Alternating current arc fault weak signal detection method, device, equipment and medium

Abstract

The invention discloses an AC arc fault weak signal detection method, device and equipment and a medium, and the method comprises the steps: obtaining a bus current signal of a low-voltage power distribution system, carrying out the amplitude standardization preprocessing of the bus current signal, and generating a normalized signal; performing time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector; performing fault diagnosis on the comprehensive feature vector by adopting a nested sliding window mechanism to generate a fault diagnosis mark; and based on the comprehensive feature vector and the fault diagnosis mark, utilizing a pre-trained classification model to generate an arc fault classification result. According to the method, through the synergistic effect of signal noise reduction, feature fusion, real-time diagnosis and intelligent classification, the detection sensitivity is improved to reliable identification in a low signal-to-noise ratio-10dB scene, the real-time performance meets the UL1699 standard requirement, and the core problem that weak alternating current arc fault detection sensitivity and real-time performance are insufficient is solved.

Description

Technical Field

[0001] This invention relates to the technical field of electrical safety monitoring, and in particular to a method, device, equipment, and medium for detecting weak signals of AC arc faults. Background Technology

[0002] With the development of safety monitoring technology for low-voltage power distribution systems, arc fault detection technology has gradually become a core research direction in the field of electrical fire prevention. Traditional technologies mainly rely on hardware protection devices such as miniature circuit breakers (MCBs) and residual current devices (RCDs). These devices are based on current amplitude thresholds or ground leakage current-triggered tripping mechanisms, and are characterized by fast response speed and low deployment cost. However, in weak AC arc fault scenarios (such as current distortion rate less than 15% and duration less than 100ms), traditional hardware protection devices suffer from serious detection omissions because they cannot capture nonlinear fault characteristics. For example, the current change amplitude of a series arc fault is only 10%-30% of the load current, far below the operating threshold of the MCB; when the system is effectively grounded, the detection sensitivity of the RCD decreases significantly in parallel arc faults. Current software-based arc fault detection methods attempt to compensate for hardware limitations through signal analysis algorithms. Mainstream solutions include: time-domain threshold detection, which judges faults by calculating the rate of change of current or zero crossover distortion, but the false alarm rate is as high as 40% under disturbances such as motor start-up and switching operations; and frequency-domain harmonic analysis, which relies on specific harmonic components of arc faults (such as the 3rd / 5th / 7th harmonics), but the high-frequency energy of weak arcs accounts for less than 5% and is easily drowned out by background noise. The core technical bottleneck is that the above solutions have real-time defects. Existing algorithms have a processing delay of >200ms, which cannot meet the half-cycle (such as 10ms@50Hz) diagnostic timeliness required by the UL1699 standard. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method, device, equipment and medium for detecting weak signals of AC arc faults, which can improve the diagnostic timeliness.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A method for detecting weak signals of AC arc faults, comprising the following steps:

[0006] Obtain the bus current signal of the low-voltage power distribution system, perform amplitude standardization preprocessing on the bus current signal, and generate a normalized signal;

[0007] The normalized signal is subjected to time-frequency domain feature fusion extraction to generate a comprehensive feature vector;

[0008] A nested sliding window mechanism is used to perform fault diagnosis on the comprehensive feature vector, generating fault diagnosis markers;

[0009] Based on the comprehensive feature vector and the fault diagnosis label, an arc fault classification result is generated using a pre-trained classification model.

[0010] To solve the above-mentioned technical problems, the present invention adopts other technical solutions as follows:

[0011] A device for detecting weak signals of AC arc faults, comprising:

[0012] The signal acquisition and preprocessing module is used to acquire the bus current signal of the low-voltage power distribution system and perform amplitude standardization preprocessing to generate a normalized signal.

[0013] The time-frequency domain feature fusion and extraction module is used to perform time-frequency domain feature fusion and extraction on the normalized signal to generate a comprehensive feature vector;

[0014] The fault diagnosis marker generation module is used to perform fault diagnosis on the comprehensive feature vector using a nested sliding window mechanism and generate fault diagnosis markers.

[0015] The classification result generation module is used to generate arc fault classification results based on the comprehensive feature vector and the fault diagnosis label, using a pre-trained classification model. The classification results include normal state, series arc fault and parallel arc fault.

[0016] To solve the above-mentioned technical problems, the present invention adopts other technical solutions as follows:

[0017] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for detecting weak signals of AC arc faults.

[0018] To solve the above-mentioned technical problems, the present invention adopts other technical solutions as follows:

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for detecting weak signals of AC arc faults.

[0020] The beneficial effects of this invention include at least the following: providing a method, apparatus, device, and medium for detecting weak AC arc fault signals; acquiring bus current signals of low-voltage power distribution systems; eliminating the influence of environmental noise floor and signal amplitude fluctuations through amplitude normalization preprocessing; generating a normalized signal; improving the signal-to-noise ratio (e.g., from -10dB to a detectable level); and avoiding the difficulty in capturing weak arc fault signals due to noise overload. The normalized signal undergoes time-frequency domain feature fusion extraction, specifically combining progressive singular value decomposition (PSVD) recursively to extract aperiodic components and fast Fourier transform (FFT) to extract periodic features, generating a comprehensive feature vector. By fusing multi-dimensional time-frequency domain features, the distinguishability of fault features is enhanced, avoiding the confusion problem of traditional single-domain features under load harmonic interference (such as the 40% false alarm rate caused by motor starting or switching operations), and improving detection sensitivity. A nested sliding window mechanism is used for fault diagnosis based on the comprehensive feature vector. The long-time window covers a signal segment of fixed length (e.g., 200ms), and multiple short-time windows slide continuously within its range (e.g., a 40ms window extracts features in real time with a 20ms step). A pre-trained classifier (e.g., SVM) is used for preliminary judgment through the small windows. If any small window triggers a fault, the state determination of the large window is activated, and a fault diagnosis label is generated. This can compress the diagnosis delay to half a cycle (10ms@50Hz) and address the real-time defects caused by excessively long delays. Based on the comprehensive feature vector and the fault diagnosis label, a pre-trained classification model is used to dynamically output classification results including normal state, series and parallel arc faults. Through feature recombination triggered by fault state (e.g., frequency domain component amplitude enhancement) and a two-level classification collaborative mechanism, the classification ambiguity problem of weak arc faults (current distortion rate <10%) is overcome. Through the synergistic effect of signal denoising, feature fusion, real-time diagnosis and intelligent classification, the detection sensitivity is improved to reliable identification in low signal-to-noise ratio scenarios of -10dB, and the real-time performance meets the requirements of UL1699 standard, solving the core problem of insufficient sensitivity and real-time performance in weak AC arc fault detection. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for detecting weak signals of AC arc faults according to an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of a weak signal detection device for AC arc faults according to an embodiment of the present invention;

[0023] Label Explanation:

[0024] 101. Signal acquisition and preprocessing module; 102. Time-frequency domain feature fusion and extraction module; 103. Fault diagnosis marker generation module; 104. Classification result generation module. Detailed Implementation

[0025] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0026] Before detailing the embodiments of this application, some related concepts will first be explained:

[0027] In existing technologies, time-domain threshold detection judges faults by calculating the rate of change of current or zero crossover distortion, but the false alarm rate is as high as 40% under disturbances such as motor starting and switching operations; frequency-domain harmonic analysis relies on specific harmonic components of arc faults (such as the 3rd / 5th / 7th harmonics), but the high-frequency energy of weak arcs accounts for less than 5% and is easily drowned out by background noise; single-window machine learning uses classifiers such as SVM to directly process the original signal features, but it is sensitive to the sampling window position and the false negative rate exceeds 25%. Its core technical bottlenecks are: signal weakness, when the fault current distortion rate is <10%, the signal-to-noise ratio (SNR) is as low as -10dB, and traditional feature extraction methods fail; environmental interference coupling, the harmonics generated by nonlinear loads (such as frequency converters) overlap with the arc spectrum, resulting in feature confusion; real-time defects, the processing delay of existing algorithms is >200ms, which cannot meet the half-cycle (such as 10ms@50Hz) diagnostic timeliness required by the UL1699 standard.

[0028] To at least solve the above problems, please refer to Figure 1 This invention provides a method for detecting weak signals of AC arc faults, comprising the following steps:

[0029] S01. Obtain the bus current signal of the low-voltage power distribution system, perform amplitude standardization preprocessing on the bus current signal, and generate a normalized signal.

[0030] The process includes: bus current signal (current data on the low-voltage distribution bus monitored in real time by acquisition devices such as Hall current sensors, used to capture system operating status); amplitude normalization preprocessing (a signal processing technique aimed at eliminating the environmental noise floor and unifying the signal scale to improve the signal-to-noise ratio of weak arc faults); calculation and comparison of the original bus current signal with the bus current reference signal under any historical normal operating conditions to obtain a normalized signal with filtered steady-state noise; and through dynamic noise reduction and normalization, solving the key problem that weak arc fault signals are difficult to distinguish when the signal-to-noise ratio is as low as -10dB, ensuring that the entire detection process can maintain high sensitivity and robustness in noisy environments.

[0031] S02. Perform time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector;

[0032] A nested sliding window mechanism is used to perform fault diagnosis on the comprehensive feature vector, generating fault diagnosis markers;

[0033] The time-frequency domain feature fusion extraction (the core process of constructing a multi-dimensional fault characterization system by collaboratively analyzing the non-periodic features of the signal in the time dimension and the periodic features in the frequency dimension) can be implemented through two paths. The main path can employ a collaborative mechanism of progressive singular value decomposition (PSVD) and fast Fourier transform (FFT). Singular value decomposition is recursively performed on the normalized signal, and the entropy exponent is calculated through a preset sliding window (size 5). When the entropy exponent is lower than the threshold of 0.5, the corresponding singular values ​​are selected as non-periodic component features. Simultaneously, fast Fourier transform is performed in parallel on the same normalized signal to extract the amplitude ratio of the fundamental wave and the 3rd to 7th harmonics, as well as the energy proportion of the high-frequency band (>5kHz), as periodic features. Finally, the non-periodic singular value sequence and the periodic spectral parameters are concatenated into a comprehensive feature vector. The alternative path integrates wavelet analysis and singular value decomposition. It performs 8-level wavelet decomposition on the normalized signal to obtain the component coefficients of each frequency band, constructs the Hankel matrix, and performs SVD decomposition to obtain the singular value sequence. The component filtering is realized by entropy exponent calculation, and the filtered singular value sequence is reconstructed into a time domain signal as a comprehensive feature vector.

[0034] S03. A nested sliding window mechanism is used to perform fault diagnosis on the comprehensive feature vector to generate fault diagnosis markers;

[0035] Among them, the nested sliding window mechanism (is a two-layer window diagnostic architecture, consisting of a long-time window (e.g., 200ms corresponding to 10 power cycles @ 50Hz) covering a signal segment of a first fixed time length and at least one short-time window (each covering a second fixed time length, e.g., 40ms corresponding to 2 power cycles @ 50Hz) that slides continuously within the time range of the large window)

[0036] When implementing fault diagnosis marking (a Boolean flag representing the final diagnostic state of a long-term window (0 for normal, 1 for fault)), the sizes of the large and small windows are configured (the large window size is larger than the small window size). On the time-series data of the comprehensive feature vectors, the small window typically slides with a step size half the window size. For each sliding position of the small window, feature sub-vectors within its range are extracted and input into a pre-trained first support vector machine classifier (SVM model) for preliminary binary classification (outputting a normal or faulty judgment). If any small window is judged to be in a faulty state, the fault state judgment of the corresponding large window is triggered; otherwise, when all small windows are normal, the large window is marked as normal, and a fault diagnosis mark for that large window is generated. Through dense sampling and real-time judgment of the small windows (processing delay compressed to within half a cycle), the problem of false negatives caused by the difficulty in capturing the initial signals of weak arc faults is solved. Simultaneously, the comprehensive judgment of the large window avoids false alarms caused by load fluctuations, meeting the dual constraints of the UL1699 standard on the real-time performance and reliability of fault diagnosis.

[0037] S04. Based on the comprehensive feature vector and the fault diagnosis label, generate arc fault classification results using a pre-trained classification model; wherein, the classification results include normal state, series arc fault and parallel arc fault.

[0038] The pre-trained classification model (using a two-level classifier architecture trained on historical datasets (the first level is a Support Vector Machine (SVM) classifier, and the second level is a parallel fault-specific classifier)) generates a state feature vector by concatenating the fault diagnosis label (Boolean value) with the comprehensive feature vector when the arc fault classification result (final fault type label determined by a confidence threshold) is implemented. When the diagnosis label represents the fault state, the amplitude enhancement operation is performed on the frequency domain feature component generated by the Fast Fourier Transform in the comprehensive feature vector, replacing the original feature component to generate a reconstructed state feature vector. The final fault type label determined by the confidence threshold is input into the first-level SVM classifier, which outputs the series arc fault probability value. If it is lower than the preset threshold of 0.7, the amplitude-enhanced frequency domain component is input into the second-level parallel fault-specific classifier to output the parallel arc fault probability value. When the diagnosis label is in a normal state, the fault judgment result of normal state is directly output. When the label is in a fault state, the series arc fault probability value is compared with the parallel arc fault probability value, and the graded alarm information is generated by combining the confidence level. By combining feature recombination triggered by fault state with a two-level classification collaboration mechanism, the problem of classification ambiguity caused by feature overlap in series / parallel weak arc faults at a signal-to-noise ratio of -10dB has been overcome.

[0039] As can be seen from the above description, the beneficial effects of this invention are as follows: It acquires the bus current signal of a low-voltage power distribution system and eliminates the influence of environmental noise floor and signal amplitude fluctuations through amplitude normalization preprocessing, generating a normalized signal and improving the signal-to-noise ratio (e.g., from -10dB to a detectable level), solving the problem of weak arc fault signals being difficult to capture due to noise overload; it performs time-frequency domain feature fusion extraction on the normalized signal, combining progressive singular value decomposition (PSVD) to recursively extract aperiodic components and fast Fourier transform (FFT) to extract periodic features, generating a comprehensive feature vector. By fusing multi-dimensional time-frequency domain features, it enhances the distinguishability of fault features, avoiding the confusion problem of traditional single-domain features under load harmonic interference (such as the 40% false alarm rate caused by motor starting or switching operation), and improving detection sensitivity; it adopts a nested sliding window mechanism for comprehensive... Fault diagnosis is performed using feature vectors. A long-time window covers a signal segment of fixed length (e.g., 200ms), and multiple short-time windows slide continuously within this range (e.g., a 40ms window extracts features in real time with a 20ms step). A pre-trained classifier (e.g., SVM) is used for preliminary judgment through the small windows. If any small window triggers a fault, the state determination of the large window is activated, generating a fault diagnosis label. This can compress the diagnosis delay to half a cycle (10ms@50Hz), addressing the real-time defects caused by excessively long delays. Based on the comprehensive feature vectors and fault diagnosis labels, a pre-trained classification model dynamically outputs classification results including normal states, series and parallel arc faults. Through feature recombination triggered by fault states (e.g., frequency domain component amplitude enhancement) and a two-level classification collaborative mechanism, the classification ambiguity problem of weak arc faults (current distortion rate <10%) is overcome. Through the synergistic effect of signal denoising, feature fusion, real-time diagnosis and intelligent classification, the detection sensitivity is improved to reliable identification in low signal-to-noise ratio scenarios of -10dB, and the real-time performance meets the requirements of UL1699 standard, solving the core problem of insufficient sensitivity and real-time performance in weak AC arc fault detection.

[0040] In some implementations, the amplitude normalization preprocessing of the bus current signal to generate a normalized signal specifically includes:

[0041] S11. Calculate the mean μ and standard deviation σ of the bus current signal X, and use the following formula to calculate the initial normalized signal.

[0042]

[0043] S12, Select any normal operating condition bus current signal X from the historical database of the same low-voltage power distribution system. ref Calculate the mean μ ref and standard deviation σ ref The reference normalized signal is calculated using the following formula.

[0044]

[0045] S13, based on the initial normalized signal and reference normalized signal The following formula is used to perform dynamic noise basis cancellation to generate a normalized signal extracted from time-frequency domain features.

[0046]

[0047] in, This is the normalized vector extracted by time-frequency domain feature fusion.

[0048] Specifically, the bus current signal is the low-voltage power distribution system bus current timing data X = {x1, x2, ..., x...} acquired in real time by Hall sensors. N}, where N is the number of sampling points, and its mean μ and standard deviation σ represent the DC offset and fluctuation intensity of the signal, respectively; the initial normalized signal is obtained through the formula Calculations are performed to eliminate the interference of load current amplitude differences on feature extraction. Reference normalized signal. The generation requires selecting any normal operating current signal X from the historical database of the same system. ref Calculate its mean μ ref and standard deviation σ ref Substitute formula It also performs dynamic noise basis cancellation operation to normalize the initial signal. Normalized signal with reference Subtract elements by formula Generate normalized vectors extracted from time-frequency domain features fusion By using differential operations to remove steady-state noise (such as motor harmonics and switching transients), the signal-to-noise ratio is improved by about 15dB. Before processing, the WASVD coefficients of the normal signal and the arc fault signal highly overlap in the low-frequency band. After dynamic noise elimination by this step, the fault characteristics show significant separation in the high-frequency non-periodic components, which can effectively suppress the noise floor.

[0049] In some implementations, the step of performing time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector includes a first implementation:

[0050] S21. Perform asymptotic singular value decomposition on the normalized signal to obtain the first singular value sequence; wherein, the time complexity of asymptotic singular value decomposition is O(n);

[0051] S22. Perform a Fast Fourier Transform on the first singular value sequence to obtain periodic feature components; wherein, the time complexity of the Fast Fourier Transform is O(nlog n);

[0052] S23. Calculate the entropy index based on the first singular value sequence, select the singular values ​​that meet the first preset entropy threshold, and use them as the non-periodic feature components of the first singular value sequence.

[0053] The aperiodic feature components and the periodic feature components are fused to generate the comprehensive feature vector. The total time complexity of the time-frequency domain feature fusion extraction is O(nlog n).

[0054] As described above, Progressive Singular Value Decomposition (PSVD) is an algorithm that recursively performs Singular Value Decomposition (SVD). It constructs a Hankel matrix by segmenting the normalized signal into segments of a preset window size (e.g., 200 points), recursively folds the residual components layer by layer, and repeats the SVD operation to generate the first singular value sequence S = {s1, s2, ..., s...}. i}, where the singular value s i The projection intensity of signal energy onto an orthogonal basis represents the amplitude distribution of unsteady-state fault components. Asymptotic decomposition compresses the time complexity to O(n) (where n is the time length) through recursive dimensionality reduction, compared to the O(n) time complexity of traditional SVD. 3 Significantly improves real-time performance. A Fast Fourier Transform (FFT) is performed on the first singular value sequence. For the same computational spectrum F(ω), the amplitude ratio of the fundamental wave (50Hz) and the 3rd-7th harmonics, along with the energy proportion of the high-frequency band (>5kHz), are extracted as periodic feature vectors to capture the harmonic distortion patterns caused by faults. The FFT operation has a time complexity of O(nlog n), and parallel execution with PSVD avoids timing delays. Singular values ​​that meet the first preset entropy threshold are selected using the entropy exponent calculation, which can be achieved through the following formula: Where, p i =s i / ∑s j Let s be the energy percentage of the i-th singular value. i Let j be the i-th singular value and j be the summation index variable. Based on a first preset entropy threshold (statistically obtained from the UL1699 standard test dataset: the average entropy value of the singular value sequence under normal operating conditions is 0.82±0.15, and the entropy value drops sharply to 0.32±0.12 under arc faults; taking 0.5 as the threshold can separate 99.2% of fault events), key singular values ​​are screened: when the window entropy value drops sharply below the threshold, the singular values ​​before that position are retained as non-periodic components. Low-frequency steady-state components related to environmental interference are eliminated through adaptive filtering, and aperiodic components and periodic features are fused together to generate a comprehensive feature vector V through vector concatenation. fused =[S aperiodic V periodicThis solution achieves feature fusion through parallel processing of Progressive Singular Value Decomposition (PSVD) and Fast Fourier Transform (FFT), resolving false alarms due to load fluctuations. The processing delay is ≤5ms (PSVD and FFT are calculated in parallel), meeting the timeliness requirements of UL1699 standard for half-cycle (10ms@50Hz) diagnosis. It is suitable for resource-constrained environments such as civil power distribution and embedded terminals.

[0055] In some embodiments, the step of performing time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector further includes a second embodiment:

[0056] S21. Perform wavelet decomposition on the normalized signal to obtain the component coefficients of multiple sub-frequency bands;

[0057] S22. Perform singular value decomposition on the component coefficients of each sub-band to generate a second singular value sequence;

[0058] S23. Calculate the entropy exponent of the second singular value sequence within the preset sliding window, and filter out the second singular values ​​that are lower than the second preset entropy threshold.

[0059] S24. Perform differential filtering operation based on the physical properties of the sub-frequency band of the second singular value that is lower than the second preset entropy threshold:

[0060] S241. For sub-frequency bands located within the preset power frequency range, filter them out after performing phase correction processing on the corresponding component coefficients;

[0061] S242. For sub-bands located in the preset near-zero frequency range, their singular value sequences are directly filtered out to obtain the second singular value sequence after differential filtering.

[0062] S25. The second singular value sequence after differential filtering is reconstructed into a time-domain signal and used as the comprehensive feature vector.

[0063] As can be seen from the above description, wavelet decomposition can be used to normalize the signal. Perform Discrete Wavelet Transform (DWT), and decompose into 8 levels using the Daubechies-4 wavelet basis to obtain the coefficients C = {c1, c2, ..., c8} for each frequency band component, where the coefficients c i Encapsulates the signal energy distribution of a specific frequency band (such as the frequency band corresponding to the i-th layer) to separate high-frequency transient fault components from low-frequency steady-state noise; a Hankel matrix H can be constructed based on the component coefficients. k (Matrix dimensions are dynamically set based on component lengths, with the number of rows being...) The number of columns is Where n is the length of the coefficient sequence, and it is decomposed into H using singular value decomposition (SVD). k =U∑V TExtract the second singular value sequence ∑={σ1,σ2,…,σ r (r is the matrix rank, and the singular values ​​are arranged in descending order to represent the energy intensity of the principal components of the signal). The sparse representation of fault features is enhanced through matrix reconstruction. The entropy exponent of the second singular value sequence can be calculated using the following formula, based on a preset sliding window size (e.g., 200 points): Where, p i =s i / ∑s j Let s be the energy percentage of the i-th singular value. i (where i is the i-th singular value and j is the summation index variable). By comparing and obtaining the second singular value below the second preset entropy threshold (based on historical dataset statistical optimization and fault physical characteristic verification, for example, according to the frequency band coupling characteristics, for the power frequency sub-band (e.g., 50Hz): the entropy value can be 0.3-0.6 (strong harmonic interference); for the near-zero frequency sub-band (<5Hz): the entropy value is 0.1-0.4 (DC drift); for the fault sensitive band (>5kHz): the entropy value is >0.7 (weak noise)), differential filtering is performed according to the physical properties of the corresponding sub-band: power frequency range sub-band processing, for sub-bands located in the preset power frequency range (e.g., 50Hz±5Hz), phase correction processing of the corresponding component coefficients is performed (phase distortion is eliminated through linear phase offset compensation), and then the singular value sequence is filtered out to solve the false alarm problem caused by the overlap of power frequency harmonics and arc characteristics; near-zero frequency range sub-band processing, for sub-bands located in the preset near-zero frequency range (e.g., <5Hz), their singular value sequences are directly filtered out. To address DC drift noise and prevent low-frequency interference from drowning out weak fault signals, a differential filtering method is employed, resulting in a second singular value sequence ∑filtered after component filtering. The innovation lies in customizing the filtering strategy based on frequency band physical properties, improving the signal-to-noise ratio by approximately 15dB compared to traditional uniform filtering methods, with a processing delay of ≤15ms, making it suitable for high-noise and complex scenarios. The differentially filtered second singular value sequence ∑filtered is then reconstructed into a Hankel matrix using inverse SVD, and further mapped to a time-domain signal via inverse wavelet transform, serving as the output of the comprehensive feature vector. This reconstruction process preserves the high-frequency non-stationary characteristics of arc faults while filtering out environmental steady-state components, solving the problem of power frequency harmonics drowning out high-frequency fault signals. It can be deployed on cloud servers to process high-noise scenario data and is suitable for power distribution in industrial and mining enterprises and high-noise, complex load scenarios.

[0064] In some implementations, the step of using a nested sliding window mechanism to perform fault diagnosis on the comprehensive feature vector and generate fault diagnosis markers includes:

[0065] S31. Set up a nested sliding window mechanism, and use a long-time window in the nested sliding window mechanism to cover a signal segment of a first fixed time length; use at least one short-time window in the nested sliding window mechanism to slide continuously within the time range of the long-time window; each of the short-time windows covers a second fixed time length;

[0066] S32. Real-time feature extraction is performed on the comprehensive feature vector through each sliding short-time window;

[0067] S33. Input the features extracted from each short-term window into the pre-trained first classifier to perform preliminary fault judgment.

[0068] S34. If any short-term window is determined to be in a fault state, then the corresponding long-term window is triggered to determine the fault state.

[0069] S35. Generate a fault diagnosis marker based on the state determination.

[0070] As described above, a long-term window can be defined as a signal segment covering a first fixed time length (e.g., 200ms corresponding to 10 power frequency cycles @ 50Hz), and a short-term window is at least one sub-window that slides continuously within the time range of the long-term window (each covering a second fixed time length, e.g., 40ms corresponding to 2 power frequency cycles). These two constitute a nested sliding window mechanism. Real-time feature extraction is performed on the comprehensive feature vector through each sliding short-term window (sliding step size is half the window size), extracting feature sub-vectors (e.g., time-domain statistics or frequency-domain energy) within its range. The features extracted from each short-term window are input into a pre-trained first classifier (e.g., a support vector machine (SVM) model) for preliminary fault judgment, outputting a binary classification result (0 indicating normal, 1 indicating fault). If any short-term window is judged as a fault state, the corresponding long-term window fault state judgment is triggered; otherwise, when all small windows are normal, the large window is marked as normal. Based on this state judgment, a Boolean fault diagnosis label F is generated. flag (0 indicates normal, 1 indicates fault), which can compress the diagnostic delay to half a cycle (e.g., 10ms@50Hz), solving the real-time defect of excessively long processing delays in algorithms. Millisecond-level response to initial fault signals is achieved through dense sampling with a small window (avoiding false negatives), while comprehensive judgment with a large window suppresses load fluctuation interference (e.g., the 40% false alarm rate caused by motor starting), ensuring that the generation of fault diagnosis markers still meets the UL1699 standard aging requirements even at a signal-to-noise ratio as low as -10dB.

[0071] In some implementations, generating arc fault classification results based on the integrated feature vector and the fault diagnosis marker using a pre-trained classification model includes:

[0072] S41. Concatenate the fault diagnosis marker with the comprehensive feature vector to generate a state feature vector;

[0073] S42. When the fault diagnosis mark represents the fault state, the frequency domain feature component generated by the fast Fourier transform in the state feature vector is subjected to amplitude enhancement operation to generate amplitude-enhanced frequency domain component.

[0074] S413. Replace the frequency domain feature component in the state feature vector with the amplitude-enhanced frequency domain component to generate a recombined state feature vector;

[0075] S44. Input the recombined state feature vector into the second classifier and output the series arc fault probability value.

[0076] S45. When the probability value of the series arc fault is lower than the preset threshold, the amplitude enhancement frequency domain component is input into the third classifier and the parallel arc fault probability value is output.

[0077] S46. Perform a classification operation based on the status of the fault diagnosis marker:

[0078] S461. If the label represents the normal state, then output the arc fault classification result for the normal state.

[0079] S462. If the label represents the fault state, then the arc fault classification result of series or parallel fault is generated based on the series arc fault probability value and the parallel arc fault probability value.

[0080] As can be seen from the above description, the fault diagnosis mark F can be used. flag With the comprehensive feature vector V fused =[S aperiodic V periodic [Concatenate to generate state feature vector V] state =[F flag V fused ], Introducing prior knowledge of fault states to enhance classification guidance. When F flag When V = 1, an amplitude enhancement operation is performed on V. periodic Amplitude-enhanced frequency domain components are generated with a gain factor k = 1.5. By amplifying the fault harmonic characteristics, the signal-to-noise ratio is improved, and V state V in periodic Replace with Forming the recombined state feature vector V recombined V recombined The first classifier outputs the probability value P of a series arc fault. series If P series If the value is less than 0.7 (the preset threshold has been optimized experimentally), then... The second classifier (an SVM model specifically designed for parallel faults) is input separately, and the output is the probability value P of the parallel arc fault. parallel Based on F flag Perform a classification operation on the state, if F flag =0, directly output the classification result of the normal state; if F flag =1, compare P series and P parallel When P series Output series arc fault when ≥0.7, otherwise when P parallel When the value is >0.6, a parallel arc fault is output. Redundant calculations under normal conditions are avoided by recombining conditional features triggered by fault markers, improving real-time performance (processing delay compressed to 10ms); amplitude enhancement is used to specifically amplify fault-sensitive frequency components, solving the problem of weak parallel arcs being difficult to identify in grounding systems; the hierarchical diagnosis strategy for series / parallel faults reduces the feature space dimension, significantly optimizing the misjudgment rate compared to traditional single-model approaches.

[0081] In some implementations, the generated arc fault classification result is followed by:

[0082] S51. Based on the type of the arc fault classification result, calculate the confidence level and generate graded alarm information;

[0083] S52. The time-domain waveform of the normalized signal, the time-frequency feature spectrum corresponding to the comprehensive feature vector, the time-series change of the fault diagnosis mark and the evolution curve of the classification confidence are dynamically visualized and fused to generate a fault visualization interface.

[0084] As described above, the graded alarm information (priority alarms generated based on the arc fault classification result type (normal / series fault / parallel fault), where the priority can be divided into first-level alarms and second-level alarms) has a confidence level C = max(P) series P parallel The system calculates probability values ​​from two-level classifiers. When C > 0.9, a Level 1 alarm (audio-visual emergency alarm) is triggered; when 0.7 ≤ C ≤ 0.9, a Level 2 alarm (system log recording and early warning) is triggered; when C < 0.7, only diagnostic logs are recorded. This mechanism addresses the high false alarm rate of the background technology. Dynamic visualization fusion processing is performed, combining four data sets: normalized signal... The time-domain waveform (showing the signal purity after noise floor elimination); the time-frequency feature spectrum corresponding to the comprehensive feature vector (such as the comparison of aperiodic components before and after PSVD entropy filtering and the FFT spectrum energy distribution); the time-series changes of fault diagnosis markers: F flag (The fault triggering process of nested windows is shown using a 0 / 1 Boolean sequence); Classification confidence evolution curve (showing P) series With P parallelThe dynamic game theory is used to generate a fault visualization interface through spatiotemporal alignment algorithm, which renders the time-domain waveform change point, the time-frequency spectrum feature separation area, the fault mark trigger time and the confidence inflection point, so as to realize transparent diagnosis of fault tracing.

[0085] The aforementioned method for detecting weak AC arc fault signals based on time-frequency fusion technology addresses the issues of missed detection and false alarms in traditional hardware protection devices under weak arc fault conditions (signal-to-noise ratio as low as -10dB, current distortion rate <10%) through a four-order linkage architecture. The technical steps include: acquiring the bus current signal of the low-voltage power distribution system; generating a normalized signal through dynamic noise floor elimination and amplitude standardization; and performing differential calculations by calculating the mean / standard deviation of the current signal and historical normal signals. Removing steady-state noise improves the signal-to-noise ratio by approximately 15dB; time-frequency domain feature fusion is performed on the normalized signal: two real-time methods can be adopted depending on the actual situation: Method 1: Parallel processing of Progressive Singular Value Decomposition (PSVD) + Fast Fourier Transform (FFT), suitable for low-noise environments such as civil power distribution, with a time complexity of O(n log n). n)(PSVD and FFT parallel optimization) improves real-time performance (delay ≤ 5ms); Method 2 adopts wavelet decomposition + singular value decomposition (SVD) cascade processing, which is suitable for high-noise complex environments such as industrial and mining enterprises, with a delay ≤ 15ms (wavelet and SVD cascade processing), thereby enhancing anti-interference performance (signal-to-noise ratio improved by 15dB); a nested sliding window mechanism is adopted for real-time fault diagnosis: a sliding short window is set within a long window, and the fault status is judged in real time by the SVM classifier of the small window. When any small window triggers a fault, the fault mark of the large window is activated, compressing the diagnosis delay period; based on the comprehensive feature vector and fault diagnosis mark, dynamic classification is performed through a pre-trained classification model: under the fault status, the amplitude of the FFT frequency domain component is enhanced, and after reorganizing the feature vector, the SVM classifier outputs the probability value of series / parallel arc fault, which is combined with the confidence threshold to generate a graded alarm and visualization interface. By overcoming the coupling problem of environmental interference through dynamic noise cancellation and feature fusion, the fault features are significantly separated in the high-frequency non-periodic components; the nested window mechanism reduces the false alarm rate from 40% to near zero, and the real-time performance meets the UL1699 standard; a two-level classification strategy is adopted to solve the problem of accurate identification and classification ambiguity of weak arcs under a signal-to-noise ratio of -10dB.

[0086] Please refer to Figure 2 In another embodiment of the present invention, a device for detecting weak AC arc fault signals based on time-frequency fusion technology is provided, comprising:

[0087] The signal acquisition and preprocessing module 101 is used to acquire the bus current signal of the low-voltage power distribution system and perform amplitude standardization preprocessing to generate a normalized signal.

[0088] The time-frequency domain feature fusion extraction module 102 is used to perform time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector.

[0089] The fault diagnosis marker generation module 103 is used to perform fault diagnosis on the comprehensive feature vector using a nested sliding window mechanism and generate fault diagnosis markers.

[0090] The classification result generation module 104 is used to generate arc fault classification results based on the comprehensive feature vector and the fault diagnosis label, using a pre-trained classification model. The classification results include normal state, series arc fault and parallel arc fault.

[0091] Based on the same inventive concept, this application also provides an AC arc fault weak signal detection device for implementing the AC arc fault weak signal detection method described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more AC arc fault weak signal detection device embodiments based on time-frequency fusion technology provided below can be found in the limitations of the AC arc fault weak signal detection method based on time-frequency fusion technology described above, and will not be repeated here.

[0092] In one embodiment, the device further includes a visual alarm module for:

[0093] Based on the type of arc fault classification results, the confidence level is calculated and graded alarm information is generated.

[0094] The time-domain waveform of the normalized signal, the time-frequency feature spectrum corresponding to the comprehensive feature vector, the time-series changes of the fault diagnosis mark, and the evolution curve of the classification confidence are dynamically visualized and fused to generate a fault visualization interface.

[0095] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method for detecting weak signals of AC arc faults.

[0096] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for detecting weak signals of AC arc faults.

[0097] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0098] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for detecting weak signals of AC arc faults, characterized in that, Including the following steps: Obtain the bus current signal of the low-voltage power distribution system, perform amplitude standardization preprocessing on the bus current signal, and generate a normalized signal; The normalized signal is subjected to time-frequency domain feature fusion extraction to generate a comprehensive feature vector; A nested sliding window mechanism is used to perform fault diagnosis on the comprehensive feature vector, generating fault diagnosis markers; Based on the comprehensive feature vector and the fault diagnosis label, an arc fault classification result is generated using a pre-trained classification model.

2. The method for detecting weak signals of AC arc faults according to claim 1, characterized in that: The amplitude normalization preprocessing of the bus current signal to generate a normalized signal specifically includes: Obtain the mean and standard deviation of the bus current signal, and calculate the initial normalized signal based on the mean and standard deviation; The reference normalized signal is calculated based on the bus current signal under normal operating conditions selected from the historical database of the same low-voltage power distribution system. Based on the initial normalized signal and the reference normalized signal, a dynamic noise basis elimination operation is performed to generate a normalized signal for the time-frequency domain feature fusion extraction.

3. The method for detecting weak signals of AC arc faults according to claim 2, characterized in that: The step of performing time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector includes: Perform asymptotic singular value decomposition on the normalized signal to obtain the first singular value sequence; Perform a Fast Fourier Transform on the first singular value sequence to obtain periodic feature components; The entropy index is calculated based on the first singular value sequence, and singular values ​​that meet the first preset entropy threshold are selected and used as the non-periodic feature components of the first singular value sequence. The aperiodic feature component and the periodic feature component are fused to generate the comprehensive feature vector.

4. The method for detecting weak signals of AC arc faults according to claim 2, characterized in that: The step of performing time-frequency domain feature fusion extraction on the normalized signal to generate a comprehensive feature vector includes: The normalized signal is decomposed by wavelet decomposition to obtain the component coefficients of multiple sub-bands; Singular value decomposition is performed on the component coefficients of each sub-band to generate a second singular value sequence; Calculate the entropy exponent of the second singular value sequence within a preset sliding window, and filter out the second singular values ​​that are lower than the second preset entropy threshold; Perform differential filtering based on the physical properties of the sub-band of the second singular value that is below the second preset entropy threshold: For sub-bands located within a preset power frequency range, the corresponding component coefficients are phase corrected and then filtered out; for sub-bands located within a preset near-zero frequency range, their singular value sequences are directly filtered out to obtain a second singular value sequence after differential filtering. The second singular value sequence after differential filtering is reconstructed into a time-domain signal and used as the comprehensive feature vector.

5. The method for detecting weak signals of AC arc faults according to claim 1, characterized in that, The step of using a nested sliding window mechanism to perform fault diagnosis on the comprehensive feature vector and generate fault diagnosis markers includes: A nested sliding window mechanism is set up, and a long-time window in the nested sliding window mechanism is used to cover a signal segment of a first fixed time length; at least one short-time window in the nested sliding window mechanism slides continuously within the time range of the long-time window; each of the short-time windows covers a second fixed time length. Real-time feature extraction is performed on the comprehensive feature vector through each sliding short-time window; The features extracted from each short-term window are input into the pre-trained first classifier for preliminary fault judgment. If any short-term window is determined to be in a fault state, then the corresponding long-term window is triggered to determine the fault state. Fault diagnosis markers are generated based on the state determination.

6. The method for detecting weak signals of AC arc faults according to claim 5, characterized in that, The process of generating arc fault classification results based on the comprehensive feature vector and the fault diagnosis marker using a pre-trained classification model includes: The fault diagnosis marker is concatenated with the comprehensive feature vector to generate a state feature vector; When the fault diagnosis mark characterizes the fault state, an amplitude enhancement operation is performed on the frequency domain feature component generated by the fast Fourier transform in the state feature vector to generate an amplitude-enhanced frequency domain component. The frequency domain feature components in the state feature vector are replaced with the amplitude-enhanced frequency domain components to generate a reconstructed state feature vector. The recombined state feature vector is input into the second classifier, which outputs the probability value of series arc fault. When the probability value of the series arc fault is lower than the preset threshold, the amplitude enhancement frequency domain component is input into the third classifier, and the parallel arc fault probability value is output. Perform a classification operation based on the status of the fault diagnosis markers: If the label represents a normal state, the arc fault classification result for the normal state is output; if the label represents a fault state, the arc fault classification result for the series or parallel fault is generated based on the series arc fault probability value and the parallel arc fault probability value.

7. The method for detecting weak signals of AC arc faults according to claim 6, characterized in that, Following the generated arc fault classification results, the following are also included: Based on the type of the arc fault classification result, the confidence level is calculated and graded alarm information is generated; The time-domain waveform of the normalized signal, the time-frequency feature spectrum corresponding to the comprehensive feature vector, the time-series changes of the fault diagnosis mark, and the evolution curve of the classification confidence are dynamically visualized and fused to generate a fault visualization interface.

8. A device for detecting weak signals of AC arc faults, characterized in that, include: The signal acquisition and preprocessing module is used to acquire the bus current signal of the low-voltage power distribution system and perform amplitude standardization preprocessing to generate a normalized signal. The time-frequency domain feature fusion and extraction module is used to perform time-frequency domain feature fusion and extraction on the normalized signal to generate a comprehensive feature vector; The fault diagnosis marker generation module is used to perform fault diagnosis on the comprehensive feature vector using a nested sliding window mechanism and generate fault diagnosis markers. The classification result generation module is used to generate arc fault classification results based on the comprehensive feature vector and the fault diagnosis label, using a pre-trained classification model. The classification results include normal state, series arc fault and parallel arc fault.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for detecting weak signals of AC arc faults according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for detecting weak signals of AC arc faults according to any one of claims 1 to 7.

Images