Wavelet analysis-based arc fault signal feature extraction method and related device

Abstract

The invention provides an arc fault signal feature extraction method based on wavelet analysis and a related device, and relates to the technical field of electrical safety monitoring. According to the method, self-adaptive filtering is achieved by collecting current and voltage signals and combining load type recognition, and an analysis window is dynamically adjusted according to the signal-to-noise ratio; carrying out multi-resolution analysis by adopting wavelet transform, and extracting typical features such as energy distribution and spectrum abrupt change points and wavelet entropy auxiliary features; and fault judgment is completed through deep learning and a support vector machine cascade model, confidence coefficient verification is carried out by fusing adjacent user data, and finally early warning information is uploaded. According to the scheme, the detection sensitivity and the anti-interference capability of weak arc are improved, the problems that a traditional circuit breaker has many recognition blind areas and the misoperation rate is high are effectively solved, and the circuit breaker is suitable for early fire early warning of an intelligent power distribution system.

Description

Technical Field

[0001] This invention belongs to the field of electrical safety monitoring technology, and relates to a method and related device for extracting features of arc fault signals based on wavelet analysis. Background Technology

[0002] With the continuous growth of residential electricity load, electrical fires are occurring frequently, with the proportion of fires caused by arc faults increasing year by year. In low-voltage power distribution systems, metering boxes, as key nodes for user-side power access, undertake the functions of electricity metering and line protection. Traditional circuit breakers mainly rely on current overload or short-circuit protection mechanisms, making it difficult to effectively identify series arc faults with small voltage fluctuations (5%-8%) and weak energy. Existing arc detection technologies mostly use time-domain threshold judgment or simple spectrum analysis, lacking the ability to extract refined features of non-stationary and transient arc signals, resulting in problems such as numerous detection blind spots and high false alarm rates in complex load environments.

[0003] The aforementioned existing technologies, due to their lack of adaptive processing for interference caused by different load types and the absence of high-resolution time-frequency analysis methods, cannot accurately capture the unique transient characteristics of arc faults. They are particularly prone to missed detections or false alarms under high-frequency noise interference, such as from switching power supplies. Furthermore, fixed-parameter filtering and static judgment thresholds are difficult to adapt to dynamic power environments, resulting in poor system robustness and severely impacting the reliability and practicality of early warning systems for arc faults. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and related device for extracting features of electric arc fault signals based on wavelet analysis, so as to solve the technical problems of low detection sensitivity, weak anti-interference ability and high false alarm rate caused by the crude signal processing and simple judgment mechanism of the traditional electric arc detection technology.

[0005] To achieve the above objectives, the present invention employs the following technical solution: A method for feature extraction of arc fault signals based on wavelet analysis, comprising: Acquire current and voltage signals; Based on the current signal, the type of dominant load is identified through a predefined load feature library. The corresponding filtering method is configured for the filtering unit corresponding to the dominant load. The current signal and voltage signal are filtered by the corresponding filtering method to obtain the preprocessed signal. During the filtering process, the width of the sliding window is adjusted according to the signal-to-noise ratio of the current signal and voltage signal. The preprocessed signal is analyzed using a wavelet transform algorithm to obtain a time-frequency domain feature map. Typical characteristic parameters of arc faults are extracted from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points and singular value decomposition index, and auxiliary features are calculated simultaneously. Deep learning is performed on typical feature parameters, and anomaly scoring is performed. The anomaly scores are compared with preset thresholds to obtain preliminary identification results. The preliminary identification results and auxiliary features are processed by support vector machine to determine whether there is an arc fault and generate an early warning signal. The warning signals and fault data are output to the cloud platform or local monitoring terminal.

[0006] A further improvement of the present invention is that: Preferably, the dominant load types include resistive loads, inductive loads, capacitive loads, and switching power supply loads; In the process of configuring the corresponding filtering method for the filtering unit corresponding to the dominant load, a hybrid strategy combining notch filter and Kalman filter is adopted for the high-frequency noise of switching power supply type loads.

[0007] Preferably, when the signal-to-noise ratio is lower than a first threshold, the window width is automatically narrowed to improve the ability to capture transient signals; when the signal-to-noise ratio is higher than a second threshold, the window width is widened to optimize computational efficiency and smooth random noise.

[0008] Preferably, the wavelet transform algorithm uses Daubechies wavelet basis functions for discrete wavelet transform, with a decomposition level of 6-8 levels.

[0009] Preferably, the auxiliary features include wavelet coefficient entropy and energy entropy.

[0010] Preferably, the support vector machine is periodically incrementally trained using an incremental learning sample library, which includes confirmed fault samples and their contextual features.

[0011] Preferably, the warning signal and fault data are analyzed simultaneously with the anonymized power consumption anomaly data of adjacent users under the same power distribution circuit. When the judgment result of this end is correlated with the anomaly data of the adjacent end in time, the confidence level of this judgment result is increased; otherwise, a review procedure is initiated.

[0012] An arc fault signal feature extraction device based on wavelet analysis, comprising: The signal acquisition module is used to acquire current and voltage signals; The signal preprocessing module is used to identify the type of dominant load based on the current signal through a predefined load feature library, configure the corresponding filtering method for the filtering unit corresponding to the dominant load, and filter the current signal and voltage signal through the corresponding filtering method to obtain the preprocessed signal; during the filtering process, the width of the sliding window is adjusted according to the signal-to-noise ratio of the current signal and voltage signal. The wavelet analysis module is used to perform multi-resolution analysis on the preprocessed signal using wavelet transform algorithm to obtain time-frequency domain feature maps. The feature extraction module is used to extract typical feature parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points and singular value decomposition index, and to calculate auxiliary features simultaneously. The fault judgment module is used to learn typical feature parameters through deep learning and perform anomaly scoring. The anomaly score is compared with a preset threshold to obtain a preliminary identification result. The preliminary identification result and auxiliary features are processed by a support vector machine to determine whether an arc fault exists and generate an early warning signal. The data communication module is used to output the warning signals and fault data to the cloud platform or local monitoring terminal.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements any one of the above-described wavelet analysis-based methods for extracting features from arc fault signals.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a wavelet analysis-based method for extracting features of an arc fault signal as described in any of the preceding claims.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a wavelet analysis-based method for extracting features from arc fault signals. This method introduces a load-aware adaptive filtering mechanism, which dynamically configures filtering strategies according to different load types, significantly improving signal preprocessing accuracy. Combined with signal-to-noise ratio-driven sliding window dynamic calibration, it balances transient capture capability with computational efficiency. Utilizing the multi-resolution analysis capability of Daubechies wavelets, it accurately extracts time-frequency local abrupt change features caused by arcs. By fusing typical features and entropy-based auxiliary features, it enhances sensitivity to concealed arcs. Employing a cascaded judgment structure of deep learning and support vector machines improves classification stability and generalization ability. Furthermore, it reduces the false alarm rate through a multi-node collaborative verification mechanism. This solves the problems of numerous detection blind spots, weak anti-interference capabilities, and high false alarm rates in traditional technologies, achieving high-sensitivity, low-false-alarm early warning of arc faults.

[0016] This invention discloses a wavelet analysis-based arc fault signal feature extraction device, comprising a signal acquisition module, a signal preprocessing module, a wavelet analysis module, a feature extraction module, a fault judgment module, and a data communication module. The signal acquisition module is connected to the metering box of the user-side power distribution line for real-time acquisition of current and voltage signals. The signal preprocessing module filters and reduces noise in the acquired raw signals to eliminate load fluctuations and high-frequency interference. The wavelet analysis module uses a wavelet transform algorithm to perform multi-resolution analysis on the preprocessed signal, generating a time-frequency domain feature spectrum. This invention achieves high-precision feature extraction of arc fault signals through wavelet analysis technology, significantly improving the detection sensitivity and anti-interference capability of weak arcs, and solving the problems of numerous blind spots and high false alarm rates in traditional circuit breakers. The fault judgment module can adapt to complex user-side power environments, reduce false alarm rates, and meet early warning requirements. Simultaneously, through digital communication and self-diagnosis mechanisms, it enhances the intelligence level of the community electrical safety control system. Attached Figure Description

[0017] Figure 1 This is a flowchart of a wavelet analysis-based method for extracting features from electric arc fault signals according to the present invention. Figure 2 This is a structural diagram of an arc fault signal feature extraction device based on wavelet analysis according to the present invention. Figure 3 This is a system structure diagram of Embodiment 1 of the present invention. Detailed Implementation

[0018] Hereinafter, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature.

[0019] The synchronization method provided in this application can be applied to mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, and ultra-mobile personal computers. In this application, the specific type of terminal device is not limited to terminal devices such as mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs).

[0020] It should be noted that the terms "first," "second," etc., used in the specification and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. 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 comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] See Figure 1 The first aspect of this invention discloses a method for extracting features from arc fault signals based on wavelet analysis, the method comprising the following steps: S1, acquire current and voltage signals; S2, based on the current signal, identifies the type of the dominant load through a predefined load feature library, configures the corresponding filtering method for the filtering unit corresponding to the dominant load, and filters the current signal and voltage signal through the corresponding filtering method to obtain the preprocessed signal; during the filtering process, the width of the sliding window is adjusted according to the signal-to-noise ratio of the current signal and voltage signal. S3, the preprocessed signal is analyzed by multi-resolution using wavelet transform algorithm to obtain time-frequency domain feature map; S4 extracts typical characteristic parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points and singular value decomposition index, and calculates auxiliary features simultaneously; S5. Perform deep learning on typical feature parameters and perform anomaly scoring. Compare the anomaly scores with a preset threshold to obtain preliminary identification results. Process the preliminary identification results and auxiliary features using a support vector machine to determine whether an arc fault exists and generate an early warning signal. S6, output the warning signal and fault data to the cloud platform or local monitoring terminal.

[0022] In the above process, S1, acquiring current and voltage signals refers to the real-time acquisition of raw electrical signals by high-precision sensors deployed inside the metering box of the user-side power distribution line. The signal acquisition module includes high-precision current and voltage sensors with a sampling frequency of no less than 100kHz, capable of capturing microsecond-level arc signal changes. Specifically, the current signal can be non-invasively sampled using a Rogowski coil or Hall effect current sensor, while the voltage signal is acquired through resistive voltage division or capacitive coupling. In an optional embodiment, three-phase current and voltage signals can be acquired simultaneously at different phases, suitable for comprehensive monitoring under three-phase unbalanced load scenarios.

[0023] In S2, based on the current signal, during the process of identifying the dominant load type through a predefined load feature library, a machine learning classification model or rule matching algorithm can be combined to identify the dominant load category in the current circuit online. The predefined load feature library stores standard feature templates for resistive loads (such as electric heaters), inductive loads (such as motors), capacitive loads (such as compensation capacitors), and switching power supply loads (such as LED drivers and chargers). The identification process can use Dynamic Time Warping (DTW) or Support Vector Machine (SVM) to achieve rapid classification. For example, configuring appropriate filtering methods for the filtering unit corresponding to the dominant load involves using low-pass filtering to suppress high-frequency noise for resistive loads, while considering phase compensation for inductive loads.

[0024] In S3, during the filtering operation, the time window length used for moving average or windowed Fourier transform is dynamically adjusted based on the real-time calculated signal-to-noise ratio (SNR). Alternatively, an exponentially weighted moving average method can be used to estimate the SNR trend, improving the robustness of the adjustment.

[0025] In S4, mathematical wavelet functions are used to decompose the signal at different scales, achieving a joint representation in the time and frequency dimensions. Specifically, the Discrete Wavelet Transform (DWT) is used to decompose the signal layer by layer into approximate components (low frequency) and detail components (high frequency), with each layer corresponding to a different frequency bandwidth and time-domain granularity. The two-dimensional matrix image formed by arranging the wavelet coefficients of each layer along the time axis intuitively reflects the distribution of signal energy in the time-frequency plane. This image can accurately capture the local energy abrupt changes and high-frequency oscillations caused by the occurrence of an electric arc, overcoming the limitation of traditional Fourier transform in handling non-stationary signals. In an alternative embodiment, other orthogonal wavelet bases (such as Symlets or Coiflets) can be selected to adapt to specific waveform characteristics.

[0026] In S4, among the typical characteristic parameters, the energy distribution characteristic refers to the changing trend of the signal energy ratio calculated in each wavelet sub-band. When an electric arc occurs, it is often manifested as a sudden increase in energy in the high-frequency sub-band. The spectral mutation point is located by detecting the position where the difference between wavelet coefficients at adjacent times exceeds a set threshold, reflecting the discontinuity of the signal frequency domain structure. The singular value decomposition index is to perform singular value decomposition (SVD) on the wavelet coefficient matrix, extract the largest few singular values ​​and their proportions, and use them to quantify the signal complexity and structural changes.

[0027] In S5, deep learning of typical feature parameters refers to inputting the extracted multidimensional feature sequences into models such as one-dimensional convolutional neural networks (1D-CNN) or long short-term memory networks (LSTM) for end-to-end feature abstraction and pattern recognition. "Performing anomaly scoring" involves the model outputting a value between 0 and 1, representing the degree to which the current signal segment deviates from the normal state; a higher score indicates a higher likelihood of a fault. "Comparing the anomaly score with a preset threshold" generates a preliminary judgment: if the score exceeds the threshold (e.g., 0.7), it is marked as a suspected arcing event; otherwise, it is considered normal operation. It should be understood that the threshold can be dynamically adjusted based on historical data to avoid false alarms or missed alarms caused by a fixed threshold.

[0028] In S5, a trained support vector machine (SVM) classifier is used to fuse the received primary anomaly score vectors and manually constructed auxiliary features for decision-making. SVM achieves maximum margin separation between the two classes of samples (normal / faulty) by finding the optimal hyperplane, exhibiting good generalization ability and anti-overfitting properties.

[0029] Through the above steps, this application realizes a complete closed-loop arc fault detection process of "perception-analysis-judgment-early warning". Due to the adoption of an adaptive filtering strategy based on load type, the system can effectively suppress interference and improve preprocessing accuracy under various power conditions. Through a signal-to-noise ratio-driven sliding window dynamic adjustment mechanism, the system enhances sensitivity to transient arcs under low signal-to-noise ratio conditions and improves processing efficiency under high signal-to-noise ratio conditions. Utilizing the multi-resolution analysis capability of wavelet transform, the system accurately captures the time-frequency local abrupt changes caused by arcs. Through the collaborative extraction of typical and auxiliary features, the system expands the identification coverage of concealed arcs. The use of a hybrid judgment structure combining deep learning and support vector machines balances the flexibility of feature learning with the stability of classification decisions, significantly reducing the false alarm rate. Finally, remote monitoring and data feedback are achieved through digital communication, comprehensively improving the intelligence level and practical value of arc fault detection.

[0030] In some embodiments of the present invention, the dominant load type includes resistive load, inductive load, capacitive load, and switching power supply type load; In the process of configuring the corresponding filtering method for the filtering unit corresponding to the dominant load, a hybrid strategy combining notch filter and Kalman filter is adopted for the high-frequency noise of switching power supply type loads.

[0031] Specifically, the dominant load type refers to the load characteristics of the electrical equipment that accounts for the largest proportion of power or has the most significant current characteristics in the user-side power distribution circuit during a certain power consumption period. This classification covers four common basic types: resistive loads (such as electric furnaces, incandescent lamps, etc.), where the voltage and current are in phase and the power factor is close to 1; inductive loads (such as motors, transformers, etc.), where the current lags behind the voltage and has a low power factor; capacitive loads (such as compensation capacitors, LED driver power supplies, etc.), where the current leads the voltage and exhibits negative reactive characteristics; and switching power supply loads (such as mobile phone chargers, computer power supplies, inverter air conditioners, etc.), which exhibit high-frequency chopping and pulse rectification behavior during operation, generating abundant high-order harmonics and periodic spike interference.

[0032] For the aforementioned types of loads, differentiated filtering methods need to be configured during the signal preprocessing stage to adapt to their unique electromagnetic interference patterns. In particular, when the dominant load is identified as a switching power supply, it generates strong carrier frequencies and their harmonics in the range of tens of kHz to several MHz. Conventional low-pass or band-pass filters are insufficient to effectively suppress such narrowband strong interference, easily leading to false abrupt changes or abnormal energy distributions in subsequent wavelet analysis. Therefore, this application introduces a hybrid filtering strategy combining notch filters and Kalman filters to specifically eliminate high-frequency periodic noise introduced by switching power supplies and improve dynamic signal tracking capabilities. First, a notch filter is used to pre-process the input signal, removing periodic interference concentrated in specific high-frequency bands (such as switching power supply carriers and their harmonics) and reducing the degree of signal-to-noise ratio degradation. Then, the denoised signal is input to a Kalman filter, which further smooths and reconstructs the remaining random noise, intermittent spikes, and transient disturbances. The two form a cascaded structure; the former is responsible for selective suppression in the frequency domain, while the latter undertakes dynamic tracking and optimal estimation in the time domain. Their synergistic effect is significantly better than a single filtering method.

[0033] This implementation scheme introduces differentiated filtering mechanisms for different load types and specifically designs a notch filter + Kalman filter composite noise reduction process for switching power supply loads with severe high-frequency interference. This solves the technical problem that traditional filtering methods struggle to balance fidelity and noise reduction when faced with weak arc signals and high-intensity periodic interference. As a result, it improves the quality and stability of the preprocessed signal, ensures the accuracy of time-frequency feature extraction in the subsequent wavelet transform process, and ultimately enhances the sensitivity and anti-false alarm capability of arc fault identification.

[0034] In some embodiments of the present invention, when the signal-to-noise ratio is lower than a first threshold, the window width is automatically narrowed to improve the ability to capture transient signals; when the signal-to-noise ratio is higher than a second threshold, the window width is widened to optimize computational efficiency and smooth random noise.

[0035] Specifically, regarding situations where the signal-to-noise ratio (SNR) is below the first threshold: when the detected SNR of a current or voltage signal is low, it indicates strong background noise, and the weak transient characteristics caused by the electric arc are easily masked. In this case, the system automatically narrows the width of the sliding window, i.e., reduces the number of data points participating in the current analysis. By narrowing the time window, the resolution on the time axis can be improved, enabling the algorithm to more accurately locate sudden transient processes in the signal. For cases where the signal-to-noise ratio (SNR) is higher than the second threshold: when the signal itself is relatively pure and interference is minimal, it indicates that the main noise sources have been effectively suppressed. In this case, expanding the sliding window width becomes the preferred strategy. A wider window means that more sampled data is included in the average processing range, which helps to further smooth residual random noise through time-domain integration effects, while reducing the number of data frames to be processed per unit time, thereby reducing the computational load and improving the overall system efficiency. In this implementation scheme, by establishing a closed-loop feedback mechanism between the SNR and the window width, the system can enhance time sensitivity at low SNR and focus on computational efficiency and noise suppression at high SNR. This solves the technical problem of the mutual constraint between sensitivity and robustness in traditional methods, thereby improving the overall accuracy and practicality of arc fault feature extraction, and is particularly suitable for real-world scenarios with diverse load types and complex interference sources in residential power distribution systems.

[0036] In some embodiments of the present invention, the wavelet transform algorithm uses Daubechies wavelet basis functions to perform discrete wavelet transform, with a decomposition level of 6-8 levels to cover low-frequency and high-frequency fault characteristics and enhance the ability to identify concealed electric arcs.

[0037] Furthermore, the specific type of Daubechies wavelet basis can be selected according to the actual application scenario, such as wavelet basis functions of different orders such as db4, db6, or db8. Higher-order Daubechies wavelets have smoother waveforms and stronger approximation capabilities, suitable for signal analysis under smooth background loads; while lower-order ones have faster response speeds, which is beneficial for capturing high-frequency transient components. In an optional embodiment, the system can dynamically switch the optimal wavelet basis according to the dominant load type—for example, using db6 for resistive loads to balance regularity and computational efficiency, and selecting db4 for switching power supply loads to enhance sensitivity to high-frequency glitches. Discrete Wavelet Transform (DWT) is a numerical calculation method formed by discretizing and sampling continuous wavelet transforms. By performing low-pass and high-pass filtering on the signal and downsampling layer by layer, the signal is decomposed on multiple time-frequency scales. The decomposition consists of 6-8 layers, with each layer corresponding to a frequency division of one octave: the first layer separates the highest frequency band (e.g., 25kHz–50kHz), and with each subsequent layer, the analysis band is halved downwards. A 6-layer decomposition can cover the low-frequency region below approximately 781Hz (assuming a sampling rate of 100kHz), while an 8-layer decomposition can reach approximately 195Hz, which is sufficient to cover the energy concentration frequency band of typical arc faults (usually distributed between several hundred Hz and tens of kHz).

[0038] In the above scheme, the Daubechies wavelet basis function, due to its excellent time-frequency localization performance, provides a mathematical foundation for discrete wavelet transform that adapts to the characteristics of electric arc signals. The rationally designed 6-8 level decomposition structure fully leverages the advantages of this wavelet basis, maximizing feature representation capabilities with limited computational resources. This solves the problem that general wavelet basis functions are insufficient to fully extract key features of electric arcs, thus significantly enhancing the ability to capture time-frequency abrupt changes in weak electric arc signals and improving the accuracy and robustness of subsequent feature extraction and fault diagnosis.

[0039] In some embodiments of the present invention, the auxiliary features include wavelet coefficient entropy and energy entropy. These auxiliary features are often calculated manually. Wavelet coefficient entropy measures the degree of uncertainty or disorder in the distribution of wavelet coefficients at each decomposition scale after discrete wavelet transform. In signal processing, wavelet coefficients reflect the local characteristics of a signal at different time-frequency resolutions. When an arc fault occurs, especially a low-intensity, intermittent, or highly distorted "soft fault," its current / voltage signal exhibits non-stationarity and strong abrupt changes at multiple scales, leading to a dispersed and irregular energy distribution of the wavelet coefficients, thus causing a significant increase in the wavelet coefficient entropy value. This index quantifies the change in signal structural complexity from an information theory perspective, effectively identifying hidden arc events with insignificant energy changes but intensified waveform disorder. Energy entropy, on the other hand, characterizes the uniformity of energy distribution across different frequency bands. Specifically, after completing the multi-resolution wavelet decomposition, the total signal energy is divided according to the detail components and approximate components of each level, and the energy proportion of each frequency band is calculated. Under normal operating conditions, the energy of the load current is mainly concentrated near the fundamental frequency and its harmonics, and the corresponding wavelet energy is concentrated in the lower frequency band, with a relatively concentrated distribution and low energy entropy. However, once an arc discharge occurs, the high-frequency components increase dramatically and exhibit randomness, and the energy diffuses to higher-order wavelet subbands, causing an increase in energy entropy. Therefore, energy entropy can serve as a supplementary criterion reflecting the characteristics of high-frequency oscillations induced by arcs.

[0040] In some embodiments of the present invention, the support vector machine is periodically incrementally trained using an incremental learning sample library, which includes confirmed fault samples and their contextual features.

[0041] Specifically, in the arc fault identification task, the SVM, acting as a backend fine classifier, receives the initial anomaly score vector output from the frontend deep learning module and key auxiliary features extracted manually. It then constructs the optimal classification hyperplane in a high-dimensional feature space, achieving accurate differentiation between normal load fluctuations and real arc events. This model relies on the quality and coverage of the training data; the initial training set is typically built based on historical fault data under typical load conditions. The incremental learning sample library is a dynamically updated data storage structure used to accumulate sample data and associated information of real arc faults confirmed manually or through a cloud platform during system operation. Each record contains not only raw or preprocessed current and voltage signal segments but also corresponding time-frequency domain feature maps, typical feature parameters, auxiliary features, acquisition time, dominant load type, power consumption period, signal-to-noise ratio status, and other contextual features, forming a fault case archive with rich semantic labels. When specific triggering conditions are met, the system automatically calls the incremental learning algorithm to adjust the local parameters of the existing SVM model and performs periodic incremental training.

[0042] This scheme enables online evolution of support vector machines, solving the problem that statically trained models cannot adapt to dynamic changes in the power environment. Therefore, it significantly improves the long-term stability and accuracy of arc fault identification, extends the model life cycle, reduces the frequency of manual intervention, and enhances the system's autonomous operation and maintenance capabilities and practical deployment value.

[0043] In some embodiments of the present invention, the warning signal and fault data are simultaneously analyzed with the anonymized power consumption anomaly data of adjacent users under the same power distribution circuit. When the judgment result of this end is temporally correlated with the anomaly data of the adjacent end, the confidence level of the judgment result is increased; otherwise, a review procedure is initiated.

[0044] Specifically, "anonymized abnormal electricity consumption data of adjacent users under the same power distribution circuit" refers to anomaly event records detected and reported by other user terminals deployed on the same low-voltage power supply line, which have undergone anonymization processing. This process involves matching and evaluating the correlation between the local early warning event and the abnormal data from neighboring terminals in terms of time and event characteristics. The system first constructs a time sliding window (e.g., ±3 seconds), and retrieves all abnormal events reported by neighboring users within this time window, centered on the time of the local event occurrence. Subsequently, based on the similarity of event types (such as arc-related anomalies), spatial topology (such as belonging to the same branch circuit), and electrical propagation characteristics (such as voltage disturbance propagation delay generally less than 100ms), the spatiotemporal correlation between the two is calculated. If at least one neighboring node reports the same or highly correlated anomaly within the same time period, the two are considered to have a significant temporal correlation. As a variant, a weighted correlation model can also be introduced, dynamically adjusting the data weights of each neighboring node based on factors such as proximity and load capacity ratio. If two adjacent users trigger similar alerts within the same second, the system automatically raises the confidence level of this judgment to "extremely high" and prioritizes pushing it to the monitoring platform or triggering a rapid disconnection command. If no neighboring abnormal data is detected to be temporally correlated with the local event, the system determines that the current alarm may be caused by isolated disturbances or local interference, thereby activating the review process. Through this implementation scheme, this application realizes an arc fault cross-verification mechanism based on multi-point spatiotemporal collaboration.

[0045] See Figure 2 The second aspect of the present invention discloses a wavelet analysis-based system for extracting features from arc fault signals, including... The signal acquisition module is used to acquire current and voltage signals. This module is connected to the metering box of the user-side power distribution line. The signal preprocessing module is used to identify the type of dominant load based on the current signal through a predefined load feature library, configure the corresponding filtering method for the filtering unit corresponding to the dominant load, and filter the current signal and voltage signal through the corresponding filtering method to obtain the preprocessed signal. During the filtering process, the width of the sliding window is adjusted according to the signal-to-noise ratio of the current signal and voltage signal. This module performs filtering and noise reduction processing on the acquired raw signal to eliminate load fluctuations and high-frequency interference. The wavelet analysis module is used to perform multi-resolution analysis on the preprocessed signal using wavelet transform algorithm to obtain time-frequency domain feature maps. The feature extraction module is used to extract typical feature parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points and singular value decomposition index, and to calculate auxiliary features simultaneously. The fault judgment module is used to learn typical feature parameters through deep learning and perform anomaly scoring. The anomaly score is compared with a preset threshold to obtain a preliminary identification result. The preliminary identification result and auxiliary features are processed by a support vector machine to determine whether an arc fault exists and generate an early warning signal. This module performs pattern recognition on the feature parameters through a preset threshold and a machine learning model. The data communication module is used to output the warning signals and fault data to the cloud platform or local monitoring terminal to realize remote monitoring and management.

[0046] In some embodiments of the present invention, the signal preprocessing module adopts a multimodal adaptive filtering architecture, specifically including: Load type identification unit: used to analyze current waveform and power factor in real time, and dynamically identify the type of dominant load in the current line based on a predefined load feature library. The type of dominant load includes at least resistive load, inductive load, capacitive load and switching power supply type load. Filter strategy selector: connected to the load type identification unit, it stores a set of filtering parameters corresponding to different load types; the filter strategy selector automatically configures the corresponding filtering algorithm combination and parameters for the subsequent filtering unit according to the identified load type, wherein for high-frequency noise of switching power supply type load, a hybrid strategy combining notch filter and Kalman filter is enabled. Sliding window dynamic calibration unit: The sliding window dynamic calibration unit adjusts the width of the sliding window in real time according to the signal-to-noise ratio of the signal. When the signal-to-noise ratio is detected to be lower than the first threshold, the window width is automatically narrowed to improve the ability to capture transient signals. When the signal-to-noise ratio is higher than the second threshold, the window width is widened to optimize computational efficiency and smooth random noise.

[0047] In some embodiments of the present invention, the wavelet analysis module uses Daubechies wavelet basis functions to perform discrete wavelet transform, with a decomposition level of 6-8 levels to cover low-frequency and high-frequency fault characteristics.

[0048] In some embodiments of the present invention, the feature extraction module further calculates wavelet coefficient entropy and energy entropy as auxiliary feature parameters to enhance the ability to identify concealed electric arcs.

[0049] In some embodiments of the present invention, the fault judgment module integrates a hybrid intelligent judgment engine, and its construction and operation methods include: Model Architecture: The hybrid intelligent judgment engine adopts a cascaded hybrid model. The front end deploys a primary screener based on a one-dimensional convolutional neural network, which is used to perform deep feature learning and anomaly scoring on the original feature sequence from the feature extraction module. The back end is connected to an optimized support vector machine fine classifier, which receives the output vector of the primary screener and key auxiliary features (such as the entropy feature) to complete the final fault classification decision. Online learning mechanism: The hybrid intelligent judgment engine includes an online model update loop. When the system issues multiple warnings and the warnings are confirmed by the cloud platform or maintenance personnel, the confirmed fault samples and their contextual features will be automatically stored in the incremental learning sample library. During idle periods, the system will periodically use the incremental learning sample library to incrementally train the SVM fine classifier to continuously adapt to the arc characteristics generated by new load devices. Multimodal decision fusion unit: While the hybrid model outputs the results, the multimodal decision fusion unit analyzes in parallel the anonymized power consumption anomaly data of adjacent users in the same power distribution circuit from the data communication module. When the judgment result of this end is temporally correlated with the anomaly data of the adjacent end, the confidence level of this judgment result is increased; otherwise, the review procedure is initiated.

[0050] Preferably, the fault judgment module integrates a support vector machine classifier, and trains the model using historical fault data to achieve classification and identification of series arcs and parallel arcs.

[0051] Preferably, the fault judgment module further includes a dynamic threshold adjustment unit, which automatically adjusts the judgment threshold according to the load type and power consumption period.

[0052] Preferably, the data communication module supports 4G / 5G, LoRa or Ethernet communication protocols, has edge computing capabilities, and can complete some feature extraction and judgment logic locally.

[0053] Preferably, it also includes a self-diagnostic module to periodically calibrate the sensors and verify the effectiveness of the algorithm, ensuring long-term operational stability.

[0054] The third aspect of this invention discloses a computer device, which includes a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment can be used to implement a wavelet analysis-based method for extracting features of arc fault signals, including the following steps: S1, acquiring current signals and... S2, Based on the current signal, identify the type of dominant load through a predefined load feature library, configure the corresponding filtering method for the filtering unit corresponding to the dominant load, and filter the current signal and voltage signal through the corresponding filtering method to obtain a preprocessed signal; during the filtering process, adjust the width of the sliding window according to the signal-to-noise ratio of the current signal and voltage signal; S3, perform multi-resolution analysis on the preprocessed signal through wavelet transform algorithm to obtain a time-frequency domain feature map; S4, extract typical feature parameters of arc fault from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points and singular value decomposition index, and calculate auxiliary features simultaneously; S5, perform deep learning on the typical feature parameters and perform anomaly scoring, compare the anomaly score with a preset threshold to obtain a preliminary identification result; process the preliminary identification result and auxiliary features through support vector machine to determine whether an arc fault exists and generate a warning signal; S6, output the warning signal and fault data to the cloud platform or local monitoring terminal.

[0055] A fourth aspect of this invention discloses a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement a wavelet analysis-based arc fault signal feature extraction method in the above embodiments, including the following steps: S1, acquiring current and voltage signals; S2, based on the current signal, identifying the type of dominant load through a predefined load feature library, configuring corresponding filtering methods for the filtering units corresponding to the dominant load, and filtering the current and voltage signals through the corresponding filtering methods to obtain a preprocessed signal; during the filtering process, adjusting the width of the sliding window according to the signal-to-noise ratio of the current and voltage signals; S3, through... The wavelet transform algorithm performs multi-resolution analysis on the preprocessed signal to obtain a time-frequency domain feature map; S4, typical feature parameters of the arc fault are extracted from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points, and singular value decomposition index, and auxiliary features are calculated simultaneously; S5, deep learning is performed on the typical feature parameters, and anomaly scoring is performed. The anomaly score is compared with a preset threshold to obtain a preliminary identification result; the preliminary identification result and auxiliary features are processed by a support vector machine to determine whether an arc fault exists and generate an early warning signal; S6, the early warning signal and fault data are output to the cloud platform or local monitoring terminal.

[0056] The application will be further described below with reference to a specific embodiment.

[0057] Depend on Figure 3 The present invention discloses an arc fault signal feature extraction system based on wavelet analysis, including a signal acquisition module, a signal preprocessing module, a wavelet analysis module, a feature extraction module, a fault judgment module, a data communication module, and a self-diagnosis module.

[0058] The signal acquisition module is connected to the metering box of the user-side power distribution line and is used to acquire current and voltage signals in real time. The signal preprocessing module filters and reduces noise in the acquired raw signal to eliminate load fluctuations and high-frequency interference. The wavelet analysis module uses the wavelet transform algorithm to perform multi-resolution analysis on the preprocessed signal and generate a time-frequency domain feature map. The feature extraction module extracts typical feature parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral abrupt change points, and singular value decomposition indexes. The fault judgment module performs pattern recognition on feature parameters based on preset thresholds and machine learning models to determine whether an arc fault exists and generates an early warning signal. The data communication module uploads early warning signals and fault data to the cloud platform or local monitoring terminal to achieve remote monitoring and management.

[0059] During implementation, this system first uses high-precision sensors integrated into the metering box to capture raw current and voltage signals in the user-side power distribution lines in real time. These signals are then sent to a preprocessing module, where filtering and noise reduction techniques eliminate interference from daily load fluctuations and high-frequency noise, providing a clean signal source for subsequent analysis. The core wavelet analysis module performs multi-resolution wavelet transforms on the processed signals, converting the one-dimensional time-domain signal into a feature spectrum containing rich time-frequency information, thereby accurately capturing the weak transient characteristics unique to series arc faults, with voltage fluctuations of only 5%-8%. The feature extraction module then quantifies key parameters such as energy distribution and spectral abrupt changes from this spectrum. The fault judgment module compares these feature values ​​with preset thresholds and machine learning models to achieve intelligent identification and classification of fault modes. Finally, the data communication module uploads the early warning information to the cloud platform or local monitoring center, forming a complete "perception-analysis-judgment-early warning" chain, enabling early and accurate detection of hidden arc faults. The signal acquisition module includes a high-precision current sensor and a voltage sensor, with a sampling frequency of no less than 100kHz. It can capture microsecond-level arc signal changes, ensuring that it can capture the high-frequency current mutation signal generated at the moment the arc occurs, which has a very short duration (microsecond level), providing a high-fidelity data foundation for subsequent analysis.

[0060] The signal preprocessing module adopts a multimodal adaptive filtering architecture, specifically including: Load type identification unit: used to analyze current waveform and power factor in real time, and dynamically identify the type of dominant load in the current line based on a predefined load feature library. The type of dominant load includes at least resistive load, inductive load, capacitive load and switching power supply type load. Filter strategy selector: connected to the load type identification unit, it stores a set of filtering parameters corresponding to different load types; the filter strategy selector automatically configures the corresponding filtering algorithm combination and parameters for the subsequent filtering unit according to the identified load type, wherein for high-frequency noise of switching power supply type load, a hybrid strategy combining notch filter and Kalman filter is enabled. Sliding window dynamic calibration unit: The sliding window dynamic calibration unit adjusts the width of the sliding window in real time according to the signal-to-noise ratio of the signal. When the signal-to-noise ratio is detected to be lower than the first threshold, the window width is automatically narrowed to improve the ability to capture transient signals. When the signal-to-noise ratio is higher than the second threshold, the window width is widened to optimize computational efficiency and smooth random noise.

[0061] When the above solution was implemented, it achieved accurate perception and targeted processing of the complex power environment on the user side. It can intelligently identify the types of loads connected to the line and dynamically select the optimal filtering strategy based on the noise characteristics of different loads, while adaptively adjusting the analysis window according to signal quality. This design solves the drawback of traditional fixed-parameter filters' "one-size-fits-all" approach when dealing with diverse household appliances, significantly improving the intelligence and adaptability of the signal preprocessing stage. This ensures that subsequent wavelet analysis can obtain the input signal with the optimal signal-to-noise ratio, laying a solid foundation for accurately identifying weak arc characteristics.

[0062] For example, in its implementation: when a vacuum cleaner with a variable frequency motor (inductive load) is connected to the system, the load type identification unit can correctly classify it by analyzing its current waveform and power factor characteristics. The filter strategy selector then configures the Kalman filter with parameters suitable for the inductive load to smooth the periodic interference generated by the motor commutation. At the same time, a large transient current is generated when the vacuum cleaner starts up, at which point the signal-to-noise ratio drops sharply. The sliding window dynamic calibration unit immediately narrows the window width to ensure that the detailed features of this transient process are fully captured and not filtered out as noise.

[0063] The wavelet analysis module uses the Daubechies wavelet basis function for discrete wavelet transform, with a decomposition level of 6-8 levels to cover low-frequency and high-frequency fault features. The Daubechies wavelet basis is used and decomposed into 6-8 levels because the wavelet basis has good localization characteristics and can effectively extract the edge features of transient signals.

[0064] The feature extraction module also calculates wavelet coefficient entropy and energy entropy as auxiliary feature parameters to enhance the ability to identify concealed electric arcs. Introducing these two feature parameters, wavelet coefficient entropy and energy entropy, as auxiliary features allows for the measurement of the disorder of signal distribution from an information theory perspective. When an electric arc fault occurs, the signal complexity increases significantly, and the entropy value changes accordingly, thereby enhancing the sensitivity for identifying concealed electric arcs with insignificant energy changes but waveform distortion.

[0065] The fault diagnosis module integrates a hybrid intelligent diagnosis engine, and its construction and operation methods include: Model Architecture: The hybrid intelligent judgment engine adopts a cascaded hybrid model. The front end deploys a primary screener based on a one-dimensional convolutional neural network, which is used to perform deep feature learning and anomaly scoring on the original feature sequence from the feature extraction module. The back end is connected to an optimized support vector machine fine classifier, which receives the output vector of the primary screener and key auxiliary features (such as the entropy feature) to complete the final fault classification decision. Online learning mechanism: The hybrid intelligent judgment engine includes an online model update loop. When the system issues multiple warnings and the warnings are confirmed by the cloud platform or maintenance personnel, the confirmed fault samples and their contextual features will be automatically stored in the incremental learning sample library. During idle periods, the system will periodically use the incremental learning sample library to incrementally train the SVM fine classifier to continuously adapt to the arc characteristics generated by new load devices. Multimodal decision fusion unit: While the hybrid model outputs the results, the multimodal decision fusion unit analyzes in parallel the anonymized power consumption anomaly data of adjacent users in the same power distribution circuit from the data communication module. When the judgment result of this end is temporally correlated with the anomaly data of the adjacent end, the confidence level of this judgment result is increased; otherwise, the review procedure is initiated.

[0066] It integrates the feature abstraction capabilities of deep learning with the decision-making advantages of traditional machine learning, and introduces continuous learning and collaborative verification mechanisms to overcome the problems of insufficient generalization ability of single models, difficulty in adapting to new equipment, and occasional false alarms. Furthermore, during implementation, it constructs a fault diagnosis system that can continuously evolve over time and possesses "collective intelligence," significantly improving fault identification accuracy and system robustness, thus giving the entire arc detection system long-term applicability and higher reliability.

[0067] When the above scheme is running, the primary screener (1D-CNN) first automatically learns from the input feature sequence and outputs a vector containing deep features. This vector, along with manually calculated features such as energy entropy, is fed into the SVM fine classifier to determine that the current situation is "suspected series arc". At the same time, the multimodal decision fusion unit obtains information through the LoRa network that adjacent users in the same meter box do not have similar anomalies at the same time, so the confidence of this judgment is appropriately reduced, and instead of immediately triggering an alarm, it is marked as "to be observed". Subsequently, the event is confirmed by maintenance personnel on-site to be a real arc caused by poor contact of an old socket. This sample is stored in the incremental learning sample library for subsequent model updates, enabling the system to more accurately identify such features in the future.

[0068] The fault judgment module also includes a dynamic threshold adjustment unit, which automatically adjusts the judgment threshold according to the load type and power consumption period to reduce the false alarm rate, thereby solving the problem that fixed thresholds may cause false alarms during high load periods and missed alarms during low load periods.

[0069] The data communication module supports 4G / 5G, LoRa, or Ethernet communication protocols, has edge computing capabilities, and can complete some feature extraction and judgment logic locally, ensuring the system's deployment flexibility under different community infrastructure conditions.

[0070] It also includes a self-diagnostic module for periodically calibrating sensors and verifying algorithm effectiveness, ensuring long-term operational stability. Through regular self-checks, it monitors whether sensor accuracy is drifting and whether the algorithm module is functioning correctly, thereby ensuring the system maintains the highest monitoring reliability and accuracy during long-term unattended operation, reducing maintenance costs.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting arc fault signal features based on wavelet analysis, characterized in that, The method comprises the following steps: acquiring current signals and voltage signals; identifying the type of a dominant load through a predefined load characteristic library based on the current signals, configuring a corresponding filtering method for a filtering unit corresponding to the dominant load, filtering the current signals and the voltage signals through the corresponding filtering method, and obtaining preprocessed signals; in the filtering process, adjusting the width of a sliding window according to the signal-to-noise ratio of the current signals and the voltage signals; performing multi-resolution analysis on the preprocessed signals through a wavelet transform algorithm to obtain a time-frequency domain feature map; extracting typical feature parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral mutation points, and singular value decomposition indexes, and synchronously calculating auxiliary features; performing deep learning on the typical feature parameters, performing abnormal scoring, comparing the abnormal score with a preset threshold, obtaining a preliminary identification result, processing the preliminary identification result and the auxiliary features through a support vector machine, judging whether an arc fault exists, and generating a warning signal; outputting the warning signal and fault data to a cloud platform or a local monitoring terminal.

2. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, The dominant load type includes resistive load, inductive load, capacitive load, and switching power supply type load. In the process of configuring the corresponding filtering method for the filtering unit corresponding to the dominant load, a hybrid strategy combining a notch filter and a Kalman filter is adopted for high-frequency noise of the switching power supply type load.

3. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, When the signal-to-noise ratio is lower than a first threshold, the window width is automatically narrowed to improve the ability to capture transient signals; when the signal-to-noise ratio is higher than a second threshold, the window width is widened to optimize the calculation efficiency and smooth random noise.

4. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, The wavelet transform algorithm adopts a Daubechies wavelet basis function to perform discrete wavelet transform, and the decomposition layer number is 6-8 layers.

5. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, The auxiliary features include wavelet coefficient entropy and energy entropy.

6. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, The support vector machine performs periodic incremental training through an incremental learning sample library, and the incremental learning sample library includes confirmed fault samples and context features thereof.

7. The arc fault signal feature extraction method based on wavelet analysis according to claim 1, characterized in that, The warning signal and the fault data are analyzed with anonymized power consumption abnormal data of adjacent users under the same power distribution circuit; when the local judgment result is associated with the adjacent abnormal data in time, the confidence level of the current judgment result is improved; otherwise, a review program is started.

8. An arc fault signal feature extraction apparatus based on wavelet analysis, characterized by, The method comprises the following steps: a signal acquisition module for acquiring current signals and voltage signals; a signal preprocessing module for identifying the type of a dominant load through a predefined load characteristic library based on the current signals, configuring a corresponding filtering method for a filtering unit corresponding to the dominant load, filtering the current signals and the voltage signals through the corresponding filtering method, and obtaining preprocessed signals; in the filtering process, adjusting the width of a sliding window according to the signal-to-noise ratio of the current signals and the voltage signals; a wavelet analysis module for performing multi-resolution analysis on the preprocessed signals through a wavelet transform algorithm to obtain a time-frequency domain feature map; a feature extraction module for extracting typical feature parameters of arc faults from the time-frequency domain feature map, including energy distribution characteristics, spectral mutation points, and singular value decomposition indexes, and synchronously calculating auxiliary features; The fault judgment module is used for deep learning of typical characteristic parameters, abnormal scoring, comparison of the abnormal score with a preset threshold, and obtaining of a preliminary identification result; and the support vector machine is used for processing of the preliminary identification result and auxiliary features, judgment of whether there is an arc fault, and generation of a warning signal; The data communication module is used for outputting the warning signal and fault data to a cloud platform or a local monitoring terminal.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the wavelet analysis-based arc fault signal feature extraction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the wavelet analysis-based arc fault signal feature extraction method according to any one of claims 1 to 7.

Images