Low-voltage arc fault detection method and device and electronic equipment

Abstract

The invention provides a low-voltage arc fault detection method and apparatus, and an electronic device. The method comprises the steps of obtaining m current data of a target low-voltage power distribution network within a preset time period; m is a positive integer greater than or equal to 1, feature extraction is performed on the m current data based on a preset feature extraction method, m first feature sets are obtained, and each first feature set comprises n first features; n is a positive integer greater than or equal to 24, performing correlation calculation on n first features in the target first feature set to obtain a correlation index scores, and performing feature selection on the target first feature set according to the a correlation index scores by adopting a preset feature selection method to obtain a target second feature set; and inputting the target second feature set into a preset arc detection model to obtain a target arc detection result. When arc fault detection is carried out, the accuracy and the stability of fault detection can be improved.

Description

Technical Field

[0001] This application relates to the field of power safety detection technology, and in particular to a method, device and electronic equipment for detecting low-voltage arc faults. Background Technology

[0002] Currently, with the continuous innovation of smart grid technology and the rapid increase in electricity load demand, the safety hazards involved in low-voltage distribution networks are becoming increasingly complex. Among these safety hazards, electric arc faults, due to their instantaneous high temperature and electromagnetic impact, are prone to causing electrical fires and have now become a key target for prevention in low-voltage distribution network safety. Existing arc detection technologies mainly extract a series of features from the time and frequency domains, or steady-state and transient perspectives, and select highly correlated features to detect arc faults. One approach uses the ReliefF algorithm, which evaluates the correlation between features and categories by calculating the feature distance between samples, and further utilizes these correlations for arc prediction. Another approach uses feature selection methods such as FCBF, ReliefF, and Fisher-Score to rank the feature set, combine the outputs of each feature selection method, and determine the final selected features through a weighted average ranking. A threshold is set to filter out the higher-ranked features for arc detection in low-voltage distribution networks. However, existing feature selection methods for arc fault detection, such as ReliefF, information gain, gain ratio, and Fisher-Score, only consider the importance of the selected arc fault features to the arc fault labels, without taking into account the redundancy between features. In addition, most studies simply weight or rank the evaluation indicators of different feature selections, failing to construct a comprehensive evaluation indicator that can simultaneously measure high relevance and low redundancy, thus affecting the screening effect and stability.

[0003] Therefore, improving the accuracy and stability of arc fault detection is an urgent issue to be addressed. Summary of the Invention

[0004] This application provides a low-voltage arc fault detection method, device, and electronic device, which improves the accuracy and stability of arc fault detection.

[0005] In a first aspect, embodiments of this application provide a low-voltage arc fault detection method, applied to electronic equipment, the method comprising:

[0006] Obtain m current data points from the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1.

[0007] Based on a preset feature extraction method, features are extracted from the m current data to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24;

[0008] The correlation of the n first features in the target first feature set is calculated to obtain a correlation index scores; the target first feature set is any one of the n first feature sets.

[0009] A preset feature selection method is used to select features from the first feature set of the target based on the scores of the a relevance indicators, thereby obtaining the second feature set of the target;

[0010] The target second feature set is input into a preset electric arc detection model to obtain the target electric arc detection result.

[0011] Secondly, embodiments of this application provide a low-voltage arc fault detection device, applied to electronic equipment, the device comprising:

[0012] The acquisition unit is used to acquire m current data points of the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1.

[0013] The calculation unit is used to extract features from the m current data based on a preset feature extraction method to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24; perform correlation calculation on the n first features in the target first feature set to obtain a correlation index scores; the target first feature set is any one of the n first feature sets; and use a preset feature selection method to select features from the target first feature set according to the a correlation index scores to obtain a target second feature set.

[0014] The control unit is used to input the second feature set of the target into a preset arc detection model to obtain the target arc detection result.

[0015] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs include instructions for performing steps in any method of the first aspect of this application.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in any method of the first aspect of this application.

[0017] Fifthly, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in any method of the first aspect of this application. The computer program product may be a software installation package.

[0018] By implementing the embodiments of this application, the following beneficial effects are achieved:

[0019] This application describes a low-voltage arc fault detection method, device, and electronic equipment. It acquires m current data points from a target low-voltage distribution network within a preset time period, where m is a positive integer greater than or equal to 1. Based on a preset feature extraction method, it extracts features from the m current data points to obtain m first feature sets, where each first feature set includes n first features, where n is a positive integer greater than or equal to 24. It then performs correlation calculations on the n first features in the target first feature sets to obtain a correlation index scores. The target first feature set is any one of the n first feature sets. A preset feature selection method is used to select features from the target first feature set based on the a correlation index scores to obtain a target second feature set. Finally, the target second feature set is input into a preset arc detection model to obtain the target arc detection result. Thus, on the one hand, a feature comprehensive evaluation index that simultaneously considers the relationship between features and arc faults as well as the redundancy between features is constructed, and a feature selection process is proposed based on this index. This process can effectively select beneficial arc fault features from the high-dimensional feature set, thereby significantly reducing the feature dimensionality and improving the stability of subsequent arc fault detection. On the other hand, an arc fault detection model based on the CNN-SENet-HKELM hybrid model is proposed. This model first extracts deep features from arc fault features through the CNN-SENet module, and replaces the fully connected layers in the neural network with a hybrid extreme learning machine. The kernel of this hybrid extreme learning machine is composed of a mixture of Gaussian kernels and multinomial kernels, which can improve the accuracy of arc fault detection. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a system architecture diagram of a low-voltage arc fault detection method provided in an embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application;

[0023] Figure 3 This is a schematic flowchart of a low-voltage arc fault detection method provided in an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of a feature selection process provided in an embodiment of this application;

[0025] Figure 5 This is a flowchart illustrating another low-voltage arc fault detection method provided in an embodiment of this application;

[0026] Figure 6 This is a framework diagram of an arc detection model provided in an embodiment of this application;

[0027] Figure 7 This is a schematic flowchart of an arc detection method provided in an embodiment of this application;

[0028] Figure 8 This is a block diagram of the functional modules of a low-voltage arc fault detection device provided in an embodiment of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0030] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0031] It should be understood that the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document indicates that the preceding and following related objects are in an "or" relationship. In the embodiments of this application, "multiple" refers to two or more.

[0032] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0033] In this application, the term "connection" refers to various connection methods, such as direct connection or indirect connection, to achieve communication between devices. This application does not impose any limitations on this.

[0034] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0035] The following is an explanation of the relevant terms used in this application:

[0036] The Laplacian matrix: The Laplacian matrix is ​​a matrix used in graph theory to represent the structure of a graph. It is also called the admittance matrix or Kirchhoff matrix. It is defined as the difference between the degree matrix D and the adjacency matrix A, i.e., L = DA, where D is the matrix representing the degree of vertices corresponding to the diagonal elements, and A is the adjacency matrix of the graph. The Laplacian matrix is ​​symmetric and all its eigenvalues ​​are non-negative real numbers.

[0037] Fisher score: Fisher's discriminant analysis is a method of discriminant analysis, and the Fisher score is used to measure the performance of Fisher discriminant analysis. Fisher uses the idea of ​​analysis of variance to construct one or more linear discriminant functions using the p-dimensional observations of samples drawn from each population. The coefficient is determined by maximizing the deviation between different populations and minimizing the deviation within the same population. This coefficient is the Fisher score.

[0038] When detecting arc faults in low-voltage distribution networks, feature selection technology is used to extract current data features. However, most existing technologies have problems such as insufficient consideration of feature redundancy, significant influence of ranking thresholds on the feature selection process, and lack of a comprehensive feature evaluation index that can balance high relevance and redundancy. These issues affect the screening effect and stability of feature selection in low-voltage arc detection, and consequently affect the accuracy and stability of low-voltage arc detection.

[0039] To address the aforementioned problems, this application provides a low-voltage arc fault detection method, apparatus, and electronic device. Applied to electronic equipment, it acquires m current data points from a target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1. Based on a preset feature extraction method, features are extracted from the m current data points to obtain m first feature sets, where each first feature set includes n first features; n is a positive integer greater than or equal to 24. Correlation calculations are performed on the n first features in the target first feature set to obtain a correlation index scores. A preset feature selection method is then used to select features from the target first feature set based on the a correlation index scores, resulting in a target second feature set. The target second feature set is input into a preset arc detection model to obtain the target arc detection result. This approach can improve the accuracy and stability of arc fault detection.

[0040] The following is combined with Figure 1 The system architecture of a low-voltage arc fault detection method according to an embodiment of this application is described below. Figure 1 This is a system architecture diagram of a low-voltage arc fault detection method provided in an embodiment of this application. The low-voltage arc fault detection system 100 includes a low-voltage arc detection device 110, a data acquisition module 120, a feature processing module 130, and an arc fault detection module 140.

[0041] The low-voltage arc detection device 110 is used to monitor the operating status of low-voltage electrical lines in real time, collect electrical signal data when an arc fault occurs, and provide original information for subsequent fault detection. The low-voltage arc detection device 110 includes multiple arc detection devices, including: arc detection device A111, arc detection device B112, and arc detection device C113 (the number can be set according to actual needs). As a core component of the system's perception layer, it can be deployed in key locations such as low-voltage distribution boxes and the incoming lines of electrical equipment. Through built-in current and voltage sensors, it captures abnormal electrical characteristics such as sudden current changes and voltage distortions, covering various low-voltage power consumption scenarios (such as residential homes, small commercial spaces, and low-voltage circuits in industrial workshops), accurately identifying the signal characteristics of the initiation and development stages of arc faults.

[0042] In one possible embodiment, the low-voltage arc detection device 110 includes an arc detection device A111, an arc detection device B112, and an arc detection device C113. The arc detection devices A111, B112, and C113 may have the same structure and function, or they may all be different, or at least one may be the same; this is not limited here. Arc detection device A111 focuses on time-domain feature monitoring. Based on a high-speed sampling circuit (sampling frequency up to 10kHz-100kHz), it acquires time-domain waveform data of current and voltage signals, extracts time-domain features such as current mutation rate and voltage sag depth, and is used to initially identify transient distortions in electrical signals caused by arc faults. Arc detection device B112 focuses on frequency-domain feature analysis. Through a built-in Fourier transform module, it converts the acquired time-domain electrical signals into frequency-domain signals, extracts frequency-domain features such as harmonic components and spectral energy distribution, and captures spectral anomalies caused by arc faults, such as a surge in harmonic content at specific frequencies. Arc detection device C113 focuses on multi-feature fusion monitoring. It integrates time-domain and frequency-domain acquisition circuits and algorithms, extracts multi-dimensional features and makes preliminary fusion judgments, quickly screens suspected fault signals, reduces the data processing pressure of subsequent modules, and the various sub-devices work together to comprehensively cover the multi-domain manifestations of arc fault characteristics.

[0043] The data acquisition module 120 receives the raw electrical signal data transmitted by the low-voltage arc detection device 110 and performs data preprocessing and preliminary integration. It first performs synchronous calibration on the acquired multi-channel, multi-type electrical signals to ensure consistency in the time base of data acquired from different devices and channels. Then, it performs denoising processing, employing wavelet denoising and adaptive filtering algorithms to remove electromagnetic interference in the electrical environment (such as high-frequency noise generated by frequency converters and switching power supplies) and noise introduced by measurement errors, thus restoring the arc fault characteristic signals. Simultaneously, it constructs a data buffer queue to store the preprocessed data in a time sequence, providing a stable and orderly data stream for the feature processing module 130, ensuring the continuity of subsequent feature extraction and fault detection. In actual operation, the data acquisition module 120 monitors the working status of the low-voltage arc detection device 110 in real time. When it detects that the data collected by a certain sub-device (such as arc detection device A111) is abnormal (such as a sudden change in signal caused by sensor failure), it automatically triggers the redundant data switching mechanism, calls historical normal data or complementary data from other sub-devices of the same type (such as arc detection device B112), maintains the reliability of data acquisition, and avoids the overall detection process of the system being affected by the failure of a single device.

[0044] The feature processing module 130 is used to perform feature extraction and feature selection operations on the preprocessed data transmitted by the data acquisition module 120. First, 24 initial arc fault features, such as RMS current, voltage peak factor, harmonic order, and spectral entropy, are extracted from the preprocessed electrical signal according to the time domain, frequency domain, and time-frequency domain. Then, using the feature selection technique based on maximum correlation and minimum redundancy, the redundancy index between features (obtained by linear weighting of Pearson correlation coefficient and Spearman correlation coefficient) and the correlation index with arc fault are calculated. The set of candidate features that are highly correlated with arc fault and have low redundancy is selected to provide accurate and concise feature input for the arc fault detection module 140, thereby improving the computational efficiency and accuracy of the fault detection model.

[0045] In one possible embodiment, the feature processing module 130 has online feature updating and adaptive filtering capabilities. When a new electrical device is connected to the low-voltage power distribution system (such as adding a variable frequency speed control device, which will change the distribution of electrical signal characteristics), causing the original feature set to decrease in its ability to identify arc faults, the module automatically triggers a feature re-extraction and re-filtering process to re-analyze the electrical signals, mine arc fault characteristics under the new operating conditions, update the candidate feature set, and ensure the system's adaptability to changing power consumption scenarios. For example, after adding charging pile equipment to a commercial building, the feature filtering strategy can be adjusted in a timely manner to accurately identify the feature changes in the low-voltage circuit caused by arc faults during the operation of the charging pile.

[0046] The arc fault detection module 140 receives candidate features output by the feature processing module 130 and performs arc fault detection tasks based on the constructed CNN-SENet-HKELM hybrid model. This module first loads the trained model parameters, inputs the candidate features into the CNN-SENet network (composed of a single-layer one-dimensional convolutional layer and an SE module), and extracts deep-level fault features. Then, it uses a hybrid kernel extreme learning machine (composed of a linearly weighted multinomial kernel and a Gaussian kernel) to classify the deep-level features and outputs the arc fault detection result (determined as fault or normal). Simultaneously, the module incorporates a model evaluation and self-optimization mechanism. It periodically tests the model's detection accuracy using newly collected labeled data. When the detection accuracy, recall, or other metrics decline, it automatically triggers a model retraining process to update the model parameters and maintain high reliability in fault detection.

[0047] In one possible embodiment, the low-voltage arc detection device 110 is deployed in the household distribution box and key branch circuits (such as air conditioning and kitchen power circuits) to collect electrical signals during peak and off-peak electricity consumption periods. The data acquisition module 120 denoises and synchronizes the signals before transmitting them to the feature processing module 130, which filters out key features such as current harmonic distortion rate and voltage fluctuation coefficient. The arc fault detection module 140 uses a trained hybrid model to determine in real time whether there is an arc fault in the circuit. When a fault is detected, it quickly triggers a local alarm (such as an audible and visual alarm) and a remote notification (pushed to the user's mobile APP and property management system) to eliminate potential low-voltage arc fault hazards in a timely manner and ensure household electricity safety.

[0048] As can be seen, the system architecture of the aforementioned low-voltage arc fault detection system 100 constructs a complete process from low-voltage arc signal perception, data processing, feature optimization to fault detection and model self-maintenance. The modules work collaboratively to achieve accurate and real-time detection of arc faults in low-voltage power distribution systems. It is adaptable to diverse low-voltage power consumption scenarios such as residential, commercial, and industrial applications, effectively identifying arc fault characteristics and improving the accuracy and stability of low-voltage electrical systems. Furthermore, it can provide technical support for early warning and timely handling of low-voltage arc faults, contributing to the construction of a safe low-voltage power distribution network.

[0049] The following is combined with Figure 2 The electronic devices in the embodiments of this application will be described. Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 2 As shown, the electronic device 200 includes one or more processors 210, a memory 220, a communication interface 230, and one or more programs 221. The processor 210 is communicatively connected to the memory 220 and the communication interface 230 via an internal communication bus.

[0050] The processor 210 can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, cells, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication unit can be a communication interface, a transceiver, a transceiver circuit, etc., and the storage unit can be a memory.

[0051] The memory 220 can be volatile memory or non-volatile memory, or it can include both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0052] The one or more programs 221 are stored in the memory 220 and configured to be executed by the processor 210. The one or more programs 221 include instructions for performing any step in the following embodiment of the low-voltage arc fault detection method.

[0053] It is understood that the electronic device 200 may include more or fewer structural elements than those shown in the above block diagram, such as a power module, physical buttons, a Wi-Fi module, a speaker, a Bluetooth module, sensors, a display module, etc., without limitation. It is understood that the electronic device may incorporate elements such as... Figure 1 The system architecture of the low-voltage arc fault detection method is described above.

[0054] After understanding the software and hardware architecture of this application, the following will be combined with... Figure 3 This application describes a low-voltage arc fault detection method according to an embodiment. Figure 3 This is a flowchart illustrating a low-voltage arc fault detection method provided in an embodiment of this application, applied to electronic equipment. The method specifically includes the following steps:

[0055] Step S310: Obtain m current data points of the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1.

[0056] The current data refers to the set of time-series electrical quantity data collected in real time by current sensors installed at different locations during the operation of the target low-voltage distribution network, reflecting the operating status of distribution lines and electrical equipment. The current data includes both current waveforms under normal operating conditions and current characteristic waveforms during arc faults, short circuits, or other abnormal conditions, thus providing a raw data foundation for subsequent feature extraction, feature selection, and the construction of arc fault detection models. The preset time period refers to a continuous time interval determined based on the detection task requirements, distribution network operating characteristics, and data analysis granularity. This time period can be flexibly configured according to different application scenarios. For example, in rapid fault detection scenarios, short time periods ranging from minutes to hours can be selected to capture the sudden characteristics of arc faults; in operating status assessment or long-term statistical analysis scenarios, time periods ranging from days to weeks can be selected to reflect the periodic characteristics and trend changes of equipment operation. No specific limitation is imposed here.

[0057] Specifically, current data acquisition is accomplished by multi-point electronic equipment deployed throughout the target low-voltage distribution network. These devices typically consist of high-precision Hall effect current sensors, current transformers (CTs), or current measurement units based on shunt resistors, coupled with high-sampling-rate data acquisition modules to achieve continuous, real-time monitoring of line currents. For different acquisition points, the system can be deployed at feeder outgoing ends, branch lines, important load incoming ends, and inside key electrical equipment, thus forming a comprehensive monitoring system covering the entire low-voltage distribution network. The acquired analog signals are processed by analog-to-digital conversion (A / D) and transmitted to the data aggregation unit in digital signal form. During data acquisition, sampling timestamps are recorded synchronously to ensure time alignment of current data from different acquisition points, facilitating subsequent feature extraction and time-series analysis. Simultaneously, to ensure data accuracy and stability, the electronic equipment can integrate filtering modules to resist power frequency interference and transient impacts, and, when necessary, configure isolation amplifier circuits to prevent high-voltage surges from damaging the acquisition system. In addition, for multiphase systems, the instantaneous or effective values ​​of the current in each phase need to be recorded separately. If necessary, the current difference between phases and the phase sequence information can also be recorded to analyze the phase imbalance in the characteristic calculation.

[0058] It should be noted that, to accommodate the sensitivity of arc fault detection to high-frequency details, the sampling frequency should be set according to the Nyquist sampling theorem, typically no less than 100 times the fundamental frequency of the power grid (50Hz), i.e., 5kHz or higher, to capture the high-frequency components and waveform distortion characteristics generated by the arc fault. Simultaneously, the acquired data must meet quantization accuracy requirements; the analog-to-digital converter resolution should ideally be 12 bits or higher to ensure accurate recording of even minute current changes. After acquisition, the m current data points are stored in a multi-dimensional time-series data matrix according to the acquisition point number and time sequence. The rows of the matrix represent sampling time points, and the columns represent different acquisition points or different feature quantities. For example, when m=8, it indicates that a total of 8 current acquisition channels are deployed within the target low-voltage distribution network, with each channel corresponding to a complete set of time-series data. In addition, considering the potential problems of noise interference, sampling packet loss and transmission delay in the field environment, a real-time data verification and repair mechanism can be integrated into the data aggregation unit, including CRC (cyclic redundancy check) for detecting transmission errors, interpolation compensation algorithm for lost data, and outlier removal method based on sliding window, thereby improving the availability and reliability of current data.

[0059] Step S320: Based on a preset feature extraction method, feature extraction is performed on the m current data to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24.

[0060] Feature extraction refers to obtaining a set of feature parameters that can effectively characterize the differences between arc faults and normal operation from m collected raw current time-series data through feature engineering, based on preset mathematical analysis methods and signal processing algorithms. The goal of feature extraction is to transform high-dimensional, redundant raw waveform data into representative, information-rich, and easily processed low-dimensional feature vectors, thus providing high-quality data input for subsequent feature selection and fault detection model training.

[0061] The n first features are a multi-dimensional feature set constructed according to time-domain and frequency-domain features, which can comprehensively reflect the amplitude characteristics, waveform shape, spectral distribution, and statistical regularities of the current signal. The time-domain features may include parameters such as RMS, peak value, current crest coefficient, zero-sequence time, integral value, skewness, kurtosis, average value, and standard deviation. These indicators can directly reflect the energy distribution and shape characteristics of the current waveform on the time axis. The frequency-domain features can be obtained by performing a Fast Fourier Transform (FFT) on the current waveform, including parameters such as the amplitude of the 1st to 15th harmonics and the total harmonic distortion (THD), used to characterize the energy distribution and harmonic components of the signal in the spectrum. In addition, to ensure the comparability and stability of feature calculation results, all feature values ​​need to be normalized before being output to the first feature set. For example, min-max scaling or Z-score standardization can be used to map features with different dimensions and different value ranges to a unified numerical range. This is to avoid the bias caused by differences in feature dimensions to subsequent feature selection and model training, and to improve the balance of features in the comprehensive evaluation index.

[0062] In one possible embodiment, the step of extracting features from the m current data based on a preset feature extraction method to obtain m first feature sets specifically includes the following steps:

[0063] 321. Preprocess the m current data to obtain m first current data;

[0064] 322. Determine the time-domain statistical data of each of the m first current data to obtain m time-domain statistical feature sets; each of the m time-domain statistical feature sets includes p time-domain statistical features; p is an integer less than n;

[0065] 323. Generate m current frequency spectrum data based on the m first current data using Fourier transform;

[0066] 324. Extract the harmonic component features of each current spectrum data in the m current frequency spectrum data to obtain m frequency domain harmonic feature sets; each frequency domain harmonic feature set in the m frequency domain harmonic feature sets includes q frequency domain harmonic features; q is an integer less than n; p + q = n;

[0067] 325. Determine the m first feature sets based on the m frequency domain harmonic feature sets and the m time domain statistical feature sets.

[0068] Where m is a positive integer greater than or equal to 1, the current data is acquired by current sensors installed at different sampling points in the target low-voltage distribution network, and the acquired content includes multi-phase current waveform information of each sampling point within a preset time period. The purpose of preprocessing is to improve the accuracy and stability of subsequent feature extraction, including data denoising, outlier removal, missing value completion, and time synchronization alignment, which are not limited here. In the data denoising process, low-pass filtering, band-pass filtering, or wavelet denoising methods can be used to suppress high-frequency interference and random noise other than power frequency. In the outlier removal process, abrupt changes can be detected and replaced using the sliding window midpoint method or the three-standard-deviation method. In the missing value completion process, linear interpolation, spline interpolation, or prediction based on adjacent sampling windows can be combined to fill missing sampling points. In the time synchronization process, the timing consistency of the current data from multiple sampling points is ensured by unifying the timestamp and correcting the deviation of the sampling clock. After preprocessing, m first current data points with continuous waveforms, suppressed noise, and amplitude accuracy meeting the requirements are obtained.

[0069] Among them, time-domain statistical characteristics are statistical quantities calculated directly based on the amplitude distribution of the current waveform on the time axis, used to characterize the energy level, waveform morphology, and amplitude variation characteristics of the signal. For example, p can be taken as 8, including: effective value (RMS), peak value, crest coefficient, zero-sequence time, integral value, average value, skewness, kurtosis, etc. Among them, the effective value reflects the overall energy of the waveform, the peak value reflects the instantaneous maximum amplitude of the waveform, the crest coefficient is used to describe the ratio of the peak value to the effective value, which is used to identify impulsive loads or fault transients, the zero-sequence time represents the duration of the current waveform near zero, which reflects the conduction or disconnection characteristics, the integral value can be used to measure the total current over a certain period of time, and the average value, skewness, and kurtosis characterize the waveform symmetry, skewness, and peak sharpness from a statistical perspective. These time-domain characteristics will show significant changes when an arc fault occurs, such as an increase in peak value, a change in skewness, or a shortening of the zero-sequence time. The Fast Fourier Transform (FFT) is used to transform the first current data of each channel to obtain the corresponding frequency spectrum. The transformation process can be based on whole-cycle sampling data to ensure the accuracy of amplitude and phase calculation of spectral components. Window functions (such as Hanning windows and Hamming windows, which are not limited here) are used to suppress spectral leakage. The frequency spectrum data contains amplitude and phase information from the fundamental wave to several harmonics, which can reflect the energy distribution of the current signal at different frequency components. Frequency domain harmonic characteristics are important parameters for characterizing the distribution characteristics of the current signal in the spectrum, revealing phenomena such as the increase of high-frequency components and the aggravation of harmonic distortion caused by arc faults. For example, q can be 16, including indicators such as the amplitude of the 1st to 15th harmonics and the total harmonic distortion (THD). Among them, the amplitude of each harmonic is used to identify the intensity of non-fundamental components in the current. The variation law of different harmonics under different fault types is different. The total harmonic distortion comprehensively reflects the energy ratio of all harmonic components relative to the fundamental wave and is an important indicator for measuring power quality.

[0070] Each first feature set contains n features (n = p + q), preferably n = 24, consisting of 8 time-domain statistical features and 16 frequency-domain harmonic features. This retains the ability to characterize the overall energy and waveform of the signal while integrating the ability to analyze harmonic structure and spectral distribution, achieving complementarity of time and frequency information. Before feature combination, normalization can be performed. For example, min-max normalization can be used to map all feature values ​​to the [0,1] interval, eliminating the influence of differences in dimensions and amplitudes between different features. Feature smoothing and noise reduction methods can also be introduced to reduce the interference of abnormal fluctuations on the stability of the feature set. Furthermore, the feature extraction method is not limited to the above-mentioned combination of time-domain and frequency-domain features. Time-frequency analysis features can also be introduced into the feature set, such as short-time Fourier transform (STFT) energy distribution, wavelet packet decomposition (WPD) band energy, and empirical mode decomposition (EMD) intrinsic mode component mean, to further enhance the sensitivity to transient signals of arc faults. The specific values ​​of p, q, and n can be adjusted according to the accuracy requirements of the detection task, the limitations of computing resources, and the feature selection strategy to achieve a balance between computational efficiency and detection accuracy. No restrictions are imposed here.

[0071] Step S330: Perform correlation calculation on the n first features in the target first feature set to obtain a correlation index scores; the target first feature set is any one of the n first feature sets; a is greater than or equal to n.

[0072] Correlation calculation assesses the similarity in numerical distribution or trend among features within the same feature set to measure feature redundancy. Correlation scores reflect the linear or non-linear dependency between any two features, providing a basis for subsequent maximum correlation minimum redundancy (mRMR) feature selection. The n×n correlation scores are represented in matrix form, where rows and columns correspond to feature numbers, and the element in the i-th row and j-th column represents the correlation score between the i-th and j-th features. Correlation scores can be calculated in various ways, such as using the Pearson correlation coefficient to assess linear relationships; using Spearman's rank correlation coefficient to assess the consistency of feature ranking; or using a weighted fusion of the Pearson and Spearman coefficients to preserve linear relationship information while also capturing non-linear monotonic trends.

[0073] In one possible embodiment, the step of calculating the correlation of the n first features in the target first feature set to obtain a correlation index scores specifically includes the following steps:

[0074] 331. Calculate the Fisher scores of the n first features based on the preset first formula to obtain n first Fisher scores;

[0075] 332. Based on the preset second formula, determine the Laplace scores of the n first features to obtain n first Laplace scores;

[0076] 333. Based on the preset third formula, the n first Fisher scores, and the n first Laplace scores, determine the scores of the a correlation indicators.

[0077] Fisher score measures the separability of each feature across different classes, reflecting the feature's ability to distinguish between categories. A higher score indicates that the mean difference between categories is more significant than the within-class variance, and thus contributes more to the classification. The first formula is the formula for calculating Fisher score:

[0078]

[0079] Among them, F i This represents the Fisher score for the i-th feature, where C is the total number of categories, c is the c-th category in the classification labels, and N is the number of categories. c μ is the number of samples in class c. i,c It is the average value of the i-th feature in the c-th class, μ i This represents the mean of the i-th feature across all samples. It is the variance of the i-th feature in the c-th class.

[0080] The Laplacian score is used to evaluate the ability of features to preserve the local structure of the data, reflecting the fidelity of features in preserving the similarity structure between samples under unsupervised conditions. The second formula is used to calculate the Laplacian score, and its expression is:

[0081]

[0082] Among them, L i It is the Laplace score of the i-th feature, x j,i x is the i-th feature value of the j-th sample. k,i W is the i-th feature value of the k-th sample. jk f represents the weight between the j-th sample and the k-th sample in the Laplace adjacency graph. i D represents the expected value obtained from the i-th feature. jj This indicates the degree of clustering of other samples around the j-th sample.

[0083] Specifically, the Fisher score is first positively normalized to map it to the [0,1] interval; then the Laplacian score is negatively normalized so that a higher score indicates a stronger feature correlation. The normalization formula is as follows:

[0084]

[0085] in, F is the Fisher score after normalization of the i-th feature, and max(F) and min(F) represent the maximum and minimum Fisher scores for all features, respectively. Let L be the normalized Laplace score for the i-th feature, and max(L) and min(L) represent the maximum and minimum Laplace scores for all features, respectively.

[0086] Finally, the Fisher score and Laplace score of each feature are linearly weighted using the third formula to obtain the relevance index, which represents the relevance score. The third formula is:

[0087]

[0088] Among them, R i Let α be the relevance index of the i-th feature, and α be the weighting factor, where α∈[0,1].

[0089] In one possible embodiment, determining the Laplace scores of the n first features based on a preset second formula to obtain n first Laplace scores specifically includes the following steps:

[0090] 3321. Calculate the similarity score of the n first features in the m first feature set to obtain m*n similarity scores;

[0091] 3322. Based on the m*n similarity scores and the n features, construct a weight matrix between the m first feature sets and the n first features to obtain a Laplacian matrix; the Laplacian matrix is ​​an m*n dimensional matrix;

[0092] 3323. Generate a Laplace adjacency graph based on the Laplace weight matrix; any edge in the Laplace adjacency graph corresponds to a similarity score in the m*n;

[0093] 3324. Determine the first weight of the target's first feature based on the Laplace adjacency graph; the target's first feature is any one of the n first features;

[0094] 3325. Determine the target expected value of the first feature of the target;

[0095] 3326. Based on the second formula, the target expected value, and the m*n similarity scores, determine the n first Laplace scores.

[0096] The similarity score measures the degree of similarity between two features in the sample space, reflecting the feature's ability to preserve the local structure of the data. The similarity score can be calculated using the Gaussian radial basis function.

[0097]

[0098] Among them, S u,v Indicates sample x u and sample x v The similarity score, where t is the kernel width parameter; S u,v The range is (0,1], S u,v The closer the value is to 1, the higher the similarity of the special diagnosis on the sample.

[0099] In this matrix, the elements correspond to the similarity scores of the respective features on the sample pairs, representing the adjacency relationships between feature dimensions. The Laplacian matrix is ​​typically constructed from the weight matrix W and the degree matrix D, and is calculated using the following formula:

[0100] L=DW

[0101] Where L is the Laplacian matrix, and D is the degree matrix, which is a diagonal matrix; the i-th diagonal element of D is the sum of the weights of the i-th node, i.e. W ij This represents the weight coefficient of the i-th row and j-th column.

[0102] Here, the m×n-dimensional Laplacian matrix represents the ability of each feature to preserve local structure across all sample pairs. The Laplacian adjacency graph is an undirected weighted graph where nodes represent features or samples, and edge weights represent the similarity scores between nodes. Adjacency graphs are used to model the local structural relationships between features; the larger the edge weight between any two nodes, the closer they are in the feature space, thus receiving a higher evaluation for local structure preservation in the Laplacian score calculation. The first weight refers to the correlation strength between the target first feature and other features in the adjacency graph, which can be obtained by summing the edge weights between the feature node and all its adjacent nodes. The magnitude of the first weight reflects the centrality of the feature in the local structure of the feature space; a larger weight indicates a higher contribution of the feature to preserving the local manifold structure. The target expected value is used to measure the average value level of the target first feature across the entire sample space, and its calculation formula is:

[0103]

[0104] in, E(x) represents the value of the i-th feature in the k-th sample, where m is the total number of samples. i ) represents the expected value of the i-th feature.

[0105] Specifically, in calculating the n first Laplace scores, the values ​​of each feature are first standardized to have a mean of 0 and a variance of 1, thus eliminating the influence of different feature units. Next, the weighted squared difference of each feature over all sample pairs is calculated, and then summed using adjacency graph weights. Finally, this weighted squared difference is divided by the overall variance of the feature to obtain the Laplace score. It should be noted that during the feature selection stage, features with low Laplace scores are preferentially retained to ensure that the final feature subset is highly correlated with the classification label while maximizing the preservation of local structural information between samples.

[0106] Step S340: Using a preset feature selection method, feature selection is performed on the first feature set of the target based on the scores of the a relevance indicators to obtain the second feature set of the target.

[0107] Feature selection, in particular, involves selecting a subset of features from n features in the target feature set that are highly relevant to the target task and have low redundancy, based on a pre-defined feature comprehensive evaluation criterion. The goal of this method is to retain as much useful information as possible while reducing the computational overhead and model overfitting risk caused by redundant features, thereby improving the generalization ability and operational efficiency of the subsequent arc fault detection model. The feature selection method can employ the maximum relevance-minimum redundancy (mRMR) criterion, which prioritizes selecting candidate features with the highest relevance score to the classification label and the lowest redundancy with any feature in the already selected feature set in each iteration, thus constructing an incremental feature subset.

[0108] Specifically, all n features from the target first feature set are placed into the candidate feature set, and an empty selected feature set is initialized. The feature with the highest relevance score to the classification label is selected from the candidate feature set, added to the selected feature set, and then removed from the candidate feature set. For the remaining candidate features, their relevance score to the classification label is calculated, and their redundancy score with each feature in the selected feature set is also calculated. The maximum redundancy score is taken as the redundancy index of that candidate feature.

[0109] According to the preset comprehensive evaluation formula:

[0110] Score(f)=Rel(f,Label)-λ×Redundancy(f,Selected)

[0111] Where Rel(f,Label) represents the relevance score between feature f and label, Redundancy(f,Selected) represents the maximum redundancy score between feature f and any feature in the selected feature set, and λ is the redundancy penalty coefficient, with a value ranging from 0 to 1.

[0112] Then, the candidate feature with the highest comprehensive evaluation score is selected, added to the selected feature set, and removed from the candidate set; this process is repeated until the number of selected features reaches the preset feature quantity limit, or the candidate feature set is empty, or the comprehensive evaluation score is lower than the preset threshold. Finally, the selected feature set is output as the target second feature set, which will be used as input features in subsequent model training and inference.

[0113] It should be noted that feature selection methods can employ other algorithms based on relevance and redundancy evaluation, such as Sequential Forward Selection (SFS), Recursive Feature Elimination (RFE), or feature ranking methods based on mutual information. In different practical applications, a suitable feature selection strategy can be flexibly chosen based on the feature dimension size, computational resource constraints, and real-time requirements.

[0114] In one possible embodiment, the step of using a preset feature selection method to select features from the first target feature set based on the scores of the a relevance indicators to obtain the second target feature set includes:

[0115] 341. Determine the correlation coefficient between each of the n first features to obtain a correlation coefficients;

[0116] 342. Determine the redundancy index a based on the scores of the a correlation indicators and the a correlation coefficients;

[0117] 343. Based on the preset fourth formula, the scores of the a correlation indicators, and the a redundancy indicators, determine a comprehensive evaluation indicators;

[0118] 344. Extract the target comprehensive evaluation indicators that are greater than the preset first threshold, obtain the target first feature corresponding to the target comprehensive evaluation indicator, and obtain the target second feature set.

[0119] The correlation coefficient is a statistical indicator that measures the degree of linear correlation between two features. Its value ranges from -1 to 1, with positive values ​​indicating a positive correlation and negative values ​​indicating a negative correlation. The closer the absolute value is to 1, the stronger the correlation. It is calculated using the Pearson correlation coefficient, and its formula is as follows:

[0120]

[0121] Among them, CO(X) i ,X j ) represents the i-th feature X i and the j-th feature X j The Pearson correlation coefficient is obtained between them, where n is the total number of samples and x is the number of samples. k,i X represents i The observation value of the k-th sample, x k,j X represents j The observation value of the kth sample; X represents i X j The sample mean.

[0122] The redundancy index measures the degree of information overlap between features, reflecting the potential repetitive information carried by each feature. If two features have high absolute correlation coefficients and similar correlation index scores, it indicates a high degree of information redundancy. The purpose is to combine feature relevance (measured by the correlation coefficient) with feature importance (measured by the correlation index score) to provide a more comprehensive assessment of redundancy. The redundancy index is calculated using the following formula:

[0123]

[0124] P(X i ,X j )=β·CO(X i ,X j )+(1-β)·SC(X i ,X j ),β∈[0,1]

[0125] Among them, SC(X) i ,X j ) represents the i-th feature X i and the j-th feature X j Spearman correlation coefficient between them, P(X) i ,X j ) represents the i-th feature X i and the j-th feature X j Redundancy index between samples, where n represents the total number of samples, dk β represents the difference in the order number of the k-th sample between the two features, and β represents the weight.

[0126] The fourth formula is:

[0127]

[0128] in, Represents the i-th feature X in the candidate feature set. i X in the selected feature set T j The maximum value of the redundancy index between them.

[0129] The comprehensive evaluation index measures the overall quality of features. Its principle is to maximize relevance to the target task while minimizing redundancy with existing features, representing an implementation of the maximum relevance-minimum redundancy (mRMR) feature selection approach. During calculation, the relevance score and redundancy score are first normalized to ensure consistent value ranges, avoiding the influence of dimensional differences on the comprehensive evaluation results. The first threshold is a screening boundary set according to task requirements. Its function is to filter out feature pairs with low comprehensive evaluation scores and limited contribution to the overall feature selection. The first threshold can be the mean of the comprehensive evaluation index or the mean plus one standard deviation, or it can be adaptively determined through methods such as cross-validation to balance model accuracy and feature dimensionality. When the comprehensive evaluation index is greater than the first threshold, it indicates that the feature has high relevance to other features and low redundancy, making it suitable for inclusion in the final feature subset. By extracting feature pairs that meet the conditions and adding the relevant features to the target second feature set, feature redundancy can be effectively reduced, improving the training efficiency and prediction accuracy of subsequent models.

[0130] Specifically, in the feature selection process, the first target feature set needs to be preprocessed, including missing value imputation, outlier removal, and feature normalization, to ensure the accuracy of the correlation coefficient and relevance index scores. Subsequently, an optimized feature subset (i.e., the target second feature set) is obtained. This feature set not only retains feature information highly relevant to the target task but also minimizes information overlap between different features, thereby reducing the risk of overfitting and improving generalization ability during subsequent model training.

[0131] To make it easier to understand, please participate Figure 4 , Figure 4This is a flowchart illustrating a feature selection process provided in an embodiment of this application. As can be seen, the first step is to "extract a series of fault arc features" as the basic input for feature selection, obtaining the original feature set to be screened. Next, "all features are placed into the candidate feature set U, and the selected feature set T is empty." Initially, all extracted fault arc features constitute the candidate feature set U, and the selected feature set T is empty, preparing for subsequent iterative screening. Then, the process proceeds to the step of "calculating the correlation index Ri of all features in the feature set and selecting the feature with the highest Ri to be placed into the selected feature set T" (Ri is calculated according to the third formula mentioned above), constructing the correlation between features within the feature set. By calculating the correlation of each feature in the candidate feature set, the most relevant feature is selected and added to the selected feature set T, establishing the initial key feature foundation. Then, "calculate the comprehensive feature evaluation index Si of the candidate feature set, find the feature corresponding to max(Si), and if max(Si)≥0, put it into the selected feature set T" (Si is calculated according to the fourth formula above). Calculate the correlation index between the candidate feature and the arc fault, as well as the redundancy index between the candidate feature and the features in the selected feature set T. Then, calculate the comprehensive feature evaluation index Si according to the formula, and select the feature with the largest Si and ≥0, moving it from the candidate feature set U to the selected feature set T to achieve iterative feature inclusion. Finally, "the candidate feature set is empty or max(Si)<0" is used as the loop termination condition. If the maximum value of the comprehensive feature evaluation index Si in the candidate feature set is less than or equal to 0, it indicates that the remaining features have weak correlation with the arc fault or high redundancy and do not need to be included; or the candidate feature set is empty, with no features to filter, and the loop ends. At this point, "the features in the selected feature set are the result of feature selection," completing the arc fault feature selection process and obtaining the final key feature set for tasks such as arc fault analysis. Through iterative screening, an optimized feature set that is closely related to arc faults and reduces redundancy between features is gradually constructed, providing high-quality feature input for subsequent applications such as arc fault detection.

[0132] Step S350: Input the target second feature set into the preset arc detection model to obtain the target arc detection result.

[0133] The arc detection model can be a pattern recognition model based on machine learning or deep learning, such as Support Vector Machine (SVM), Random Forest, Gradient Boosting Tree (GBDT), Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), or a time-series analysis model based on attention mechanisms. No limitation is imposed here. Preferably, the arc detection model is a hybrid neural network model that integrates convolutional feature extraction and time-series dependency analysis, capable of simultaneously extracting the correlation patterns of input features in both the time and frequency domains.

[0134] The training data for the arc detection model includes electrical signal features collected under various operating conditions, such as current waveform features, harmonic components, spectral features, and instantaneous power features. Supervised training is then performed using known labels indicating whether an arc has occurred or not, enabling the model to distinguish between normal operation and arc faults. The training process may include steps such as data augmentation, feature normalization, cross-validation, and hyperparameter optimization to improve the model's generalization ability and robustness under different operating conditions.

[0135] In one possible embodiment, the arc detection model includes a cumulative excitation model and a hybrid kernel extreme learning machine classification model. The step of inputting the target second feature set into the preset arc detection model to obtain the target arc detection result specifically includes the following steps:

[0136] 351. Input the target second feature set into the accumulation excitation model to obtain the deep feature vector;

[0137] 352. Input the deep feature vector into the hybrid kernel extreme learning machine classification model to obtain the target electric arc detection result.

[0138] The stacked activation model is used to perform deep nonlinear mapping and multi-scale feature fusion on the input target second feature set to extract high-dimensional deep feature vectors that can more fully represent the characteristics of the electric arc. The idea behind the stacked activation model is to introduce nonlinear activation units during the multi-layer feature transformation process. By stacking convolutional layers, fully connected layers, and normalization layers, it progressively mines the local patterns and global correlations in the target second feature set. Simultaneously, the model introduces skip connections and residual structures to alleviate the gradient vanishing or gradient exploding problems that may occur in deep networks, improving the stability and generalization ability of feature representation. Specifically, this can be formalized as follows:

[0139] u = Cov(x)

[0140]

[0141] s = F ex (Z,W)=Sig(W2·relu(W1·z))F=s·u

[0142] Where x represents the feature set X * A training sample feature vector, Cov(x), represents a one-dimensional convolution operation on the input feature x, where u is the output of the one-dimensional convolutional layer, and its dimension is represented as H×C, where H is the feature sequence length, C is the number of feature channels, and F... sq (u c ) is the squeeze operation in the SE block, u c It is the c-th channel component of u, u c(i) is the i-th element of the c-th channel dimension of u, z c is u c The output after the compression operation, z is the output obtained after the compression operation of all channel components in u, F ex (Z,W) denotes the activation operation in the SE block, W denotes the weight matrix, W1 is the weight coefficient matrix of the first fully connected layer, W2 is the weight coefficient matrix of the second fully connected layer, Sig(·) denotes the Sigmoid activation function, relu(·) denotes the Relu activation function, s denotes the feature vector, u denotes the feature vector adapted to the s dimension, and F denotes the final output.

[0143] The hybrid kernel extreme learning machine classification model is a classification algorithm based on Extreme Learning Machine (ELM). It utilizes a hybrid kernel function (such as a linear combination of a Gaussian radial basis function and a polynomial kernel function) to map deep feature vectors to a high-dimensional feature space, thereby improving the nonlinear fitting ability of the classification boundary. During the model training phase, the hybrid kernel parameters and kernel weight coefficients are optimized through cross-validation and grid search to ensure that the classifier has high-precision discrimination ability between different arc types and non-arc signals.

[0144] Specifically, the second feature set of the target is first normalized, mapping each feature value to the interval [0,1] or [-1,1] to eliminate the influence of feature dimension differences on the model's computational stability. Then, the normalized feature set is input into the input layer of the cumulative activation model. The input layer can be directly connected to a one-dimensional convolutional layer or a fully connected layer for preliminary feature mapping. In the intermediate feature extraction layer, multiple nonlinear activation units (such as ReLU, LeakyReLU, ELU, etc.) are used, combined with batch normalization and residual connection mechanisms, to perform multi-scale fusion and deep representation learning on the features. In the output layer, the multi-layer feature mapping results are flattened or global average pooling is performed to obtain a fixed-length deep feature vector, which is then used as the input to the hybrid kernel extreme learning machine classification model. A classifier is rapidly constructed using a stochastic hidden layer weight initialization strategy of Extreme Learning Machine (ELM), and the output weight matrix is ​​analytically solved using the least squares method, achieving efficient training and inference of the classification model. Based on the output of the hybrid kernel ELM, the target arc detection result is obtained, which may include: arc state label (present / absent); arc type classification result (e.g., series arc, parallel arc, etc.); and detection confidence score (probability value between 0 and 1). Deep features are extracted from the target's second feature set using a cumulative excitation model, and then efficiently classified using a hybrid kernel ELM. This not only improves the separability of arc features but also significantly reduces computational complexity while maintaining high detection accuracy, thus meeting the engineering requirements for real-time performance and robustness in arc detection.

[0145] With the above Figure 3 For embodiments consistent with those described, please refer to [link / reference]. Figure 5 , Figure 5 This is a flowchart illustrating another low-voltage arc fault detection method provided in this application embodiment. Before inputting the target second feature set into a preset arc detection model to obtain the target arc detection result, the method specifically includes the following steps:

[0146] S510. Divide the m first feature sets according to a preset ratio to obtain a training set and a test set;

[0147] S520. Train the first arc detection model based on the training set to obtain the first arc detection result and the second arc detection model;

[0148] S530. Test the second arc detection model based on the test set to obtain the second classification result;

[0149] S540. Fine-tune the second arc detection model according to the second classification result to obtain the arc detection model.

[0150] The preset ratio is typically determined based on the sample size and model generalization requirements, and can be 7:3 or 8:2. The partitioning process requires a stratified sampling strategy to ensure that the proportions of normal operating condition samples and arc fault samples in the training and test sets are consistent, thus avoiding model training bias towards the majority class due to data distribution imbalance. For example, if the original m first feature sets contain 60% normal samples and 40% fault samples, then the training and test sets must maintain the same proportion. Furthermore, the feature sets must be randomly shuffled before partitioning to eliminate the potential impact of the data collection order on model training and ensure that sample independence satisfies the statistical assumptions.

[0151] The first arc detection model is an initially constructed CNN-SENet-HKELM hybrid model, which consists of a single-layer one-dimensional convolutional layer, an SE module, and a hybrid kernel extreme learning machine (HKELM). The training process is divided into two stages: the first stage is the pre-training of the CNN-SENet network, where feature vectors from the training set are input into a one-dimensional convolutional layer, and feature mapping is performed through 3×1 convolutional kernels (the number is set to 16 or 32 depending on the feature dimension) to generate intermediate features with dimensions of [feature sequence length, number of channels]. Subsequently, the SE module dynamically adjusts the channel weights. The squeezing operation uses global average pooling to calculate the mean of each channel feature, and the activation operation uses two fully connected layers (the hidden layer dimension is 1 / 4 of the number of channels) and a sigmoid activation function to generate channel attention weights, thereby enhancing key feature channels. The second stage involves training HKELM, using the deep features output by CNN-SENet as input. A kernel matrix is ​​constructed using a linear weighted combination of multinomial and Gaussian kernels, and the output weights are solved using the least squares criterion to complete the fault classification mapping. The first arc detection result is the model's classification output on the training set during training, used to calculate the cross-entropy loss function. The Adam optimizer (with an initial learning rate of 0.001 and a decay rate of 0.9) is then used for backpropagation to update the convolutional kernel parameters and SE module weights of CNN-SENet, ultimately resulting in a second arc detection model with initially converged parameters.

[0152] The testing process based on the test set aims to evaluate the model's generalization ability by calculating multi-dimensional metrics to generate secondary classification results. These metrics include accuracy, precision, recall, and F1 score. During testing, the test set features are input into the second arc detection model, and the fault labels output by the model (1 for fault, 0 for normal) are recorded. These labels are then compared with the true labels to generate a confusion matrix, and the aforementioned metrics are calculated. If fluctuations in the metrics occur during testing (e.g., recall rate below 95%), the feature distribution of misclassified samples needs to be analyzed to determine if feature selection bias or model overfitting exists.

[0153] Specifically, the fine-tuning process aims to optimize model parameters based on the second classification results to improve detection performance, mainly adjusting three key modules: First, the kernel function parameters of HKELM, optimizing the order d of the polynomial kernel (range 1-3) and the bandwidth γ of the Gaussian kernel (range 10⁻³ to 10⁻³) using a grid search method. 1 The first step is to optimize the model's performance. The second step involves adjusting the convolutional layer parameters of CNN-SENet. If insufficient feature extraction exists, the kernel size can be fine-tuned or the number of kernels increased, and the attention weights of the SE module can be retrained. The third step is an early stopping strategy. By monitoring changes in the test set loss, training is terminated when the loss does not decrease for five consecutive epochs to avoid overfitting. Fine-tuning needs to be iterative. After each adjustment, the model's performance on the test set is re-evaluated until accuracy ≥ 98%, recall ≥ 99%, and F1 score ≥ 98.5%, meeting the high reliability requirements for low-voltage arc fault detection. The final arc detection model needs to have its network parameters, kernel function configuration, and feature mapping relationships saved for direct loading and use in subsequent iterations.

[0154] For easier understanding, please refer to Figure 6 , Figure 6This is a framework diagram of an arc detection model provided in this application embodiment. As can be seen, this model is a hybrid model based on CNN-SENet-HKELM, used to achieve deep extraction of arc fault features and fault detection. The specific process and principle are as follows: The model takes "arc fault features" as input, first processed by "one-dimensional convolution." The input feature vector is processed by one-dimensional convolution to generate intermediate features. After completing the initial feature mapping, local correlation information of the features is mined. Next, the "squeeze and excitation (SE) module" further processes the one-dimensional convolution output. First, a global average pooling operation is performed on each channel component through a squeezing operation to obtain channel-level statistics; then, an excitation operation is performed through a fully connected layer containing ReLU and Sigmoid activation functions to generate channel attention weights, dynamically weighting each channel of the intermediate features to highlight key feature channels and output "deep-level features," strengthening the feature expression with discriminative power for arc faults. After this, the "deep-level features" are input into a "hybrid kernel extreme learning machine classification model." After freezing the CNN-SENet network parameters, a hybrid kernel extreme learning machine replaces the fully connected layer and softmax layer. The hybrid kernel is composed of a linearly weighted multinomial kernel and a Gaussian kernel. After receiving deep features, it is trained and outputs the "arc fault detection result". First, a CNN-SENet network is trained to extract deep features. Then, a preliminary fault detection result is output through fully connected layers and softmax layers. The network parameters are updated using a classification loss function (such as cross-entropy loss) combined with the backpropagation algorithm. The CNN-SENet parameters are then frozen, and the hybrid kernel extreme learning machine is connected and trained. Finally, a CNN-SENet-HKELM hybrid model is formed, realizing the complete process from arc fault feature input to detection result output. Through the collaboration of deep feature extraction and the hybrid kernel classifier, the accuracy and robustness of arc fault detection are improved.

[0155] For easier understanding, please refer to Figure 7 , Figure 7 This is a schematic flowchart of an arc detection method provided in an embodiment of this application. As can be seen, Figure 7To achieve accurate arc fault detection, the process involves data processing, feature engineering, and model building. Specifically, the first step is to "collect arc fault data and form a dataset," which serves as the foundational input for the entire process. This data includes information on various electrical parameters and operating states at the time of the arc fault, providing raw material for subsequent analysis and covering multi-dimensional feature records under arc fault scenarios. Next, "data preprocessing" involves cleaning, denoising, and imputing missing values ​​in the collected data to ensure data quality and prevent abnormal or erroneous data from affecting subsequent feature extraction and model training, thus laying a solid data foundation for accurate analysis. Then, the process moves to "feature extraction, extracting 24 arc fault features from the time and frequency domains." This involves mining 24 features that characterize arc faults from both the time domain (such as the amplitude, mean, and variance of current and voltage, reflecting the characteristics of signal changes over time) and the frequency domain (such as spectral components obtained through Fourier transform and energy distribution, showcasing the frequency domain characteristics of the signal). This comprehensively depicts the performance of arc faults in different domains and enriches the feature dimensions. The next step is "feature selection," employing a feature selection technique based on maximum correlation and minimum redundancy to obtain beneficial features. This technique filters features, retaining those highly correlated with arc faults (maximum correlation) while eliminating redundant and repetitive information (minimum redundancy). From the 24 original extracted features, the most valuable candidate feature set for arc fault detection is selected, improving the quality of subsequent model input and reducing computational complexity. Finally, the process of "constructing an arc fault detection model based on a CNN-SENet-HKELM hybrid model" is executed. Using the previously processed candidate features, following the model training process (e.g., first training a CNN-SENet network to extract deep features, then connecting it to a hybrid kernel extreme learning machine for classification), a detection model integrating convolution, attention mechanisms, and hybrid kernel classification is built and trained. This achieves accurate identification of arc faults and outputs arc fault detection results, completing a full closed loop from data acquisition to model construction and application. Through multi-stage collaboration, the accuracy and effectiveness of arc fault detection are ensured.

[0156] As can be seen, by implementing the low-voltage arc fault detection method provided in this application embodiment, m current data points of the target low-voltage distribution network within a preset time period are obtained, where m is a positive integer greater than or equal to 1. Based on a preset feature extraction method, features are extracted from the m current data points to obtain m first feature sets, where each first feature set includes n first features, where n is a positive integer greater than or equal to 24. Correlation calculation is performed on the n first features in the target first feature set to obtain a correlation index scores. The target first feature set is any one of the n first feature sets. A preset feature selection method is used to select features from the target first feature set based on the a correlation index scores to obtain a target second feature set. The target second feature set is then input into a preset arc detection model to obtain the target arc detection result. Thus, on the one hand, a feature comprehensive evaluation index that simultaneously considers the relationship between features and arc faults as well as the redundancy between features is constructed, and a feature selection process is proposed based on this index. This process can effectively select beneficial arc fault features from the high-dimensional feature set, thereby significantly reducing the feature dimensionality and improving the stability of subsequent arc fault detection. On the other hand, an arc fault detection model based on the CNN-SENet-HKELM hybrid model is proposed. This model first extracts deep features from arc fault features through the CNN-SENet module, and replaces the fully connected layers in the neural network with a hybrid extreme learning machine. The kernel of this hybrid extreme learning machine is composed of a mixture of Gaussian kernels and multinomial kernels, which can improve the accuracy of arc fault detection.

[0157] The above primarily describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the electronic device includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0158] This application embodiment can divide the electronic device into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0159] When dividing each function into modules according to its corresponding function. Figure 8 This is a functional module block diagram of a low-voltage arc fault detection device provided in an embodiment of this application. The low-voltage arc fault detection device 800 includes:

[0160] Acquisition unit 810 is used to acquire m current data points of the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1.

[0161] The calculation unit 820 is used to extract features from the m current data based on a preset feature extraction method to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24; perform correlation calculation on the n first features in the target first feature set to obtain a correlation index scores; the target first feature set is any one of the n first feature sets; and use a preset feature selection method to select features from the target first feature set according to the a correlation index scores to obtain a target second feature set.

[0162] The control unit 830 is used to input the target second feature set into a preset arc detection model to obtain the target arc detection result.

[0163] In one possible embodiment, the computing unit 820, in terms of extracting features from the m current data based on a preset feature extraction method to obtain m first feature sets, is specifically used for:

[0164] The m current data are preprocessed to obtain m first current data.

[0165] Determine the time-domain statistical data for each of the m first current data points to obtain m time-domain statistical feature sets; each of the m time-domain statistical feature sets includes p time-domain statistical features; p is an integer less than n;

[0166] m current frequency spectrum data are generated based on the m first current data using Fourier transform;

[0167] Harmonic component features are extracted from each of the m current frequency spectrum data to obtain m frequency domain harmonic feature sets; each of the m frequency domain harmonic feature sets includes q frequency domain harmonic features; q is an integer less than n; p + q = n;

[0168] The m first feature sets are determined based on the m frequency domain harmonic feature sets and the m time domain statistical feature sets.

[0169] In one possible embodiment, the calculation unit 820, in terms of performing correlation calculations on the n first features in the target first feature set to obtain m correlation index scores, is specifically used for:

[0170] Fisher scores for the n first features are calculated based on a preset first formula to obtain n first Fisher scores;

[0171] Based on the preset second formula, the Laplace scores of the n first features are determined, resulting in n first Laplace scores;

[0172] The scores of the m correlation indicators are determined based on the preset third formula, the n first Fisher scores, and the n first Laplace scores.

[0173] In one possible embodiment, the calculation unit 820, in determining the Laplace scores of the n first features based on a preset second formula to obtain n first Laplace scores, is specifically used for:

[0174] Calculate the similarity score of the n first features in the m first feature set to obtain m*n similarity scores;

[0175] Based on the m*n similarity scores and the n features, a weight matrix is ​​constructed between the m first feature sets and the n first features to obtain the Laplacian matrix; the Laplacian matrix is ​​an m*n dimensional matrix.

[0176] A Laplace adjacency graph is generated based on the Laplace weight matrix; any edge in the Laplace adjacency graph corresponds to a similarity score in the m*n matrix.

[0177] The first weight of the target's first feature is determined based on the Laplace adjacency graph; the target's first feature is any one of the n first features.

[0178] Determine the target expected value of the first feature of the target;

[0179] The n first Laplace scores are determined based on the second formula, the target expected value, and the m*n similarity scores.

[0180] In one possible embodiment, the computing unit 820, in the aspect of performing feature selection on the target first feature set according to the scores of the a relevance indicators using a preset feature selection method to obtain the target second feature set, is specifically used for:

[0181] Determine the correlation coefficient between each of the n first features to obtain a correlation coefficients;

[0182] The redundancy index a is determined based on the scores of the a correlation indicators and the a correlation coefficients.

[0183] Based on the preset fourth formula, the scores of the a correlation indicators, and the a redundancy indicators, a comprehensive evaluation indicators are determined.

[0184] Extract the target comprehensive evaluation indicators that are greater than a preset first threshold, obtain the target first feature corresponding to the target comprehensive evaluation indicator, and obtain the target second feature set.

[0185] In one possible embodiment, the arc detection model includes a cumulative excitation model and a hybrid kernel extreme learning machine classification model. Specifically, the computing unit 820, in inputting the target second feature set into the preset arc detection model to obtain the target arc detection result, is used for:

[0186] The target second feature set is input into the accumulation excitation model to obtain the deep feature vector;

[0187] The deep feature vector is input into the hybrid kernel extreme learning machine classification model to obtain the target electric arc detection result.

[0188] In one possible embodiment, before inputting the target second feature set into a preset arc detection model to obtain the target arc detection result, the computing unit 820 is further configured to:

[0189] The m first feature sets are divided according to a preset ratio to obtain a training set and a test set;

[0190] The first arc detection model is trained based on the training set to obtain the first arc detection result and the second arc detection model.

[0191] The second arc detection model is tested based on the test set to obtain the second classification result;

[0192] The second arc detection model is fine-tuned based on the second classification result to obtain the arc detection model.

[0193] As can be seen, the low-voltage arc fault detection device provided in this application acquires m current data points from a target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1. Based on a preset feature extraction method, features are extracted from the m current data points to obtain m first feature sets, where each first feature set includes n first features; n is a positive integer greater than or equal to 24. Correlation calculations are performed on the n first features in the target first feature set to obtain a correlation index scores. A preset feature selection method is then used to select features from the target first feature set based on the a correlation index scores, resulting in a target second feature set. The target second feature set is then input into a preset arc detection model to obtain the target arc detection result. This device can improve the accuracy and stability of arc fault detection.

[0194] This application also provides a computer-readable storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes an electronic device.

[0195] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include an electronic device.

[0196] It should be noted that, for the sake of simplicity, the above embodiments are all described as a series of actions. Those skilled in the art should understand that this application is not limited to the described order of actions, as some steps in the embodiments of this application can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions, steps, modules, or units involved are not necessarily essential to the embodiments of this application.

[0197] In the above embodiments, the descriptions of each embodiment in this application have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0198] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0199] The steps of the methods or algorithms described in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in RAM, flash memory, ROM, EPROM, electrically erasable programmable read-only memory (EEPROM), registers, hard disk, portable hard disk, read-only optical disk (CD-ROM), or any other form of storage medium well known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. Furthermore, the ASIC can reside in a terminal device or management device. Alternatively, the processor and storage medium can exist as discrete components in the terminal device or management device.

[0200] The modules / units included in the various devices and products described in the above embodiments can be software modules / units, hardware modules / units, or a combination of both. For example, for devices and products applied to or integrated into a chip, all modules / units can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits. For devices and products applied to or integrated into a terminal device, all modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components within the terminal device, or at least some modules / units can be implemented using software programs that run on a processor integrated within the terminal device, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0201] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above descriptions are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A method for detecting low-voltage arc faults, characterized in that, Applied to electronic devices, the method includes: Obtain m current data points from the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1. Based on a preset feature extraction method, features are extracted from the m current data to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24; The correlation of the n first features in the target first feature set is calculated to obtain a correlation index scores; the target first feature set is any one of the n first feature sets; a is greater than or equal to n; A preset feature selection method is used to select features from the first feature set of the target based on the scores of the a relevance indicators, thereby obtaining the second feature set of the target; The target second feature set is input into a preset electric arc detection model to obtain the target electric arc detection result.

2. The method as described in claim 1, characterized in that, The preset feature extraction method is used to extract features from the m current data to obtain m first feature sets, including: The m current data are preprocessed to obtain m first current data. Determine the time-domain statistical data for each of the m first current data points to obtain m time-domain statistical feature sets; each of the m time-domain statistical feature sets includes p time-domain statistical features; p is an integer less than n; m current frequency spectrum data are generated based on the m first current data using Fourier transform; Harmonic component features are extracted from each of the m current frequency spectrum data to obtain m frequency domain harmonic feature sets; each of the m frequency domain harmonic feature sets includes q frequency domain harmonic features; q is an integer less than n; p + q = n; The m first feature sets are determined based on the m frequency domain harmonic feature sets and the m time domain statistical feature sets.

3. The method as described in claim 1, characterized in that, The correlation calculation of the n first features in the target first feature set to obtain a correlation index scores includes: Fisher scores for the n first features are calculated based on a preset first formula to obtain n first Fisher scores; Based on the preset second formula, the Laplace scores of the n first features are determined, resulting in n first Laplace scores; The scores of the a correlation indicators are determined based on the preset third formula, the n first Fisher scores, and the n first Laplace scores.

4. The method as described in claim 3, characterized in that, The Laplace scores of the n first features are determined based on a preset second formula, resulting in n first Laplace scores, including: Calculate the similarity score of the n first features in the m first feature set to obtain m*n similarity scores; Based on the m*n similarity scores and the n features, a weight matrix is ​​constructed between the m first feature sets and the n first features to obtain the Laplacian matrix; the Laplacian matrix is ​​an m*n dimensional matrix. A Laplace adjacency graph is generated based on the Laplace weight matrix; any edge in the Laplace adjacency graph corresponds to a similarity score in the m*n matrix. The first weight of the target's first feature is determined based on the Laplace adjacency graph; the target's first feature is any one of the n first features. Determine the target expected value of the first feature of the target; The n first Laplace scores are determined based on the second formula, the target expected value, and the m*n similarity scores.

5. The method according to any one of claims 1-4, characterized in that, The step of using a preset feature selection method to select features from the first feature set of the target based on the scores of the a relevance indicators to obtain the second feature set of the target includes: Determine the correlation coefficient between each of the n first features to obtain a correlation coefficients; The redundancy index a is determined based on the scores of the a correlation indicators and the a correlation coefficients. Based on the preset fourth formula, the scores of the a correlation indicators, and the a redundancy indicators, a comprehensive evaluation indicators are determined. Extract the target comprehensive evaluation indicators that are greater than a preset first threshold, obtain the target first feature corresponding to the target comprehensive evaluation indicator, and obtain the target second feature set.

6. The method according to any one of claims 1-4, characterized in that, The arc detection model includes a cumulative excitation model and a hybrid kernel extreme learning machine classification model. The step of inputting the target second feature set into the preset arc detection model to obtain the target arc detection result includes: The target second feature set is input into the accumulation excitation model to obtain the deep feature vector; The deep feature vector is input into the hybrid kernel extreme learning machine classification model to obtain the target electric arc detection result.

7. The method according to any one of claims 1-4, characterized in that, Before inputting the target second feature set into the preset arc detection model to obtain the target arc detection result, the method further includes: The m first feature sets are divided according to a preset ratio to obtain a training set and a test set; The first arc detection model is trained based on the training set to obtain the first arc detection result and the second arc detection model. The second arc detection model is tested based on the test set to obtain the second classification result; The second arc detection model is fine-tuned based on the second classification result to obtain the arc detection model.

8. A low-voltage arc fault detection device, characterized in that, Applied to electronic devices, the device includes: The acquisition unit is used to acquire m current data points of the target low-voltage distribution network within a preset time period; m is a positive integer greater than or equal to 1. The calculation unit is used to extract features from the m current data based on a preset feature extraction method to obtain m first feature sets; the first feature sets include n first features; n is a positive integer greater than or equal to 24; perform correlation calculation on the n first features in the target first feature set to obtain a correlation index scores; the target first feature set is any one of the n first feature sets; and use a preset feature selection method to select features from the target first feature set according to the aa correlation index scores to obtain a target second feature set. The control unit is used to input the second feature set of the target into a preset arc detection model to obtain the target arc detection result.

9. An electronic device, characterized in that, include: Processor, memory, communication interface, and one or more programs; The one or more programs are stored in the memory and configured to be executed by the processor, the programs including instructions for performing the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-7.

Images