0185] First, the variances of adjacent threshold particles and the current threshold particle across the two health indicator sequences are calculated. Specifically, for the contact health indicator sequence, the indicator data of adjacent threshold particles and the current threshold particle are extracted over multiple consecutive detection cycles, and the dispersion of the two is calculated using the variance formula to obtain the first variance.

[0186] Similarly, the second variance is calculated based on the corresponding data of the mechanical health index sequence. For example, the detection data of adjacent threshold particles and the current threshold particle in the contact health index sequence are [0.72, 0.70, 0.68] and [0.71, 0.69, 0.67], respectively, and the first variance is calculated to be 0.0001; the detection data in the mechanical health index sequence are [0.61, 0.59, 0.57] and [0.60, 0.58, 0.56], respectively, and the second variance is calculated to be 0.0001. The two types of variances quantify the fluctuations in different dimensions.

[0187] Furthermore, the average of the first variance and the second variance is calculated as a comprehensive volatility measure. This comprehensive volatility measure fully reflects the overall volatility level of adjacent threshold particles and the current threshold particle across the two health status assessment dimensions; a smaller value indicates that the two states are closer and the stability is stronger.

[0188] For example, the average of the first variance and the second variance is 0.0001, that is, the comprehensive fluctuation measure is 0.0001, indicating that the overall fluctuation of the adjacent threshold particle and the current threshold particle is small and the state fit is high.

[0189] Furthermore, the ratio of the value 1 to the comprehensive volatility metric is used as a volatility compensation factor to adjust the weights of particles with different volatility levels. Specifically, the volatility compensation factor is inversely proportional to the comprehensive volatility metric; the smaller the comprehensive volatility metric, the larger the volatility compensation factor, and the higher the weight of the corresponding particle will be; conversely, the weight will decrease.

[0190] For example, when the overall volatility metric is 0.0001, the volatility compensation factor is 1 / 0.0001 = 10000. This factor amplifies the weight influence of stable particles and weakens the interference of highly volatile particles on the fusion result.

[0191] Furthermore, the Euclidean distance between the current threshold particle and each of its neighboring threshold particles on the two health index sequences is calculated as the inter-particle distance. The inter-particle distance reflects the spatial fit between neighboring threshold particles and the current threshold particle; a smaller distance indicates a more similar threshold configuration and a stronger correlation.

[0192] For example, the two index sequence points of the current threshold particle are (0.71, 0.60), and the corresponding point of a certain adjacent threshold particle is (0.72, 0.61). The distance between the particles is calculated to be 0.0141 using the Euclidean distance formula, which quantifies the spatial difference between the two.

[0193] Furthermore, the product of the reciprocal of the distance between each particle and the fluctuation compensation factor is used as the initial fusion weight for the corresponding adjacent threshold particles. This initial fusion weight considers both the correlation and stability between particles; the smaller the distance between particles and the larger the fluctuation compensation factor, the higher the initial fusion weight, indicating a greater contribution of that particle to threshold fusion. For example, the reciprocal of the particle distance 0.0141 is approximately 70.92, which, when multiplied by the fluctuation compensation factor 10000, yields an initial fusion weight of approximately 709200, highlighting the importance of this adjacent threshold particle.

[0194] Finally, the initial fusion weights of all adjacent threshold particles are normalized to obtain the final particle fusion weights. The normalization process divides each initial fusion weight by the sum of all initial fusion weights, ensuring that the final fusion weights are within the range of [0, 1] and that the sum of the final fusion weights of all particles is 1, thus ensuring the rationality and operability of the weight allocation.

[0195] For example, if there are three adjacent threshold particles with initial fusion weights of 709200, 638280, and 567360 respectively, and a total of 1914840, the final fusion weights obtained after normalization are 0.37, 0.33, and 0.30 respectively. By using these weights to perform weighted fusion on adjacent threshold particles, a preliminary optimized threshold that takes into account both stability and correlation can be generated.

[0196] Furthermore, after obtaining the final particle fusion weights, a weighted summation is performed based on the threshold values ​​of each adjacent threshold particle and its corresponding fusion weight. This weighted fusion of adjacent threshold particles generates a preliminary optimized threshold.

[0197] Specifically, the contact health threshold and mechanical health threshold of each adjacent threshold particle are multiplied by their final fusion weights, and then the weighted thresholds of all particles are summed to obtain the preliminary optimized contact health threshold and mechanical health threshold.

[0198] For example, the contact health thresholds of three adjacent threshold particles are 0.72, 0.70, and 0.69, respectively, with corresponding fusion weights of 0.37, 0.33, and 0.30. The preliminarily optimized contact health threshold after weighted fusion is 0.72×0.37+0.70×0.33+0.69×0.30≈0.70; the mechanical health thresholds are 0.61, 0.59, and 0.58, respectively. The preliminarily optimized mechanical health threshold after weighted fusion is 0.61×0.37+0.59×0.33+0.58×0.30≈0.59. This preliminarily optimized threshold integrates the advantages of each related particle and has strong stability.

[0199] Furthermore, the trend of health indicators over multiple consecutive detection cycles is analyzed. When there is a false intersection of the trend's reverse extension line, the preliminary optimization threshold is dynamically adjusted based on the predicted deterioration time corresponding to the false intersection, generating an adaptive threshold for this round of early warning judgment.

[0200] Specifically, the contact health index and mechanical health index of the past 5 testing cycles are extracted, and their respective trend lines are fitted. The two trend lines are extended in opposite directions. If there is an intersection, it is a dummy intersection. The deterioration time is estimated and predicted by the time point corresponding to the intersection.

[0201] For example, the contact health index for five consecutive cycles is [0.80, 0.78, 0.75, 0.72, 0.70], and the mechanical health index is [0.70, 0.68, 0.65, 0.62, 0.59]. The illusory intersection point after the trend line is extended in the opposite direction corresponds to a predicted degradation time of eight detection cycles. At this point, the threshold needs to be adjusted according to this predicted time. If the predicted degradation rate is relatively fast, the initially optimized contact health threshold is increased from 0.70 to 0.72, and the mechanical health threshold is increased from 0.59 to 0.61 to make the warning more sensitive. If the predicted degradation rate is slow, the threshold is maintained or slightly adjusted to ensure the accuracy of the warning.

[0202] Finally, the currently detected contact health index and mechanical health index are compared with their corresponding adaptive thresholds. If either index is lower than the corresponding adaptive threshold, a corresponding warning is triggered.

[0203] Specifically, the values ​​of both the contact health index and the mechanical health index are in the range of [0, 1]. The closer the value is to 1, the better the health status. If the value is lower than the adaptive threshold, it indicates that the equipment has a corresponding risk of failure.

[0204] For example, in the adaptive threshold, the contact health threshold is 0.72 and the mechanical health threshold is 0.61. If the currently detected contact health index is 0.75 (higher than the contact health threshold) and the mechanical health index is 0.58 (lower than the mechanical health threshold), then a mechanical defect warning is triggered.

[0205] In addition, if the current contact health index is 0.71 (below the contact health threshold) and the mechanical health index is 0.60 (below the mechanical health threshold), a dual warning for contact degradation and mechanical defects will be triggered simultaneously. The warning information will clearly indicate the fault type, providing maintenance personnel with targeted repair directions, thereby achieving accurate assessment and early warning of health status.

[0206] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects:

[0207] This application proposes a non-intrusive detection method for high-voltage circuit breakers based on electromagnetic signal analysis. First, a broadband microstrip antenna array deployed around the target circuit breaker synchronously captures electromagnetic pulse signals from the contact arc during the opening process, acquiring multiple raw signals to ensure a complete reflection of the arc discharge and equipment operation characteristics. Next, the raw signals undergo baseline calibration and bandpass filtering preprocessing to extract arc duration, peak energy, and energy decay rate. After outlier correction, an initial feature vector is constructed. Subsequently, based on the parameter distribution of the energy decay rate, a preset distance metric is corrected to separate two feature subsets: contact degradation and mechanical defects. Statistical distance is calculated to obtain the current feature discrimination, quantifying the differences between the two types of fault characteristics. Then, with the goal of maximizing feature discrimination, a gradient-guided weight optimization algorithm iteratively adjusts the antenna spatial perception weights to generate an optimized feature vector that enhances fault feature identification. Finally, the optimized feature vector is input into a pre-trained dual-branch multi-state classifier, outputting contact health and mechanical health indicators. By constructing a threshold particle set, fusing adjacent thresholds, and dynamically generating adaptive thresholds based on health trend prediction, accurate health status assessment and early warning are achieved.

[0208] The method provided in this application, through the technical solution of "signal acquisition - feature construction - discrimination calculation - weight optimization - classification warning", solves the problems of traditional detection methods such as reliance on offline fixed thresholds, low accuracy of single signal recognition, and susceptibility to interference. It avoids false and missed faults, improves the accuracy and timeliness of high-voltage circuit breaker status detection, and provides technical support for the refined operation and maintenance of power equipment.

[0209] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of the non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis provided in Embodiment 1, this application also provides a non-invasive detection system for high-voltage circuit breakers based on electromagnetic signal analysis, specifically including:

[0210] Electromagnetic signal acquisition module 01 is used to capture the contact arc electromagnetic pulse signal during the opening process of the target circuit breaker based on a broadband microstrip antenna array, and to acquire multiple raw signals.

[0211] Feature extraction and construction module 02 is used to preprocess the multiple original signals and extract the arc duration, peak energy and energy decay rate of each signal to construct an initial feature vector;

[0212] The feature discrimination calculation module 03 is used to correct the preset distance metric based on the parameter distribution of the energy decay rate, and calculate the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector as the current feature discrimination.

[0213] The weight optimization generation module 04 is used to generate an optimized feature vector by iteratively adjusting the spatial sensing weights of the broadband microstrip antenna array through a gradient-guided weight optimization algorithm with the goal of maximizing the current feature discrimination.

[0214] The health assessment and early warning module 05 is used to input the optimized feature vector into a pre-trained multi-state classifier, map the output contact health index and mechanical health index, and make judgments based on adaptive threshold fusion to realize health status assessment and early warning.

[0215] In one embodiment, the feature extraction building module 02 is further configured to:

[0216] Baseline calibration and bandpass filtering are performed on each original signal to obtain the calibrated signal. Continuous signal segments exceeding a preset energy threshold in each calibrated signal are identified, and the time difference between the start point of the first signal segment and the end point of the last signal segment is calculated as the arc duration. The energy integral value of each calibrated signal in all continuous signal segments is calculated as the peak energy. The reciprocal of the time it takes for the energy of each calibrated signal to decay from the global maximum value to a preset proportion is calculated as the energy decay rate. The arc duration, peak energy, and energy decay rate extracted from all paths are summarized to form the arc duration sequence, peak energy sequence, and energy decay rate sequence, respectively. For the peak energy sequence and the energy decay rate sequence, calculate their respective quartiles, upper quartiles, and interquartile ranges; based on the quartiles, upper quartiles, and preset multiplication coefficients, calculate the statistical upper limit and statistical lower limit of each parameter sequence; mark the measurement points in the peak energy sequence and the energy decay rate sequence whose values ​​exceed the corresponding statistical upper limit or fall below the corresponding statistical lower limit as outliers; for the outliers, replace and correct them using the sliding median of the adjacent normal measurement values ​​of the corresponding parameter sequence to generate a corrected parameter sequence; based on the corrected peak energy sequence, the corrected energy decay rate sequence, and the arc duration sequence, construct the initial feature vector in a preset order.

[0217] In one embodiment, the feature discrimination calculation module 03 is further used for:

[0218] According to the fault feature mapping table, a first feature subset corresponding to contact degradation and a second feature subset corresponding to mechanical defects are separated from the initial feature vector; the statistical variance and skewness of the energy decay rate sequence are calculated, and a distribution correction factor is obtained based on their product; the distribution correction factor is multiplied by the basic weight coefficient to generate an adaptive weighting coefficient; the adaptive weighting coefficient is used to adjust the weights of each dimension in the preset distance metric item by item to obtain an adaptive distance metric; the adaptive distance metric is used to calculate the cluster centers of the first feature subset and the second feature subset in the feature space; the Euclidean distance between the two cluster centers is calculated and divided by the sum of the maximum feature dispersions within each of the first feature subset and the second feature subset, and the calculation result is normalized to the [0, 1] interval, and the output is the current feature discrimination.

[0219] In one embodiment, the weight optimization generation module 04 is further configured to:

[0220] The initial weights of each antenna in the broadband microstrip antenna array are used to perform weighted fusion with the multiple original signals to generate a weighted signal set. Based on the weighted signal set, a trial feature vector under the current weight is obtained, and the corresponding trial feature discrimination is calculated. The gradient direction and gradient magnitude of the trial feature discrimination with respect to the weights of each antenna are calculated. According to the gradient direction and gradient magnitude, combined with a preset learning rate, the weights of each antenna are updated in the direction that increases the feature discrimination. The iterative process of weighted fusion, feature extraction, discrimination calculation, and weight update is repeated until the growth rate of the trial feature discrimination is lower than a preset convergence threshold, and the optimal weight configuration is obtained. The optimal weight configuration is used to perform final weighted fusion of the multiple original signals to generate the optimized feature vector.

[0221] Furthermore, the weight optimization generation module 04 also includes:

[0222] Keeping the weights of the other antennas in the broadband microstrip antenna array unchanged, a small positive perturbation is added to the current weight of the target antenna, and the discrimination after the first perturbation is recalculated; the same small negative perturbation is added to the current weight of the target antenna, and the discrimination after the second perturbation is recalculated; the discrimination after the first perturbation is subtracted from the discrimination after the second perturbation, and then divided by twice the small perturbation value, and the result is used as an approximate value of the partial derivative of the target antenna weight; the perturbation calculation process is repeated for each antenna in the broadband microstrip antenna array to obtain a set of partial derivative values ​​of all antenna weights; the vector direction of the set of partial derivative values ​​is used as the gradient direction, and the absolute value of each partial derivative is used as the gradient magnitude.

[0223] In one embodiment, the health assessment and early warning module 05 is also used for:

[0224] A large number of electromagnetic signal samples from historical circuit breakers under known contact and mechanical states are collected. These samples are processed to obtain an optimized feature vector set, which is then labeled with corresponding true contact health and mechanical health labels. Based on a fault feature mapping table, each optimized feature vector is divided into a contact-related feature subset and a mechanical-related feature subset. A multi-state classifier network architecture is constructed, wherein the input layer of the network architecture is configured with a dual-branch structure, with the first branch channel receiving the contact-related feature subset and the second branch channel receiving the mechanical-related feature subset. The output layer of the network architecture is configured with dual output nodes to generate predicted contact health and predicted mechanical health values, respectively. The multi-state classifier is then trained under supervised supervision using the optimized feature vector set and corresponding health labels until the combined error between the model's output health prediction value and the true label is lower than a preset accuracy threshold, thus completing the model's pre-training.

[0225] Furthermore, the health assessment and early warning module 05 also includes:

[0226] Based on the initial warning threshold, a threshold particle set is constructed, including the current detection cycle threshold, the previous detection cycle threshold, and the historical average threshold. The health index change trend for the current detection cycle is calculated based on the temporal changes of the contact health index and the mechanical health index, serving as the optimization direction for the current threshold particles. Multiple sets of current feature discrimination values ​​from several historical detection cycles are extracted, and their average is calculated as the cost value for the current detection cycle. The fusion radius of the current threshold particles is calculated based on the cost value, and adjacent threshold particles within the fusion radius are selected from the threshold particle set. The variance of the adjacent threshold particles and the current threshold particles on each health index sequence is calculated as a comprehensive fluctuation measure, and the fusion weight of each particle is obtained. The adjacent threshold particles are then weighted and fused to generate a preliminary optimized threshold. When there is a dummy intersection point on the reverse extension line of the health index change trends for multiple consecutive detection cycles, the preliminary optimized threshold is dynamically adjusted based on the predicted degradation time corresponding to the dummy intersection point, generating an adaptive threshold for this round of warning judgment. The contact health index and the mechanical health index are compared with their corresponding adaptive thresholds. If either index is lower than its corresponding adaptive threshold, a corresponding warning is triggered.

[0227] Furthermore, the health assessment and early warning module 05 also includes:

[0228] Calculate the first variance of the adjacent threshold particles and the current threshold particle on the contact health index sequence, and the second variance on the mechanical health index sequence; calculate the average of the first variance and the second variance as a comprehensive fluctuation measure; divide the value 1 by the comprehensive fluctuation measure as a fluctuation compensation factor; calculate the Euclidean distance between the current threshold particle and each adjacent threshold particle forming a point pair on the contact health index sequence and the mechanical health index sequence as the inter-particle distance; multiply the reciprocal of each inter-particle distance by the fluctuation compensation factor as the initial fusion weight of the corresponding adjacent threshold particle; normalize the initial fusion weights of all adjacent threshold particles to obtain the final particle fusion weights.

Claims

1. A non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis, characterized in that, The method includes: Based on the capture of electromagnetic pulse signals of contact arc during the opening process of the target circuit breaker using a broadband microstrip antenna array, multiple raw signals are obtained; The multiple raw signals are preprocessed, and the arc duration, peak energy and energy decay rate of each signal are extracted to construct an initial feature vector; Based on the parameter distribution of the energy decay rate, a preset distance metric is corrected, and the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector is calculated as the current feature discrimination, including: According to the fault feature mapping table, a first feature subset corresponding to contact degradation and a second feature subset corresponding to mechanical defects are separated from the initial feature vector; Calculate the statistical variance and skewness of the energy decay rate sequence, and obtain the distribution correction factor based on their product; The distribution correction factor is multiplied by the basic weight coefficient to generate an adaptive weighting coefficient; Using the aforementioned adaptive weighting coefficients, the weights of each dimension in the preset distance metric are adjusted item by item to obtain an adaptive distance metric; Using the adaptive distance metric, the cluster centers of the first feature subset and the second feature subset in the feature space are calculated respectively; Calculate the Euclidean distance between two cluster centers, divide it by the sum of the maximum feature dispersion within each of the first and second feature subsets, normalize the result to the [0, 1] interval, and output the current feature discrimination score. With the goal of maximizing the current feature discriminative power, a gradient-guided weight optimization algorithm iteratively adjusts the spatial sensing weights of the broadband microstrip antenna array to generate an optimized feature vector, including: The initial weights of each antenna in the broadband microstrip antenna array are used to perform weighted fusion with the multiple original signals to generate a weighted signal set; Based on the weighted signal set, the trial feature vector under the current weight is obtained, and the corresponding trial feature discrimination is calculated. Calculate the gradient direction and gradient magnitude of the discriminative power of the trial feature with respect to the weights of each antenna; Based on the gradient direction and gradient magnitude, and combined with the preset learning rate, the weights of each antenna are updated in the direction that increases the feature discrimination. Repeat the iterative process of weighted fusion, feature extraction, discrimination calculation and weight update until the growth rate of the trial feature discrimination is lower than the preset convergence threshold, and obtain the optimal weight configuration; The optimal weight configuration is used to perform final weighted fusion of the multiple original signals to generate the optimized feature vector. The optimized feature vector is input into a pre-trained multi-state classifier, which maps and outputs contact health index and mechanical health index. Based on adaptive threshold fusion judgment, health status assessment and early warning are realized. The multi-state classifier adopts a network architecture with a dual-branch input layer. The first branch channel receives the contact-related feature subset, and the second branch channel receives the mechanical-related feature subset. The output layer of the network architecture is configured with dual output nodes to generate contact health prediction value and mechanical health prediction value respectively.

2. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 1, characterized in that, The multiple raw signals are preprocessed, and the arc duration, peak energy, and energy decay rate of each signal are extracted, including: Baseline calibration and bandpass filtering are performed on each raw signal to obtain the calibrated signal; Identify continuous signal segments exceeding a preset energy threshold in each calibrated signal, and calculate the time difference between the start point of the first signal segment and the end point of the last signal segment as the arc duration; Calculate the energy integral value of each calibrated signal over all continuous signal segments, and use it as the peak energy. Calculate the reciprocal of the time it takes for the energy of each calibrated signal to decay from the global maximum value to a preset ratio, and use this as the energy decay rate. The arc duration, peak energy, and energy decay rate extracted from all paths are summarized to form arc duration sequence, peak energy sequence, and energy decay rate sequence, respectively.

3. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 1, characterized in that, The steps for constructing the initial feature vector include: For the peak energy sequence and the energy decay rate sequence, calculate their respective quartiles, upper quartiles, and interquartile ranges; Based on the quartiles, upper quartiles, and preset multiplication coefficients, calculate the upper and lower statistical limits of each parameter sequence; In the peak energy sequence and energy decay rate sequence, measurement points whose values ​​exceed the corresponding statistical upper limit or fall below the corresponding statistical lower limit are marked as outliers. For the outlier values, the moving median of the adjacent normal measurements of the corresponding parameter sequence is used for replacement and correction to generate a corrected parameter sequence; Based on the corrected peak energy sequence, the corrected energy decay rate sequence, and the arc duration sequence, the initial feature vector is constructed in a preset order.

4. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 1, characterized in that, Calculating the gradient direction and magnitude of the trial feature discrimination with respect to each antenna weight includes: Keeping the weights of the other antennas in the broadband microstrip antenna array unchanged, a small positive perturbation is added to the current weight of the target antenna, and the discrimination after the first perturbation is recalculated; Add the same small negative perturbation to the current weights of the target antenna, and recalculate the discrimination after the second perturbation; Subtract the second perturbation-based discrimination from the first perturbation-based discrimination, and then divide by twice the small perturbation value. The result is used as an approximate value of the partial derivative of the target antenna weight. By iterating through each antenna in the broadband microstrip antenna array and repeatedly performing the perturbation calculation process, a set of partial derivative values ​​for all antenna weights is obtained. The vector direction of the set of partial derivative values ​​is taken as the gradient direction, and the absolute value of each partial derivative is taken as the gradient magnitude.

5. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 1, characterized in that, The pre-training steps of the multi-state classifier include: A large number of electromagnetic signal samples from historical circuit breakers under known contact and mechanical conditions were collected, processed to obtain an optimized feature vector set of the samples, and labeled with the corresponding real contact health label and mechanical health label. Based on the fault feature mapping table, the optimized feature vectors of each sample are divided into a contact-related feature subset and a mechanical-related feature subset; A network architecture for a multi-state classifier is constructed, wherein the input layer of the network architecture is configured as a dual-branch structure, the first branch channel receives the contact-related feature subset, the second branch channel receives the mechanical-related feature subset, and the output layer of the network architecture is configured with dual output nodes to generate contact health prediction value and mechanical health prediction value respectively. The multi-state classifier is trained under supervision using the sample optimization feature vector set and the corresponding health labels until the combined error between the health prediction value output by the model and the true label is lower than a preset accuracy threshold, thus completing the model pre-training.

6. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 1, characterized in that, Based on adaptive threshold fusion judgment, health status assessment and early warning are achieved, including: Based on the initial warning threshold, a threshold particle set is constructed, which includes the current detection cycle threshold, the previous detection cycle threshold, and the historical average threshold. Based on the temporal changes of the contact health index and the mechanical health index, the trend of health index changes in the current detection cycle is calculated as the optimization direction for the current threshold particles. Extract the current feature discrimination scores from multiple historical detection periods and calculate the average score as the cost for the current detection period; The fusion radius of the current threshold particle is calculated based on the cost value, and adjacent threshold particles within the fusion radius are selected from the set of threshold particles. Calculate the variance of the adjacent threshold particles and the current threshold particle on each health index sequence as a comprehensive fluctuation measure, obtain the fusion weight of each particle, perform weighted fusion on the adjacent threshold particles, and generate a preliminary optimized threshold. When there are false intersections in the reverse extension lines of the health indicator change trends of multiple consecutive detection cycles, the preliminary optimization threshold is dynamically adjusted based on the predicted deterioration time corresponding to the false intersection, and an adaptive threshold for this round of early warning judgment is generated. The contact health index and the mechanical health index are compared with their corresponding adaptive thresholds. If either index is lower than the corresponding adaptive threshold, a corresponding warning is triggered.

7. The non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis according to claim 6, characterized in that, Calculate the variance of the adjacent threshold particles and the current threshold particle on each health index sequence as a comprehensive fluctuation measure, and obtain the fusion weight of each particle, including: Calculate the first variance of the adjacent threshold particles and the current threshold particles on the contact health index sequence, and the second variance on the mechanical health index sequence, respectively. Calculate the average of the first variance and the second variance as a comprehensive volatility measure; Divide the value 1 by the comprehensive volatility metric as the volatility compensation factor; The Euclidean distance between the current threshold particle and each adjacent threshold particle on the contact health index sequence and the mechanical health index sequence is calculated as the inter-particle distance. The product of the reciprocal of the distance between each particle and the fluctuation compensation factor is used as the initial fusion weight of the corresponding adjacent threshold particles. The initial fusion weights of all adjacent threshold particles are normalized to obtain the final particle fusion weights.

8. A non-invasive detection system for high-voltage circuit breakers based on electromagnetic signal analysis, characterized in that, The system is used to execute the non-invasive detection method for high-voltage circuit breakers based on electromagnetic signal analysis as described in any one of claims 1-7, and the system comprises: The electromagnetic signal acquisition module is used to capture the contact arc electromagnetic pulse signal during the opening process of the target circuit breaker based on a broadband microstrip antenna array, and to acquire multiple raw signals. The feature extraction and construction module is used to preprocess the multiple original signals and extract the arc duration, peak energy and energy decay rate of each signal to construct an initial feature vector. The feature discrimination calculation module is used to correct the preset distance metric based on the parameter distribution of the energy decay rate, and calculate the statistical distance between the two feature subsets reflecting contact degradation and mechanical defects in the initial feature vector as the current feature discrimination, including: According to the fault feature mapping table, a first feature subset corresponding to contact degradation and a second feature subset corresponding to mechanical defects are separated from the initial feature vector; Calculate the statistical variance and skewness of the energy decay rate sequence, and obtain the distribution correction factor based on their product; The distribution correction factor is multiplied by the basic weight coefficient to generate an adaptive weighting coefficient; Using the aforementioned adaptive weighting coefficients, the weights of each dimension in the preset distance metric are adjusted item by item to obtain an adaptive distance metric; Using the adaptive distance metric, the cluster centers of the first feature subset and the second feature subset in the feature space are calculated respectively; Calculate the Euclidean distance between two cluster centers, divide it by the sum of the maximum feature dispersion within each of the first and second feature subsets, normalize the result to the [0, 1] interval, and output the current feature discrimination score. The weight optimization generation module is used to iteratively adjust the spatial sensing weights of the broadband microstrip antenna array using a gradient-guided weight optimization algorithm, with the goal of maximizing the current feature discrimination, to generate an optimized feature vector, including: The initial weights of each antenna in the broadband microstrip antenna array are used to perform weighted fusion with the multiple original signals to generate a weighted signal set; Based on the weighted signal set, the trial feature vector under the current weight is obtained, and the corresponding trial feature discrimination is calculated. Calculate the gradient direction and gradient magnitude of the discriminative power of the trial feature with respect to the weights of each antenna; Based on the gradient direction and gradient magnitude, and combined with the preset learning rate, the weights of each antenna are updated in the direction that increases the feature discrimination. Repeat the iterative process of weighted fusion, feature extraction, discrimination calculation and weight update until the growth rate of the trial feature discrimination is lower than the preset convergence threshold, and obtain the optimal weight configuration; The optimal weight configuration is used to perform final weighted fusion of the multiple original signals to generate the optimized feature vector. The health assessment and early warning module is used to input the optimized feature vector into a pre-trained multi-state classifier, map and output contact health index and mechanical health index, and perform health status assessment and early warning based on adaptive threshold fusion. The multi-state classifier adopts a network architecture with a dual-branch input layer. The first branch channel receives the contact-related feature subset, and the second branch channel receives the mechanical-related feature subset. The output layer of the network architecture is configured with dual output nodes to generate contact health prediction value and mechanical health prediction value, respectively.

Images