Metrology switch fault identification and warning system

Abstract

The present application relates to the technical field of power system fault monitoring, and discloses a measurement switch fault identification and early warning system, which comprises a signal acquisition unit, a feature analysis unit, a fault determination unit and an early warning triggering unit. The signal acquisition unit collects potential fluctuation sequences and excitation current pulse sequences through contact voltage sensors and coil current transformers; the feature analysis unit extracts time-frequency domain and statistical domain feature groups by means of wavelet transform and moment function analysis; the fault determination unit generates a fault determination vector by fusing features using kernel principal component analysis; and the early warning triggering unit triggers four-level early warning using dynamically updated hierarchical thresholds (including historical reference values and online dynamic values). The system realizes accurate identification and hierarchical early warning of measurement switch faults through multi-modal feature fusion, dynamic threshold adjustment and hierarchical early warning strategies, improves fault identification accuracy and early warning timeliness, and is suitable for the fields of electric power and industrial automation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system fault monitoring, in particular to a metering switch fault identification and early warning system. BACKGROUND

[0002] In the field of power systems and industrial automation, metering switches as key components, the stability of their operating state directly affects the safety and reliability of the entire system. With the development of intelligent power equipment, higher requirements are put forward for fault early warning of metering switches.

[0003] Traditional metering switch fault detection methods have many shortcomings. Most detection methods rely on manual inspection, which is not only inefficient, but also difficult to detect potential faults in real time, often not until the fault has caused significant impact, leading to equipment damage and even system downtime, causing significant economic losses; some systems based on simple electrical parameter monitoring cannot fully capture the complex feature changes of switch faults due to single monitoring dimension. For example, only monitoring a single parameter such as voltage or current, when the contact appears oxidation, poor contact, etc., the time-frequency domain characteristics of potential fluctuations and the statistical domain characteristics of excitation current will change, but traditional methods are difficult to comprehensively analyze these multi-dimensional characteristics, and are prone to misjudgment or omission.

[0004] Existing early warning systems lack dynamic adjustment mechanisms in threshold setting. Traditional thresholds are usually set based on historical experience and cannot be updated adaptively according to changes in real-time operating environment. In actual operation, fluctuations in environmental temperature, humidity, vibration, etc. will affect the electrical characteristics of the switch, making it difficult for fixed thresholds to accurately reflect the true operating state of the switch, resulting in insufficient timeliness and accuracy of the warning. For example, in an environment with high temperature or high humidity, the aging rate of switch components increases and the probability of failure increases, but the fixed threshold cannot be adjusted in time, which may delay the warning.

[0005] For faults of different severity, traditional systems lack a hierarchical warning mechanism and cannot take appropriate measures according to the severity of the fault, making it difficult to achieve precise fault management. Moreover, in terms of fault location, traditional methods can only determine whether a fault exists, but cannot accurately determine the specific components and location of the fault, making it difficult for subsequent maintenance and repair, and prolonging the downtime of the equipment.

[0006] With the development of smart grid and industrial automation technology, higher requirements are put forward for fault early warning of measuring switches, and a system capable of real-time, accurate and comprehensive identification of faults and hierarchical early warning is needed to improve the reliability and safety of power systems and industrial automation equipment. Therefore, it has important practical significance to develop a measuring switch fault identification and early warning system with multi-modal feature fusion, dynamic threshold adjustment and hierarchical early warning function. SUMMARY

[0007] The purpose of the present application is to provide a measuring switch fault identification and early warning system to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solution: a measuring switch fault identification and early warning system, the system comprising:

[0009] A signal acquisition unit: a contact voltage sensor is used to collect a contact potential fluctuation sequence, and a coil current transformer is used to obtain an excitation current pulse sequence; the potential fluctuation sequence contains amplitude-time corresponding data, and the current pulse sequence contains phase-frequency correlation information;

[0010] A feature analysis unit: a time-frequency domain feature group is extracted based on the potential fluctuation sequence, and a statistical domain feature group is extracted based on the current pulse sequence; the time-frequency domain feature group is calculated by a wavelet transform algorithm, and the statistical domain feature group is obtained by a moment function analysis technique;

[0011] A fault determination unit: the time-frequency domain feature group and the statistical domain feature group are subjected to multi-modal feature fusion to generate a fault determination vector; the multi-modal feature fusion adopts a kernel principal component analysis strategy, and the fusion weight is set according to the feature saliency;

[0012] An early warning triggering unit: a hierarchical threshold dynamically updated is used to match the fault determination vector to trigger the early warning instruction of the corresponding level; the hierarchical threshold includes a reference value trained by historical fault samples and a dynamic value adaptively adjusted online.

[0013] Preferably, the specific way of collecting the contact potential fluctuation sequence by the contact voltage sensor and obtaining the excitation current pulse sequence by the coil current transformer comprises:

[0014] The contact voltage sensor monitors the contact potential difference at a fixed sampling frequency, and converts the analog signal into digital potential data; the coil current transformer senses the coil magnetic flux change at a synchronous sampling frequency, and generates digital current data through a signal conditioning unit; the digital potential data is calibrated to a unified time axis through a time synchronization module, and the digital current data is filtered to remove high-frequency noise through a band-pass filter unit.

[0015] Preferably, the specific way of extracting the time-frequency domain feature group based on the potential fluctuation sequence comprises:

[0016] Step A, selecting a characteristic analysis window in the potential fluctuation sequence, determining the time length and frequency resolution of the window;

[0017] Step B, decomposing the analysis window signal into an energy distribution matrix on the time-frequency plane through wavelet transform, the time-frequency resolution of the energy distribution matrix being determined by the scale parameter of the wavelet base function;

[0018] Step C, constructing a three-dimensional feature tensor containing energy features, frequency features and time features based on the peak position, frequency band width and time delay information of the energy distribution matrix, as a time-frequency domain feature group.

[0019] Preferably, the specific way of extracting the statistical domain feature group based on the current pulse sequence comprises:

[0020] The statistical domain feature group contains first moment features and second moment features;

[0021] The mean and variance of the current pulse are calculated through the probability density function to obtain the first moment value and the second moment value;

[0022] The skewness and kurtosis values of the distribution are calculated based on the change trend of the first moment value and the second moment value;

[0023] The mean, variance, skewness and kurtosis values are integrated into a vector form as the statistical domain feature group.

[0024] Preferably, the specific way of performing multi-modal feature fusion on the time-frequency domain feature group and the statistical domain feature group to generate a fault judgment vector comprises:

[0025] A time-frequency feature space model is established based on the energy features, frequency features and time features in the time-frequency domain feature group;

[0026] A statistical feature distribution model is established based on the first moment features and the second moment features in the statistical domain feature group;

[0027] The time-frequency feature space model and the statistical feature distribution model are aligned through feature space mapping, the statistical information is embedded into the time-frequency space, and a fault judgment vector containing time-frequency characteristics and statistical attributes is formed.

[0028] Preferably, the specific way of performing hierarchical matching on the fault judgment vector using a dynamically updated hierarchical threshold to trigger a corresponding level of warning instruction comprises:

[0029] The time-frequency features and statistical features in the fault judgment vector are input as inputs; the input dimension is set to be consistent with the number of features;

[0030] The hidden layer is provided with multiple layers of nonlinear transformation layers; the input data is transmitted from the input layer to the first layer of transformation layers, processed by linear combination and activation function, to obtain the first layer output, which is transmitted to the next layer of transformation layers, and the nonlinear processing is repeated until the last layer of transformation layers is completed, to obtain the final transformation output;

[0031] The output layer transmits the final transformation output to the output layer, and generates a warning level value through a threshold comparison function; the warning level value includes four levels of normal, prompt, warning and failure;

[0032] The threshold weight is calculated, and the adjustment weight of the classification threshold is determined based on the occurrence probability of the historical failure sample and the stability of the current operating environment;

[0033] The warning operation is performed, the warning level value is multiplied by the adjustment weight and matched to the corresponding warning strategy, and the warning trigger is completed.

[0034] Preferably, the specific way of determining the adjustment weight of the classification threshold based on the occurrence probability of the historical failure sample and the stability of the current operating environment comprises:

[0035] The probability weight is calculated by a Poisson distribution function for the occurrence probability of the historical failure sample; the parameters of the Poisson distribution function are set according to the historical frequency of the failure type;

[0036] The stability weight is calculated by an environmental monitoring index for the stability of the current operating environment; the environmental monitoring index includes temperature fluctuation range, humidity duration and vibration acceleration value;

[0037] The adjustment weight of the classification threshold is calculated by a weighted sum formula combining the probability weight and the stability weight.

[0038] Preferably, when the warning level value is prompt, the corresponding warning strategy is executed, which specifically comprises:

[0039] When the warning type is prompt, the real-time data of the environmental monitoring device is used to obtain the running state deviation of the current switch; the monitoring frequency increment is calculated by the running state deviation and the warning level value; for the monitoring frequency increment, the sampling rate adjustment algorithm is used to improve the sampling frequency of the signal acquisition unit, and the monitoring density is improved to the level corresponding to the prompt level.

[0040] Preferably, when the warning level value is warning, the corresponding warning strategy is executed, which specifically comprises:

[0041] The current failure judgment vector corresponds to the potential failure type is identified by using the failure case library matching technology; the potential failure type includes contact oxidation and coil turn-to-turn short circuit; based on the potential failure type and the warning level value, the failure probability of the corresponding component in the switch is marked by the failure positioning rule, to complete the failure warning prompt.

[0042] Preferably, when the early warning level value is a fault, the corresponding early warning strategy is executed, specifically including:

[0043] Triggering the action of the hardware protection device, cutting off the fault circuit by controlling the relay; the action time of the relay is determined by the severity of the fault determination vector; sending fault alarm information to the monitoring terminal through the communication module, the alarm information including the fault type, the occurrence time and the component position, completing the fault level early warning execution.

[0044] Compared with the prior art, the present application has the beneficial effects that:

[0045] The system collects the contact potential fluctuation sequence and the excitation current pulse sequence through the contact voltage sensor and the coil current transformer respectively, combines the wavelet transform algorithm and the matrix function analysis technology, realizes the accurate extraction of the time-frequency domain characteristics of the potential fluctuation and the statistical domain characteristics of the current pulse, and the multi-dimensional feature collection and analysis mode can fully capture the electrical characteristic changes of the measuring switch under different fault states, overcome the limitations of the traditional single parameter monitoring, and greatly improve the accuracy and comprehensiveness of fault identification.

[0046] In terms of feature fusion, a kernel principal component analysis strategy is used for multi-modal feature fusion, the fusion weight is set according to the feature saliency, and the time-frequency characteristics and statistical characteristics are aligned through feature space mapping to generate a fault determination vector containing time-frequency characteristics and statistical properties. This fusion method fully utilizes the information complementarity of different feature domains, makes the fault determination more scientific and reasonable, and effectively reduces the probability of misjudgment and omission.

[0047] The early warning triggering unit uses the dynamically updated hierarchical threshold to perform hierarchical matching on the fault determination vector, and the hierarchical threshold includes the reference value trained by the historical fault samples and the dynamic value adjusted online. The probability weight corresponding to the occurrence probability of the historical fault sample is calculated through the Poisson distribution function, the stability weight calculated by the environmental monitoring index is combined to determine the adjustment weight of the hierarchical threshold, and the dynamic adaptive adjustment of the threshold value is realized. This makes the system can accurately adjust the early warning threshold according to the real-time running environment and historical fault rules, improves the timeliness and accuracy of early warning, and solves the problem that the traditional fixed threshold cannot adapt to environmental changes.

[0048] In the hierarchical early warning strategy, the system sets four early warning levels of normal, prompt, warning and failure, and formulates corresponding early warning strategies for different levels. When the early warning level is prompt, the sampling frequency of the signal acquisition unit is increased to increase the monitoring density, so that potential failures can be found earlier. When the early warning level is warning, the fault case library matching technology is used to identify the type of potential failure, and the failure probability of the corresponding component in the switch is marked, so that accurate failure location and early warning prompt are realized, and a clear direction is provided for maintenance. When the early warning level is failure, the hardware protection device is triggered to cut off the fault circuit, and alarm information containing the failure type, occurrence time and component location is sent to the monitoring terminal, so that the failure can be quickly isolated, the failure loss can be reduced, and the safety and reliability of the system can be improved.

[0049] In addition, in the signal acquisition process, the time synchronization module is used to calibrate the digital potential data, and the band-pass filter unit is used to filter high-frequency noise of the digital current data, so as to ensure the accuracy and reliability of the collected signals. In the feature extraction process, the three-dimensional feature tensor and the statistical domain feature vector are constructed, which provide rich feature information for subsequent failure judgment. In the early warning execution process, the action time of the control relay is determined according to the severity of the failure judgment vector, so that more refined failure processing is realized. In summary, the system improves the ability of measuring switch failure identification and early warning through multi-dimensional feature acquisition, fusion analysis, dynamic threshold adjustment and hierarchical early warning strategy, and has significant technical advantages and practical value. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The working principle diagram of the measuring switch failure identification and early warning system described in the application;

[0051] Figure 2 The flowchart for extracting the time-frequency domain feature group;

[0052] Figure 3 The flowchart for multi-modal feature fusion;

[0053] Figure 4 The strategy diagram for failure level early warning. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0055] Please refer to Figures 1-4The application relates to a measurement switch fault identification and early warning system, which comprises a signal acquisition unit, a feature analysis unit, a fault determination unit and an early warning triggering unit. The specific implementation is as follows.

[0056] The signal acquisition unit collects a contact point potential fluctuation sequence through a contact voltage sensor, and the sequence contains amplitude-time corresponding data. Meanwhile, an excitation current pulse sequence is obtained through a coil current transformer, and the sequence contains phase-frequency associated information.

[0057] The feature analysis unit calculates and extracts a time-frequency domain feature group based on the potential fluctuation sequence through a wavelet transform algorithm, and obtains a statistical domain feature group based on the current pulse sequence through a moment function analysis technology.

[0058] The fault determination unit performs multi-modal feature fusion on the time-frequency domain feature group and the statistical domain feature group, adopts a kernel principal component analysis strategy for multi-modal feature fusion, sets a fusion weight according to feature saliency, and thereby generates a fault determination vector.

[0059] The early warning triggering unit performs hierarchical matching on the fault determination vector by using a dynamically updated hierarchical threshold value, the hierarchical threshold value comprises a reference value trained by a historical fault sample and a dynamic value adaptively adjusted online, and thereby triggers a corresponding level early warning instruction.

[0060] Embodiment 1:

[0061] The specific implementation of the signal acquisition unit is as follows: the core function of the unit is to realize the collection and preprocessing of the contact point potential fluctuation sequence and the excitation current pulse sequence, so as to obtain high-quality original signal data, and provide a reliable basis for subsequent feature analysis and fault determination.

[0062] In the collection process of the contact point potential fluctuation sequence, the contact voltage sensor undertakes a key monitoring task. The sensor monitors the potential difference of the contact point in real time at a fixed sampling frequency. The fixed sampling frequency here needs to be determined according to the working characteristics of the measurement switch and the frequency components of the signal, for example, it can be set to a frequency value that can meet the Nyquist sampling theorem, so as to ensure effective sampling of the signal and avoid frequency aliasing phenomenon. During the monitoring process, the sensor converts the analog signal of the real-time sensed contact point potential difference into digital potential data through an internal analog-digital conversion module. The conversion process needs to ensure conversion accuracy to accurately reflect the actual fluctuation of the contact point potential, for example, a 16-bit analog-digital converter can be used to improve the resolution of the digital potential data.

[0063] For the acquisition of the excitation current pulse sequence, the coil current transformer plays an important role. The transformer senses the magnetic flux change of the coil at a sampling frequency synchronized with the contact voltage sensor. The setting of the synchronous sampling frequency is to ensure the consistency of the potential fluctuation sequence and the current pulse sequence in time, which is convenient for subsequent joint analysis. After the coil current transformer senses the magnetic flux change, the generated electric signal needs to be processed by the signal conditioning unit to generate digital current data that meets the requirements. The signal conditioning unit usually includes amplification, filtering and other links. The amplification link can adjust the gain according to the strength of the signal to make the signal amplitude in a suitable range; the filtering link can preliminarily filter out some low-frequency or high-frequency interference signals to improve the quality of the current data.

[0064] After the acquisition of digital potential data and digital current data, time synchronization processing is needed for these two types of data. The digital potential data is calibrated to a unified time axis through the time synchronization module. The time synchronization module can use a high-precision clock source such as a GPS clock or a Beidou clock to provide an accurate time reference. Through time synchronization processing, the potential data and current data collected at different times can be corresponded to a unified time coordinate system, ensuring the consistency and comparability of the data in the time dimension, and laying a foundation for subsequent multi-source data fusion analysis.

[0065] For digital current data, in addition to time synchronization, it also needs to be further filtered by a band-pass filter unit to filter out high-frequency noise. The passband range of the band-pass filter unit needs to be set according to the frequency characteristics of the excitation current pulse sequence, for example, it can be set to allow signals within a certain frequency range to pass through, while attenuating high-frequency noise above or below this range. The band-pass filter unit can be in the form of an active filter or a passive filter, and its filtering characteristics should meet the effective suppression of high-frequency noise while ensuring that the main frequency components of the current pulse signal can pass through smoothly to avoid signal distortion.

[0066] In practical applications, the installation positions of the contact voltage sensor and the coil current transformer need to be reasonably arranged according to the structure of the measurement switch. The contact voltage sensor should be installed near the contact point to accurately monitor the potential fluctuation of the contact point; the coil current transformer should be sleeved on the lead wire of the coil to ensure that it can sensitively sense the magnetic flux change of the coil. The signal transmission cable between the sensor and the data acquisition equipment needs to take shielding measures to reduce the influence of external electromagnetic interference on the signal and ensure that the collected signal has a high signal-to-noise ratio.

[0067] The hardware performance of the data acquisition device also needs to meet the requirements. Its sampling rate, storage capacity and processing capacity should be able to adapt to the needs of real-time acquisition and processing of a large amount of data. For example, the sampling rate of the data acquisition device should match the set sampling frequency, the storage capacity should be able to meet the storage needs of long-term continuous data acquisition, and the processing capacity should be able to timely perform preliminary preprocessing and caching on the collected data.

[0068] In addition, in order to ensure the stable operation of the signal acquisition unit, a corresponding self-checking and calibration mechanism also needs to be set. The self-checking mechanism can periodically check the working state of the sensor, the accuracy of the sampling frequency, the clock accuracy of the time synchronization module, etc., and timely issue an alarm signal when an abnormality is found. The calibration mechanism can periodically calibrate the measurement accuracy of the sensor to ensure that the collected data is accurate and reliable. For example, a standard signal source can be used every certain period of time to calibrate the contact voltage sensor and the coil current transformer, and adjust their measurement parameters to control their measurement error within the allowed range.

[0069] At the software level, a corresponding data acquisition and preprocessing program needs to be developed. This program is responsible for controlling the sampling timing of the sensor, realizing real-time acquisition, storage and preliminary processing of digital potential data and digital current data. The program should include a time synchronization algorithm to accurately implement time calibration of the potential data, and a band-pass filtering algorithm to filter out high-frequency noise from the current data. At the same time, the software program should have a good human-computer interaction interface to facilitate the operator to set and adjust the acquisition parameters, as well as to view the real-time acquisition data and processing results.

[0070] Through the coordinated work of the above various links, the signal acquisition unit can stably and reliably collect the contact potential fluctuation sequence and the excitation current pulse sequence, and perform necessary preprocessing on the collected data, providing high-quality raw data for the subsequent feature analysis unit, and ensuring the normal operation and accurate warning of the entire measurement switch fault recognition and warning system.

[0071] Example 2:

[0072] The specific implementation of the feature analysis unit based on the extraction of the time-frequency domain feature group from the potential fluctuation sequence is as follows: This process aims to extract key features from the potential fluctuation sequence that can reflect the time-frequency characteristics of the signal, providing multi-dimensional feature information for subsequent fault judgment.

[0073] The characteristic analysis window needs to be selected in the potential fluctuation sequence, which is the basic step of extracting the time-frequency domain features. The selection of the characteristic analysis window needs to consider the time characteristics and frequency characteristics of the signal. The time length of the window should be long enough to contain enough signal information and reflect the dynamic change process of the potential fluctuation; at the same time, the time length of the window cannot be too long to avoid containing too much redundant information and affecting the efficiency and accuracy of feature extraction. For example, the time length of the window can be set between tens of milliseconds to hundreds of milliseconds according to the action period of the measurement switch and the typical time scale of the potential fluctuation. The frequency resolution of the window determines the resolution ability of the frequency components of the signal, which needs to be determined according to the complexity of the frequency components that may be contained in the potential fluctuation. The higher the frequency resolution, the more accurate the resolution of the frequency components, but it will also increase the calculation amount.

[0074] After determining the time length and frequency resolution of the characteristic analysis window, the signal in the analysis window needs to be decomposed into an energy distribution matrix on the time-frequency plane through wavelet transform. Wavelet transform is a time-frequency analysis method with multi-resolution analysis characteristics, which can analyze the signal at different scales to obtain the energy distribution of the signal in the time and frequency dimensions. The selection of the wavelet basis function is one of the key steps of wavelet transform, and different wavelet basis functions have different time-frequency localization characteristics and are suitable for different types of signal analysis. For example, the db series wavelet basis function has good compactness and regularity, and is often used for analysis of power system signals; the Meyer wavelet basis function has smooth characteristics and is suitable for analyzing smooth signals. In practical applications, the appropriate wavelet basis function needs to be selected according to the characteristics of the potential fluctuation sequence. The scale parameter of the wavelet basis function determines the time-frequency resolution of the energy distribution matrix. A larger scale parameter corresponds to a lower frequency resolution and a higher time resolution, and a smaller scale parameter corresponds to a higher frequency resolution and a lower time resolution. The scale parameter needs to be adjusted according to the specific analysis requirements.

[0075] After obtaining the energy distribution matrix on the time-frequency plane, it is necessary to construct a three-dimensional feature tensor containing energy features, frequency features and time features based on the peak position, frequency band width and time delay information of the matrix. The peak position of the energy distribution matrix reflects the time and frequency point where the energy is concentrated in the signal. The frequency corresponding to these peak positions is often the main frequency component of the signal, and the time point corresponding to these peak positions is the time when the energy is concentrated in the signal. By extracting the information of these peak positions, the main frequency characteristics and time characteristics of the signal can be obtained. The frequency band width reflects the distribution range of the frequency components of the signal. A wider frequency band width indicates that the signal contains more frequency components, which may correspond to complex potential fluctuation conditions. A narrower frequency band width indicates that the frequency components of the signal are relatively simple. The time delay information reflects the shift of the signal on the time axis, which may be related to the action process of the measuring switch or the occurrence time of the fault.

[0076] The construction of energy features needs to consider the energy values of each frequency and time point in the energy distribution matrix. For example, the total energy of the energy distribution matrix and the energy proportion of each frequency band can be calculated. These energy features can reflect the intensity and energy distribution of potential fluctuations. Frequency features not only include the frequency values corresponding to peak positions, but also include the center frequency of the frequency band and the distribution range of the frequency. These features can reflect the frequency characteristics of potential fluctuations. Time features include the time points corresponding to peak positions and the specific values of time delay. These features can reflect the time characteristics of potential fluctuations.

[0077] When constructing the three-dimensional feature tensor, energy features, frequency features and time features need to be organically combined to form a multi-dimensional feature vector. For example, energy features can be used as one dimension, frequency features as one dimension, and time features as one dimension. Each dimension contains multiple specific feature parameters, forming a three-dimensional feature tensor. This three-dimensional feature tensor can comprehensively describe the time-frequency characteristics of potential fluctuation sequences, providing rich feature information for subsequent fault judgment.

[0078] In practical applications, in order to improve the efficiency and accuracy of feature extraction, the selection of feature analysis window, the selection of wavelet basis function and the adjustment of scale parameter need to be optimized. Through a large number of experiments and data analysis, the optimal feature analysis window, wavelet basis function and scale parameter combination for different types of measuring switches and different fault types can be determined. For example, for a certain type of measuring switch, by analyzing the potential fluctuation sequences under normal operation and different fault conditions, the feature analysis window size and frequency resolution that can best distinguish between normal and fault states, as well as the most suitable wavelet basis function and scale parameter can be determined.

[0079] In addition, in order to adapt to different operating environments and signal characteristics, the feature extraction process also needs to have certain self-adaptive ability. For example, according to the statistical characteristics of the real-time collected potential fluctuation sequence, the size and frequency resolution of the feature analysis window can be automatically adjusted, or the appropriate wavelet basis function and scale parameter can be automatically selected, so as to ensure that effective time-frequency domain features can be extracted in different cases.

[0080] The hardware implementation of the feature analysis unit needs to have sufficient computing power to process a large amount of potential fluctuation data and complex wavelet transform calculation. High-performance processors or dedicated digital signal processing chips can be used to improve the speed and real-time performance of feature extraction. On the software level, efficient wavelet transform algorithms and feature extraction programs need to be developed to ensure the accuracy and computational efficiency of the algorithms.

[0081] Through the above detailed implementation, the feature analysis unit can extract comprehensive and effective time-frequency domain feature groups from the potential fluctuation sequence, providing important feature input for the subsequent fault judgment unit, so as to realize accurate identification and early warning of the measurement switch fault.

[0082] Embodiment 3:

[0083] The specific implementation of extracting the statistical domain feature group based on the current pulse sequence is as follows: this process extracts feature parameters that can reflect the probability distribution characteristics and change trend of the current pulse sequence through statistical analysis of the current pulse sequence, providing feature basis based on statistical law for fault judgment.

[0084] It is clear that the statistical domain feature group contains first-order moment features and second-order moment features, which are the basis for describing the statistical characteristics of the current pulse sequence. The first-order moment feature mainly reflects the concentration trend of the sequence, and the second-order moment feature mainly reflects the dispersion degree of the sequence. In order to obtain these features, the probability density function of the current pulse sequence needs to be calculated first. The probability density function can describe the probability distribution of the current pulse value, which is estimated based on a large number of sample data of the current pulse sequence. In practical application, the kernel density estimation method can be used to calculate the probability density function. This method does not need to assume the distribution form of the data in advance, and can adaptively estimate the probability density based on the sample data, which has strong flexibility and adaptability.

[0085] After obtaining the probability density function of the current pulse sequence, the mean and variance of the current pulse can be calculated by integrating the function, thereby obtaining the first and second moment values. The mean, as the first moment value, represents the average level of the current pulse, which reflects the typical value of the coil excitation current. The variance, as the second moment value, measures the dispersion of the current pulse around the mean. The larger the variance, the greater the fluctuation range of the current pulse, and vice versa. When calculating the mean and variance, all sample data of the current pulse sequence need to be processed to ensure that the calculation results accurately reflect the overall statistical characteristics of the sequence.

[0086] After obtaining the first and second moment values, distribution fitting needs to be performed based on their variation trends. The purpose of distribution fitting is to describe the probability distribution law of the current pulse sequence through a mathematical model, so as to further calculate higher-order statistical characteristics. When performing distribution fitting, first, the variation of the first and second moment values in different time periods or different operating states needs to be observed, and whether the statistical characteristics of the current pulse sequence have changed needs to be analyzed. Then, according to the characteristics of the current pulse sequence and common probability distribution models, a suitable distribution model is selected for fitting, such as normal distribution, exponential distribution, t distribution, etc. Different distribution models have different parameters and characteristics, which need to be selected and adjusted according to the actual data situation to ensure that the fitting results can better reflect the distribution law of the current pulse sequence.

[0087] After completing the distribution fitting, the skewness and kurtosis values of the distribution are calculated next. Skewness is used to measure the symmetry of the distribution. If the skewness is 0, the distribution is symmetric. If the skewness is positive, the distribution is right-skewed. If the skewness is negative, the distribution is left-skewed. Kurtosis is used to measure the "peak" degree of the distribution. The larger the kurtosis value, the sharper the peak of the distribution, and the thicker the tail. The smaller the kurtosis value, the flatter the peak of the distribution, and the thinner the tail. Skewness and kurtosis are important characteristic parameters for describing the shape of the current pulse sequence distribution, which can provide more detailed information about the probability distribution of the current pulse sequence and help identify different types of faults.

[0088] When calculating skewness and kurtosis, they need to be calculated based on the fitted distribution model and sample data of the current pulse sequence through specific statistical formulas. These calculation processes need to accurately process each sample data to ensure the accuracy of the calculation results. For example, for a normal distribution, the skewness is 0 and the kurtosis is 3. If the calculated skewness and kurtosis values deviate greatly from the theoretical values of the normal distribution, it indicates that the distribution of the current pulse sequence may have changed, which may indicate a fault in the measuring switch.

[0089] After obtaining the mean, variance, skewness and kurtosis values, these characteristic values need to be integrated into vector form as the statistical domain feature group. The vector form of the statistical domain feature group facilitates subsequent multi-modal feature fusion and fault determination. During the integration process, these characteristic values need to be arranged in a certain order to form a unified feature vector. For example, the mean can be taken as the first element of the vector, the variance as the second element, the skewness as the third element, and the kurtosis as the fourth element, thereby constituting a four-dimensional statistical domain feature vector. This vector form can organically combine multiple statistical features together and comprehensively describe the statistical characteristics of the current pulse sequence.

[0090] In practical applications, in order to improve the reliability and effectiveness of the statistical domain feature group, the acquisition and preprocessing process of the current pulse sequence needs to be strictly controlled. The acquisition of the current pulse sequence needs to ensure sufficient sampling frequency and sampling accuracy to ensure that the detailed features of the current pulse can be accurately captured. The preprocessing process includes denoising, filtering and other operations to remove noise and interference signals in the current pulse sequence and improve the quality of the data. For example, median filtering, Gaussian filtering and other methods can be used for denoising of the current pulse sequence, and band-pass filtering method can be used to remove interference signals of specific frequency.

[0091] In addition, in order to adapt to different operating conditions and fault types, the extraction process of the statistical domain feature group also needs to be optimized and adjusted regularly. By analyzing a large amount of historical data, statistical domain feature templates under different operating conditions and fault types can be established, and when the real-time extracted statistical domain features deviate greatly from the templates, early warning signals can be sent in time. At the same time, machine learning methods can be used to train and learn the statistical domain feature group, improving the accuracy of feature extraction and the ability of fault recognition.

[0092] The extraction process of the statistical domain feature group needs to be reasonably designed on the hardware and software levels. The hardware level needs to have high-speed data acquisition and processing capability to process a large amount of current pulse data in real time; the software level needs to develop efficient statistical analysis algorithms and feature extraction programs to ensure that the statistical domain feature group can be extracted quickly and accurately. For example, parallel computing technology can be used in software to improve the efficiency of probability density function calculation and distribution fitting; optimized numerical calculation method can be used to improve the calculation accuracy of skewness and kurtosis values.

[0093] Through the above detailed embodiments, comprehensive and accurate statistical domain feature groups can be extracted from the current pulse sequence, which, combined with the time-frequency domain feature groups extracted from the potential fluctuation sequence, provide a solid foundation for multi-modal feature fusion and accurate determination of the measuring switch fault, thereby realizing early identification and early warning of the measuring switch fault.

[0094] Example 4:

[0095] The fault determination unit fuses the time-frequency domain feature set and the statistical domain feature set in a multi-modal feature and generates a fault determination vector. The specific implementation is as follows: This process needs to realize the organic combination of the two types of features, form a comprehensive feature vector with time-frequency characteristics and statistical properties through spatial mapping and feature integration, and provide multi-dimensional basis for fault level determination.

[0096] Taking a certain type of high-voltage measuring switch as an example, after the system collects the contact potential fluctuation sequence and the excitation current pulse sequence, the feature analysis unit has extracted the time-frequency domain feature set and the statistical domain feature set. Among them, the time-frequency domain feature set contains a three-dimensional feature tensor, such as the energy proportion of a certain frequency band in the energy feature, the peak frequency point in the frequency feature, and the fluctuation starting time in the time feature; the statistical domain feature set contains a four-dimensional feature vector, such as the mean, variance, skewness, and kurtosis value of the current pulse. At this time, the fault determination unit needs to fuse and process these two types of features.

[0097] Based on the energy feature, frequency feature, and time feature in the time-frequency domain feature set, a time-frequency feature space model is established. Taking the energy feature as the Z-axis, the frequency feature as the X-axis, and the time feature as the Y-axis, a three-dimensional space coordinate system is constructed. For example, the frequency value (such as 500Hz), time point (such as 0.02s), and energy intensity (such as 10mW) corresponding to the peak position of the energy distribution matrix of the potential fluctuation in a certain time period are taken as a feature point in space, and multiple such feature points can constitute a feature space model reflecting the time-frequency distribution of potential fluctuation. This model needs to cover all key parameters of the time-frequency domain feature set, ensuring that each dimension in the space can correspond to a specific physical meaning, such as the frequency dimension reflecting the characteristic frequency of the contact arc discharge and the time dimension reflecting the timing characteristics of the switch action.

[0098] Based on the first-order moment feature and the second-order moment feature in the statistical domain feature set, a statistical feature distribution model is established. Taking the mean of the current pulse as the center, measuring the degree of data dispersion with variance, and combining skewness and kurtosis value to depict the distribution form. For example, when running normally, the mean of the current pulse is 2A, the variance is 0.1A², the skewness is close to 0, and the kurtosis value is 3, at this time the statistical feature distribution model presents a standard normal distribution; if the switch has a contact oxidation fault, the mean of the current pulse may decrease to 1.8A, the variance increases to 0.3A², the skewness becomes 0.5, and the kurtosis value rises to 3.8, the distribution model presents a right-skewed and sharp-peaked form. This model can be visualized through the probability density function curve, which can intuitively reflect the statistical rule changes of the current pulse.

[0099] The time-frequency feature space model is aligned with the statistical feature distribution model by feature space mapping. Due to the differences in physical dimensions and feature dimensions of the two types of features, standardization processing is required. The energy, frequency, and time parameters in the time-frequency domain features are normalized, and the mean, variance, and other parameters in the statistical domain features are converted into dimensionless values. For example, the frequency feature (0-1000 Hz) is normalized to the [0, 1] interval, and the current mean (0-5 A) is normalized to the same interval. Then, the nonlinear mapping method in the kernel principal component analysis (KPCA) strategy is used to embed the statistical feature distribution model into the time-frequency feature space. Specifically, the statistical feature vector is mapped to a high-dimensional space through a kernel function (such as a radial basis function), and then the dimension is aligned with the original time-frequency feature space. For example, the variance parameter in the statistical feature is used as a weight to adjust the energy weight of the corresponding frequency point in the time-frequency feature space, so that the feature points with higher variance have higher contribution in the fusion space.

[0100] During the mapping process, the fusion weight needs to be set according to the feature significance. The feature significance is determined by calculating the correlation of each feature parameter with the historical fault samples. For example, by analyzing historical data, it is found that when the contact is oxidized, the energy proportion of the 500-800 Hz frequency band in the time-frequency domain feature has the highest correlation with the variance parameter in the statistical domain feature, so these two features are given higher weights (such as 0.25) during fusion, while the time delay feature with lower correlation is given lower weight (such as 0.1). The weight setting needs to be optimized through iterative training, and based on the historical fault sample set, multiple fusion experiments are performed to adjust the feature weights until the fault judgment vector after fusion can best distinguish different fault types.

[0101] After mapping, a fault judgment vector containing time-frequency characteristics and statistical attributes is formed. This vector is usually a multi-dimensional feature vector, for example, the three-dimensional feature of the time-frequency domain and the four-dimensional feature of the statistical domain are expanded to a ten-dimensional vector after mapping, each dimension corresponds to a comprehensive feature parameter after fusion. Taking the contact oxidation fault as an example, the fused vector may contain: the 500 Hz energy weight value in the time-frequency domain (which incorporates the influence of the variance feature), the coupling parameter of the statistical domain mean and frequency feature, the correlation value of the time feature and skewness, etc. Each parameter reflects the cross-influence of the two types of features, such as a parameter that reflects both the frequency characteristics of the potential fluctuation and the dispersion degree of the current pulse.

[0102] In practical applications, the feature space mapping needs to consider the operating condition changes of the measurement switch. For example, when the environmental temperature rises, the coil resistance change may cause the mean current to deviate, at which time the weight of the statistical domain feature needs to be dynamically adjusted to avoid misjudgment caused by environmental factors. The system can collect environmental monitoring data (such as temperature and humidity) in real time, establish an association model of environmental parameters and feature weights, and automatically adjust the fusion weight when the environmental parameter exceeds the threshold. For example, the weight of the mean current feature decreases by 0.05 for every 10°C rise in temperature, in order to reduce the interference of environmental factors on the fusion result.

[0103] In addition, the fusion process needs to have real-time performance to meet the timeliness requirements of fault early warning. On the hardware level, field programmable gate array (FPGA) can be used for parallel computing to accelerate kernel function mapping and weight calculation; on the software level, the algorithm process is optimized, and steps such as feature standardization, space mapping, and weight calculation are processed in a pipeline. For example, for a signal with a sampling frequency of 10 kHz, the system needs to complete feature fusion within 100 ms to ensure that the delay of fault early warning does not exceed one sampling period.

[0104] To verify the reliability of the fusion effect, historical fault data can be used for backtracking testing. For example, 100 normal samples, 50 contact oxidation samples, and 50 coil turn-to-turn short circuit samples are selected, and the clustering analysis is performed using the fused fault judgment vector. Normal samples should form a dense cluster in the feature space, and fault samples should be distributed in different outlier areas, and the vector distance of the same type of fault sample is less than that of different types of fault samples. This clustering distribution characteristic does not need to be quantified, and the discrimination ability of the fused feature can be judged through visualization.

[0105] In engineering implementation, the redundancy of feature dimension also needs to be considered. Through principal component analysis (PCA), the high-dimensional vector after fusion is reduced in dimension, and the principal components with a contribution rate of more than 90% are retained, reducing the amount of calculation while avoiding overfitting. For example, the ten-dimensional vector is reduced to five-dimensional, and the cumulative contribution rate of the first five principal components is 92%, which can retain key feature information and improve the efficiency of fault judgment.

[0106] This embodiment realizes multi-dimensional recognition of the measurement switch fault by constructing a time-frequency feature space and a statistical feature distribution model, using standardization processing and nonlinear mapping to realize feature alignment, combining dynamic weight setting and dimensionality reduction optimization, and finally generating a fault judgment vector with physical meaning and statistical law, providing accurate feature input for the subsequent early warning trigger unit level matching.

[0107] Embodiment 5:

[0108] The specific implementation of the early warning triggering unit using the dynamically updated hierarchical threshold to perform hierarchical matching on the fault determination vector and trigger the early warning instruction is as follows: the process takes the real-time monitoring of a certain 10kV high-voltage measurement switch as an application scenario, and realizes accurate mapping from the feature vector to the early warning strategy through multi-level signal processing and dynamic threshold adjustment.

[0109] When the fault determination unit generates a ten-dimensional fault determination vector containing time-frequency and statistical features, the input layer of the early warning triggering unit takes the time-frequency features (such as the energy proportion of a certain frequency band, the peak frequency point) and the statistical features (such as the mean and variance of the current pulse) in the vector as inputs, with the input dimension consistent with the number of features, for example, a ten-dimensional vector corresponding to a ten-dimensional input layer. The number of neuron nodes in the input layer matches the feature dimension, and each node receives the corresponding dimension of feature parameters, such as the first node receiving the energy feature value in the time-frequency domain, the second node receiving the frequency feature value, and so on.

[0110] The hidden layer is set to three non-linear transformation layers, each containing 20 neurons. Taking the first transformation layer as an example, when the input data is transmitted from the input layer to this layer, each neuron performs linear combination on the input features, i.e., the input of each neuron is the product of a ten-dimensional feature vector and a ten-dimensional weight matrix, plus a bias term, and then the output is processed by a ReLU activation function for non-linear processing. The activation function can introduce non-linear characteristics and enhance the model's ability to express complex fault features. The twenty-dimensional vector output by the first layer is transmitted to the second transformation layer, and the above linear combination and activation process is repeated. The weight matrix and bias term of the second layer are trained independently to further extract abstract relationships in the features. The third transformation layer also processes and outputs the final transformation result, for example, obtaining a five-dimensional feature expression vector that integrates key fault information from the original features.

[0111] The output layer transmits the final transformation output to a threshold comparison function, which contains four preset hierarchical threshold intervals corresponding to the normal, prompt, warning, and fault four levels of early warning level values. For example, when the comprehensive value of the five-dimensional output vector is less than 0.3, it is determined to be normal, 0.3-0.6 is prompt, 0.6-0.8 is warning, and greater than 0.8 is fault. The number of neuron nodes in the output layer is consistent with the number of early warning levels, and the softmax function is used to convert the transformation output into probability values of each early warning level, and the level with the maximum probability is taken as the final early warning level value.

[0112] In the threshold dynamic adjustment link, the system determines the adjustment weight of the hierarchical threshold based on the occurrence probability of historical fault samples and the stability of the current operating environment. Taking contact oxidation failure as an example, historical data shows that this failure occurs 2 times per month in the past year, which meets the characteristics of Poisson distribution, and the Poisson distribution parameter λ is set to 2. The occurrence probability weight of the failure in the current month is calculated by the Poisson distribution function. For example, when 1 failure has been recorded in the current month, the probability of failure in the remaining days is calculated, and the probability weight is 0.4. At the same time, the temperature fluctuation range of the current operating environment is ±5℃, the time length of the humidity duration exceeding 80%RH is 2 hours, and the vibration acceleration value is 0.5g. These environmental monitoring indicators are input into the stability weight calculation model. The model sets weight coefficients for temperature, humidity and vibration respectively (such as 0.4, 0.3, 0.3), and the weight increases by 0.1 for every ±2℃ of temperature fluctuation, 0.05 for every hour of humidity duration exceeding the threshold, and 0.05 for every 0.1g of vibration acceleration. Finally, the stability weight is calculated as 0.6. The adjustment weight is 0.5 by the weighted sum formula (probability weight × 0.5 + stability weight × 0.5).

[0113] When executing the early warning operation, multiply the early warning level value by the adjustment weight and match the early warning strategy. If the output layer determines that the early warning level value is prompt (such as 0.4), multiply it by the adjustment weight 0.5 to get 0.2, and trigger the prompt level early warning strategy at this time. The specific operation is: call the real-time data of the environmental monitoring device, get the temperature, humidity, vibration and other parameters of the current switch, calculate the running state deviation, such as the temperature rising by 3℃ higher than the reference value, the humidity exceeding the reference value by 10%RH, forming a state deviation vector. According to the state deviation vector and the early warning level value 0.2, calculate the monitoring frequency increment, for example, increase the original 10kHz sampling frequency by 20% to 12kHz, modify the sampling parameters of the signal acquisition unit through the sampling rate adjustment algorithm, and increase the monitoring density to the level corresponding to the prompt level, increase the data acquisition frequency to more densely monitor the switch state.

[0114] When the early warning level value is warning (such as 0.7), multiply it by the adjustment weight 0.5 to get 0.35, and trigger the warning level early warning strategy. The system uses fault case library matching technology to calculate the similarity between the current fault judgment vector and the fault characteristic templates stored in the case library, such as contact oxidation, coil turn-to-turn short circuit, etc. For example, the similarity with the contact oxidation template is calculated to be 85%, and the similarity with the coil turn-to-turn short circuit template is calculated to be 30%, so the potential fault type is determined to be contact oxidation. Based on the potential fault type and the early warning level value 0.35, the fault probability of the contact component in the switch is marked by the fault positioning rule, such as setting the fault probability of the contact component to 70% when the contact oxidation fault occurs, and highlighting the component on the monitoring interface, and generating a warning information to push to the operation and maintenance personnel, prompting to check the contact.

[0115] If the early warning level value is a fault (such as 0.9), multiplied by the adjustment weight 0.5, 0.45 is obtained, which exceeds the fault threshold 0.4, triggering the fault level early warning strategy. The system immediately sends an action instruction to the hardware protection device to control the relay to cut off the fault circuit. The action time of the relay is determined by the severity of the fault determination vector, for example, when the fault determination vector value is 0.9, the action time is set to 50ms, ensuring isolation of the fault in the shortest time. At the same time, the communication module sends fault alarm information to the monitoring terminal, including fault type (such as contact oxidation), occurrence time and component position (such as A-phase contact), and the operation and maintenance personnel can quickly locate and handle the fault according to the information.

[0116] In practical application, the calculation of the threshold adjustment weight needs to consider the type difference of the measuring switch. For example, the outdoor switch is greatly affected by environmental temperature and humidity, and the weight coefficient of the environmental monitoring index can be adjusted to 0.6, while the environmental weight coefficient of the indoor switch is set to 0.4. The probability calculation of the historical fault sample needs to be updated quarterly, and the Poisson distribution parameters are adjusted according to the latest fault data to ensure the timeliness of the probability weight.

[0117] On the hardware level, the early warning triggering unit can be implemented by an embedded processor, such as an ARM Cortex-A53 chip, with DDR4 memory to store historical fault samples and environmental parameters. The communication module selects a 4G / 5G module to ensure real-time transmission of fault alarm information. On the software level, the threshold calculation and early warning strategy matching program is developed using Python language, and the neural network model of the hidden layer is built through TensorFlow framework, and the model training data comes from the fault simulation samples and field test data provided by the switch factory.

[0118] To adapt to measuring switches of different voltage levels, the preset interval of the hierarchical threshold needs to be differentiated. The fault threshold of 10kV switch can be set to 0.8, while the fault threshold of 35kV switch can be adjusted to 0.75 due to higher insulation requirements. The threshold can be flexibly modified through the system parameter configuration interface. The collection frequency of environmental monitoring index is related to the early warning level, which is collected every 10 minutes in normal state, every 5 minutes in prompt state, and every minute in warning state, to ensure the timeliness of environmental data in high-risk state.

[0119] The neural network structure of input layer, hidden layer and output layer in this embodiment realizes the preliminary determination of fault level, dynamically adjusts the threshold weight combined with historical fault probability and real-time environmental parameters, finally triggers the hierarchical early warning strategy, forms a complete closed loop from feature processing to operation action, and realizes accurate early warning and response to measuring switch faults.

[0120] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0121] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A system for identifying and warning of switch failure, characterized in that, The method comprises the following steps: a signal acquisition unit: collecting a contact potential fluctuation sequence through a contact voltage sensor and obtaining an excitation current pulse sequence through a coil current transformer; the potential fluctuation sequence contains amplitude-time corresponding data, and the current pulse sequence contains phase-frequency correlation information; a feature analysis unit: extracting a time-frequency domain feature group based on the potential fluctuation sequence and extracting a statistical domain feature group based on the current pulse sequence; the time-frequency domain feature group is calculated through a wavelet transform algorithm, and the statistical domain feature group is obtained through a moment function analysis technique; a fault determination unit: performing multi-modal feature fusion on the time-frequency domain feature group and the statistical domain feature group to generate a fault determination vector; the multi-modal feature fusion adopts a kernel principal component analysis strategy, and the fusion weight is set according to the feature saliency; a pre-warning triggering unit: performing hierarchical matching on the fault determination vector by using a dynamically updated hierarchical threshold to trigger a pre-warning instruction of a corresponding level; the hierarchical threshold includes a reference value trained by a historical fault sample and a dynamic value adaptively adjusted online; the specific way of extracting the time-frequency domain feature group based on the potential fluctuation sequence comprises: step A: selecting a feature analysis window in the potential fluctuation sequence, and determining the time length and frequency resolution of the window; step B: decomposing the analysis window signal into an energy distribution matrix on the time-frequency plane through wavelet transform, and the time-frequency resolution of the energy distribution matrix is determined by the scale parameter of the wavelet basis function; step C: based on the peak position, frequency band width and time delay information of the energy distribution matrix, a three-dimensional feature tensor containing energy features, frequency features and time features is constructed as the time-frequency domain feature group; the specific way of performing multi-modal feature fusion on the time-frequency domain feature group and the statistical domain feature group to generate a fault determination vector comprises: establishing a time-frequency feature space model based on the energy features, frequency features and time features in the time-frequency domain feature group; establishing a statistical feature distribution model based on the first-order moment features and second-order moment features in the statistical domain feature group; aligning the time-frequency feature space model and the statistical feature distribution model through feature space mapping, embedding statistical information into the time-frequency space, and forming a fault determination vector containing time-frequency characteristics and statistical attributes.

2. The metrology switch fault identification and warning system of claim 1, wherein, The specific way of collecting the contact potential fluctuation sequence through the contact voltage sensor and obtaining the excitation current pulse sequence through the coil current transformer comprises: the contact voltage sensor monitors the contact potential difference at a fixed sampling frequency, converts the analog signal into digital potential data; the coil current transformer senses the coil magnetic flux change at a synchronous sampling frequency, generates digital current data through a signal conditioning unit; the digital potential data is calibrated to a unified time axis through a time synchronization module, and the digital current data is filtered to remove high-frequency noise through a band-pass filter unit.

3. The metrology switch fault identification and warning system of claim 1, wherein, The specific way of extracting the statistical domain feature group based on the current pulse sequence comprises: the statistical domain feature group contains first-order moment features and second-order moment features; the mean and variance of the current pulse are calculated through a probability density function to obtain the first-order moment value and the second-order moment value; the skewness and kurtosis values of the distribution are calculated based on the change trend of the first-order moment value and the second-order moment value; the mean, variance, skewness and kurtosis values are integrated into a vector form as the statistical domain feature group.

4. The metrology switch fault identification and warning system of claim 1, wherein, The specific manner of performing hierarchical matching on the fault determination vector by using the dynamically updated hierarchical threshold value includes: An input layer, which takes time-frequency features and statistical features in the fault determination vector as inputs, and sets the input dimension consistent with the number of features; A hidden layer, which sets multiple nonlinear transformation layers; the input data is transmitted from the input layer to the first transformation layer, processed by linear combination and an activation function to obtain the first layer output, and then transmitted to the next transformation layer, repeatedly processed by nonlinear functions until the last transformation layer is completed to obtain the final transformation output; An output layer, which transmits the final transformation output to the output layer to generate an early warning level value through a threshold comparison function; the early warning level value includes four levels of normal, prompt, warning, and fault; Calculating a threshold weight, determining the adjustment weight of the hierarchical threshold value based on the occurrence probability of historical fault samples and the stability of the current operating environment; Performing an early warning operation, matching the early warning level value multiplied by the adjustment weight to the corresponding early warning strategy to complete the early warning trigger.

5. The metrology switch fault identification and warning system of claim 4, wherein, The specific manner of determining the adjustment weight of the hierarchical threshold value based on the occurrence probability of historical fault samples and the stability of the current operating environment includes: For the occurrence probability of historical fault samples, the probability weight is calculated by a Poisson distribution function; the parameters of the Poisson distribution function are set according to the historical frequency of fault types; For the stability of the current operating environment, the stability weight is calculated by an environmental monitoring index; the environmental monitoring index includes temperature fluctuation range, humidity duration, and vibration acceleration value; Combining the probability weight and the stability weight, the adjustment weight of the hierarchical threshold value is calculated by a weighted sum formula.

6. The metrology switch fault identification and warning system of claim 5, wherein, When the early warning level value is prompt, the corresponding early warning strategy is executed, specifically including: When the early warning type is prompt, the real-time data of the environmental monitoring device is used to obtain the operating state deviation of the current switch; the monitoring frequency increment is calculated by the operating state deviation and the early warning level value; for the monitoring frequency increment, the sampling frequency of the signal acquisition unit is increased by using a sampling rate adjustment algorithm to improve the monitoring density to the level corresponding to the prompt level.

7. The metrology switch fault identification and warning system of claim 6, wherein, When the early warning level value is warning, the corresponding early warning strategy is executed, specifically including: The fault case library matching technology is used to identify the potential fault type corresponding to the current fault determination vector; the potential fault type includes contact oxidation and coil turn-to-turn short circuit; based on the potential fault type and the early warning level value, the fault probability of the corresponding component in the switch is marked by a fault positioning rule to complete the fault early warning prompt.

8. The metrology switch fault identification and warning system of claim 7, wherein, When the early warning level value is fault, the corresponding early warning strategy is executed, specifically including: Triggering the action of the hardware protection device to cut off the fault circuit by controlling the relay; the action time of the control relay is determined by the severity of the fault determination vector; the fault alarm information including the fault type, occurrence time, and component position is sent to the monitoring terminal through the communication module to complete the fault level early warning execution.

Images