Cross-scenario arc fault diagnosis method, system, device and storage medium

Abstract

The application discloses a cross-scene alternating current arc fault diagnosis method, system, device and storage medium. The method comprises the following steps: acquiring alternating current bus current signals and bus voltage signals under arc fault and non-fault states of multiple classic scenes to obtain original scene signals; performing wavelet packet decomposition and singular value decomposition on the original scene signals to obtain a periodic characteristic vector set; performing periodic arc structure feature screening on the periodic characteristic vector set to obtain a periodic characteristic subset; performing scene migration on the periodic characteristic subset by using a generative adversarial network to obtain a migration characteristic set; the migration characteristic set comprises real scene samples and generated samples; and training a classifier by using the periodic characteristic subset and the migration characteristic set to obtain a cross-scene alternating current arc fault diagnosis model and perform fault diagnosis. By using the method, the periodicity can be extracted, the scene can be migrated, and the high-robustness diagnosis capability can be achieved, so that the arc fault recognition demand in a multi-source environment can be met.

Description

Technical Field

[0001] This invention relates to the technical field of power system fault detection, and in particular to a method, system and storage medium for cross-scenario AC arc fault diagnosis. Background Technology

[0002] AC arc faults, a common but dangerous form of electrical fault, are characterized by high frequency, periodic instability, and difficulty in extraction, and are widely found in household power distribution, power lines, and industrial settings. This study employs signal analysis methods to extract and classify arc fault features, such as frequency domain methods based on wavelet transform, wavelet packet decomposition, and Fourier transform, combined with traditional classification algorithms like Support Vector Machine (SVM) and k-Nearest Neighbors (KNN) for identification.

[0003] Existing technologies enhance the expressive power of time-frequency features by combining matrix factorization techniques such as Singular Value Decomposition (SVD) with wavelet analysis, while also introducing deep learning models for end-to-end identification. However, these methods typically rely on large amounts of field sampling data, and the training models are highly dependent on the target equipment. Arc fault identification models trained under single equipment or fixed load conditions experience a sharp drop in recognition rate when faced with changes in power supply scenarios, exhibiting weak transferability and poor generalization ability across different equipment, loads, and voltage levels, making it difficult to meet practical deployment requirements. Existing deep learning enhancement methods do not fully consider the periodicity, energy distribution, and other physical structural characteristics of arc faults, and direct transfer can easily lead to feature distortion. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, system, device and storage medium for cross-scenario AC arc fault diagnosis, which can improve the accuracy of cross-scenario AC arc fault diagnosis.

[0005] In one aspect of the present invention, a cross-scenario AC arc fault diagnosis method is provided. The method includes: acquiring AC bus current signals and bus voltage signals under arc fault and non-fault states in multiple preset classic scenarios to obtain original scenario signals, wherein the preset classic scenarios cover different load types, power supply environments, and equipment models; performing wavelet packet decomposition and singular value decomposition on the original scenario signals to obtain a set of periodic feature vectors; filtering the set of periodic feature vectors to satisfy periodic arc structure features to obtain a subset of periodic features; using a generative adversarial network to perform scenario transfer based on the subset of periodic features to obtain a transfer feature set, wherein the transfer feature set includes real scenario samples and generated samples; training a classifier using the subset of periodic features and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model; and using the cross-scenario AC arc fault diagnosis model to perform arc fault diagnosis.

[0006] In another aspect of the present invention, a cross-scenario AC arc fault diagnosis system is provided. The system includes: a data acquisition module for acquiring AC bus current and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain original scenario signals, wherein the preset classic scenarios cover different load types, power supply environments, and equipment models; a feature extraction module for performing wavelet packet decomposition and singular value decomposition on the original scenario signals to obtain a periodic feature vector set; a periodic feature filtering module for filtering the periodic feature vector set to satisfy periodic arc structure features to obtain a periodic feature subset; a training set construction module for using a generative adversarial network to perform scenario transfer based on the periodic feature subset to obtain a transfer feature set, wherein the transfer feature set includes real scenario samples and generated samples; and a model training module for training a classifier using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, and using the cross-scenario AC arc fault diagnosis model for arc fault diagnosis.

[0007] In another aspect of the invention, a computer device is provided. The computer device includes a memory configured to store a computer program; and a processor configured to execute the computer program to perform the aforementioned cross-scenario AC arc fault diagnosis method.

[0008] In another aspect of the invention, a computer-readable medium is provided. This medium stores a computer program that is executed by a processor to implement the aforementioned cross-scenario AC arc fault diagnosis method.

[0009] The beneficial effects of this invention are as follows: By constructing a complete cross-scenario fault modeling process from the bottom up, covering the entire process from multi-scenario data acquisition to diagnostic model deployment, a virtuous cycle is formed between each step through data-driven processes, structure mapping, and model feedback. Fine-grained frequency domain processing of non-stationary signals is achieved through wavelet packet decomposition and singular value decomposition, supplemented by a periodic behavior filtering mechanism to ensure that input features are both expressive and physically interpretable. Compared to existing models trained only in a single scenario, the introduction of a structure-preserving adversarial generation mechanism effectively bridges the distribution gap between the source and target domains, exhibiting stronger robustness to scenario transfer. Dynamically weighted training combining transfer samples and partially labeled samples from the target domain, without relying on large-scale labeled data, improves the model's feasibility and ensures that generated samples do not lose accuracy in key arc features such as frequency band energy and periodic perturbations, thereby enhancing the diagnostic usability of generated samples and improving the accuracy of cross-scenario AC arc fault diagnosis. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the cross-scenario AC arc fault diagnosis method of the present invention;

[0011] Figure 2 This is a schematic diagram of the process for generating a periodic feature vector set according to the present invention;

[0012] Figure 3 This is a schematic diagram of the process for generating periodic feature subsets according to the present invention;

[0013] Figure 4 This is a structural diagram of the cross-scenario AC arc fault diagnosis system of the present invention. Detailed Implementation

[0014] 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.

[0015] Before detailing the embodiments of this application, some related concepts will first be explained:

[0016] (1) Singular value decomposition: an important matrix decomposition method in linear algebra.

[0017] (2) Wavelet Packet Decomposition (WPD): Projecting the signal onto the space spanned by wavelet packet basis functions. From the perspective of signal processing, it is to make the signal pass through a series of filters with different center frequencies but the same bandwidth.

[0018] (3) KL divergence (Kullback-Leibler Divergence): also known as relative entropy, it measures the degree of information difference between two probability distributions. The smaller the value, the more similar the distributions are, and the larger the value, the more significant the difference is.

[0019] (4) Generative Adversarial Network (GAN): This model consists of a generator and a discriminator. The generator is responsible for capturing the distribution of the sample data, while the discriminator is typically a binary classifier that distinguishes between real data and generated samples. The optimization process of this model is a binary minimax game problem. During training, one side (discriminator or generator) is fixed, and the parameters of the other side are updated. The process is repeated iteratively until the generator can estimate the distribution of the sample data.

[0020] In existing technologies, when dealing with AC arc faults, matrix factorization techniques such as singular value decomposition (SVD) are typically combined with wavelet analysis to enhance the expressive power of time-frequency features, while deep learning models are introduced for end-to-end identification. However, such methods usually rely on a large amount of field sampling data, and the training model is highly dependent on the target device. Arc fault identification models trained under single device or fixed load conditions experience a sharp drop in recognition rate when faced with changes in power supply scenarios, exhibiting weak transferability. This results in poor generalization ability across devices, loads, and voltage levels, making it difficult to meet practical deployment requirements. Furthermore, existing deep learning enhancement methods do not fully consider the periodicity, energy distribution, and other physical structural characteristics of arc faults, and direct transfer can easily lead to feature distortion.

[0021] To at least solve the above problems, please refer to Figure 1 This invention provides a cross-scenario AC arc fault diagnosis method, the method comprising:

[0022] Acquire AC bus current and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain the original scenario signal. The preset classic scenarios cover different load types, power supply environments and equipment models.

[0023] Wavelet packet decomposition and singular value decomposition are performed on the original scene signal to obtain a set of periodic feature vectors;

[0024] The periodic feature vector set is filtered to satisfy the periodic arc structure characteristics to obtain a periodic feature subset;

[0025] A scene transfer is performed using a generative adversarial network based on the periodic feature subset to obtain a transfer feature set, which includes real scene samples and generated samples.

[0026] The classifier is trained using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, which is then used for arc fault diagnosis.

[0027] As described above, by collecting bus current and voltage data under multiple preset classic scenarios and constructing a unified periodic feature vector set, the problem of data distribution offset caused by changes in power grid conditions is effectively alleviated, providing structured input for subsequent unified modeling. By introducing a periodic structure identification and filtering mechanism into the periodic feature vector set, only highly reliable features with time-frequency consistency and frequency band energy concentration trends are retained, effectively eliminating non-periodic disturbances and improving the anti-interference capability of the diagnostic model. By constructing a generative adversarial network combining adversarial loss and structure preservation loss, adaptive mapping of source scenario samples to target scenario styles is achieved, ensuring the physical authenticity of generated samples in terms of structural features, thereby expanding the applicability of the model. By constructing a hybrid training set combining real target samples and transferred generated samples, a fault identification model with controllable performance and clear classification boundaries can still be efficiently trained even when labeled data is lacking in the target scenario. The final cross-scenario diagnostic model has input adaptation and real-time identification capabilities, and can be embedded and deployed in different power supply equipment to meet the multi-scenario needs of industrial field fault monitoring and improve the accuracy of cross-scenario AC arc fault diagnosis.

[0028] Please refer to Figure 2 Furthermore, wavelet packet decomposition and singular value decomposition are performed on the original scene signal to obtain a set of periodic feature vectors, including:

[0029] The original scene signal is periodically segmented to obtain a periodic segment sequence, which includes multiple periodic segments.

[0030] Wavelet packet decomposition is performed on the periodic segment to obtain a sub-frequency band coefficient group, which includes multiple sub-frequency band nodes, and each sub-frequency band node includes signal coefficients within a corresponding local frequency range;

[0031] The sub-band coefficient groups corresponding to each periodic segment are assembled into a two-dimensional matrix to obtain the sub-band energy matrix;

[0032] The sub-band energy matrix is ​​subjected to singular value decomposition, and the time-frequency feature matrix is ​​obtained by selecting singular values ​​of a set order.

[0033] A set of periodic feature vectors is constructed based on the time-frequency feature matrices corresponding to all periodic segments within the periodic segment sequence.

[0034] As described above, wavelet packet decomposition is used to decompose signal segments of each period length. As a generalization of wavelet analysis, wavelet packet decomposition can perform complete hierarchical decomposition of the signal in the time-frequency domain, obtaining more detailed high-frequency and low-frequency sub-band coefficients, avoiding the problem of insufficient high-frequency resolution in ordinary wavelets. Furthermore, this multi-band coefficient group is used to construct a two-dimensional sub-band energy matrix, where the horizontal axis represents the band number and the vertical axis represents the sample time window. To further reduce dimensionality and preserve the main structural features, singular value decomposition is performed on this sub-band energy matrix. The resulting singular value vector reflects the energy distribution pattern of the sample along each principal axis. The previous singular values ​​are retained as compressed vectors representing the signal energy structure, i.e., the periodic feature vectors constituting the periodic segment, ultimately yielding the feature vector set corresponding to the complete set of periodic segments.

[0035] Please refer to Figure 3 Furthermore, the periodic feature vector set is filtered to satisfy the periodic arc structure characteristics to obtain a periodic feature subset, including:

[0036] A sliding window is used to construct a multi-period input on the periodic segment sequence to obtain a periodic sequence window;

[0037] Stacking the corresponding sub-band energy matrices within the periodic sequence window yields a three-dimensional tensor.

[0038] The three-dimensional tensor is expanded into a two-dimensional matrix along the periodic dimension, and progressive singular value decomposition is performed to obtain the windowed composite singular value vector.

[0039] The frequency band energy distribution index is calculated for the corresponding sub-band energy matrix within the periodic sequence window to obtain the frequency band energy normalization index.

[0040] The periodic feature subset is obtained by filtering the feature vectors corresponding to the periodic feature vector set based on the window comprehensive singular value vector of each periodic sequence window and the frequency band energy normalization index.

[0041] As described above, selecting a subset with typical periodic arc structures from the set of periodic feature vectors can provide more representative fault samples for subsequent migration and discrimination. Periodic arcs typically exhibit energy concentration and frequency band variation patterns that repeat across multiple cycles, demonstrating strong temporal structural consistency. Specifically, to capture this feature, a multi-cycle sliding window is constructed, aggregating the sub-frequency band energy matrices of multiple adjacent cycle segments within the window, and tensorizing them to preserve the time dimension. Using the Progressive Singular Value Decomposition (PSVD) method, the tensor within the window is expanded into a two-dimensional matrix, and the PSVD operation is performed to extract the principal components of cross-cycle variations, thereby measuring its periodic consistency. Furthermore, based on the average trend of energy evolution of each frequency band with the cycle, a normalized frequency-band power allocation (FPA) index is calculated, which quantifies the stability of arc energy distribution in the frequency domain. Under the conditions of satisfying singular value stability and the FPA change amplitude being less than a set threshold, periodic samples within the window are considered to have periodic arc structure characteristics and are included in the periodic feature subset.

[0042] Further, the frequency band energy distribution index is calculated for the corresponding sub-band energy matrix within the periodic sequence window to obtain the frequency band energy normalization index, including:

[0043] Based on the corresponding sub-band energy matrix within the periodic sequence window, frequency band energy statistics are performed to obtain the average energy distribution vector.

[0044] The frequency band energy normalization index is calculated based on the energy distribution of each frequency band within the period and the average energy distribution vector.

[0045] As described above, frequency band energy statistics are performed on the sub-band energy matrix corresponding to each periodic segment within the periodic sequence window to extract the average energy level of each frequency band over multiple periods. To ensure comparability of energy distributions between different frequency bands and to eliminate the influence of overall energy amplitude variations on feature determination, the average energy distribution vector is normalized to form a frequency band energy normalization index. The normalization index describes the relative distribution ratio of energy in each frequency band, and its shape and distribution pattern can be used to identify the frequency domain behavior patterns unique to arc faults. In particular, when a certain type of arc fault repeatedly accumulates energy in several frequency bands, the proportion of the corresponding frequency band in the normalization index will be significantly higher than that of other frequency bands, thus forming a recognizable structural template. This index can also effectively eliminate interference caused by differences in overall signal strength due to voltage levels, load changes, etc., allowing subsequent feature selection to focus more on the structural essence rather than signal amplitude.

[0046] Further, based on the window-comprehensive singular value vector of each periodic sequence window and the frequency band energy normalization index, the feature vectors corresponding to the periodic feature vector set are filtered to obtain a periodic feature subset, including:

[0047] The stability of the window-integrated singular value vector and the frequency band energy normalization index of each periodic sequence window is analyzed by KL divergence analysis to obtain the stability analysis results of the periodic sequence window. The stability analysis results include stable arc periodic behavior segments and unstable arc periodic behavior segments.

[0048] If the stability analysis result of the periodic sequence window is a stable arc periodic behavior segment, then the feature vector corresponding to the periodic feature vector set is selected as the periodic feature.

[0049] The periodic feature subset is constructed by constructing the periodic features corresponding to the periodic sequence windows of the multiple stable arc periodic behavior segments;

[0050] The KL divergence analysis corresponds to the following steps:

[0051] If the window's integrated singular value vector satisfies the condition of periodic change trend, and the frequency band energy normalization index satisfies the condition of fixed frequency band energy concentration trend, then the corresponding periodic sequence window is determined to be a stable arc periodic behavior segment.

[0052] If the window's integrated singular value vector does not meet the periodic change trend condition, or the frequency band energy normalization index does not meet the fixed frequency band energy concentration trend condition, then the corresponding periodic sequence window is determined to be an unstable arc periodic behavior segment.

[0053] As described above, KL divergence is introduced as the core criterion to analyze the distribution stability of the window's comprehensive singular value vector and the frequency band energy normalization index in the time series of each periodic sequence window. Specifically, the comprehensive singular value vector of each periodic sequence window is normalized and treated as a probability density distribution in a high-dimensional space. A KL divergence sequence between adjacent windows is constructed to capture the changing trend of periodic behavior. Similarly, the frequency band energy normalization index is used as a probability vector describing the frequency domain structure. The changing trend of its KL divergence across periods is used to determine whether energy is concentrated in a fixed frequency band. When the singular value vector of a periodic sequence window exhibits a clear periodic structure, that is, the KL divergence change between adjacent windows shows low amplitude and high repetition, and the corresponding frequency band energy normalization index continuously occupies a dominant proportion in a specific frequency band, it can be considered that there is stable and periodically prominent arc behavior within the window. At this time, the window is determined to be a stable arc periodic behavior segment; conversely, if any index shows violent fluctuations or lacks a concentrated trend, the window is determined to be an unstable arc periodic behavior segment.

[0054] Based on the stability analysis results, feature vector filtering was performed on the periodic sequence windows identified as stable arc periodic behavior segments. This ensured that the selected samples not only possessed energy structure characteristics within a single cycle but also exhibited pattern consistency and frequency band energy commonality across cycle sequences, thus improving the physical semantic reliability of the samples. The periodic features contained in all periodic sequence windows identified as stable arc periodic behavior segments were summarized to construct a periodic feature subset. Compared to the initial periodic feature vector set, this subset eliminated non-periodic perturbations, transient anomalies, and invalid segments, retaining the periodic segment features that remained stable and representative under multidimensional structural analysis.

[0055] Furthermore, a generative adversarial network is used to perform scene transfer based on the periodic feature subset to obtain a transfer feature set, including:

[0056] Based on minimizing the generator loss objective, feature mapping is performed on the periodic feature subset to generate target scene features; the generator loss includes adversarial loss and structure preservation loss; the structure preservation loss is used to constrain the target scene samples to retain the arc feature structure in the periodic feature subset;

[0057] Based on maximizing the discriminator loss target, the probability of authenticity of real samples and features of the target scene is determined to obtain the sample type; the sample type includes real samples and generated samples;

[0058] When the generated target scene features are the real samples, the target scene features and the real samples of the target scene are determined as a transfer feature set;

[0059] The generator loss function can be obtained using the following formula:

[0060]

[0061]

[0062] in, It is a periodic feature subset; A single periodic feature within a subset of periodic features; The generated target scene features; This represents the probability of the discriminator identifying the authenticity of the target scene features. To counteract the loss; Weighted hyperparameters for preserving the loss in the structure; To preserve the structure loss function, This represents the transformation function used for structure extraction;

[0063] The discriminator loss function can be obtained using the following formula:

[0064]

[0065] in, The set of real samples for the target scene; A single real sample within the set of real samples in the target scene; This represents the probability that the discriminator classifies a real sample as genuine. The target scene feature set generated by the generator based on a periodic feature subset; The loss term is intended to help the discriminator correctly identify real samples; The goal is to enable the discriminator to correctly identify the loss term of the generated samples.

[0066] As described above, the generator network model is first constructed. periodic feature subsets As input, the output is the transfer feature of the target scene style, i.e., the target scene feature. During the generator's training, its optimization objective includes not only making the generated results as realistic as possible in the target domain, but also achieving distribution mapping while preserving the periodic arc behavior structure, i.e., minimizing the generator's loss function. The generator's loss function consists of two parts: adversarial loss and structure preservation loss. The adversarial loss measures whether the generated sample can deceive the discriminator; the structure preservation loss constrains the generated sample in the target scene to maintain a similar physical structure to the source sample in the higher-order feature space. This loss design ensures that the generator not only focuses on the similarity of sample distribution, but also must preserve the frequency band energy structure and singular principal directions of the original arc during the generation process.

[0067] Training the discriminator model This is used to distinguish between the target scene features output by the generator and real target scene samples. The discriminator receives two input sources: a set of periodic samples from the real target scene. Each sample It should be judged as "true"; the set of target scene samples output by the generator. This should be classified as "fake". The training objective of the discriminator is to maximize its ability to distinguish between real and fake samples, that is, to maximize the discriminator loss function:

[0068] First item This indicates the desire for the discriminator to correctly determine the authenticity of the real target sample;

[0069] Second item This indicates a desire for it to be able to identify the fake attributes of the generated samples.

[0070] The discriminator and generator are optimized alternately during training, approaching the target distribution boundary through continuous game, pushing the generator to output more realistic samples, while enhancing the discriminator's sensitivity to small distribution differences.

[0071] After the discriminator training converges, the target scene features output by the generator are fused with real target samples to construct the final transfer feature set for cross-scene training. Specifically, if the discriminator outputs a probability higher than a set threshold for a generated sample, the sample is considered a high-realism sample and is determined as a transfer feature; if it is a real target scene sample, it is directly included in the transfer feature set. Through this process, the highly structured periodic features of the source scene are effectively fused with the actual distribution features of the target scene, thereby significantly improving the domain diversity and representativeness of the samples and providing high-quality support for the subsequent learning of the classifier in complex and inconsistent environments.

[0072] Furthermore, the classifier is trained using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, including:

[0073] The periodic features in the periodic feature subset are labeled with arc fault types to obtain a labeled training set; the labeled data in the labeled training set includes periodic features and their corresponding arc fault types; the arc fault types include series arc faults, parallel arc faults, and normal.

[0074] Based on the adversarial training strategy and the dynamic reweighting mechanism, the classifier's model parameters are adjusted by combining the labeled training set and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model. The cross-scenario AC arc fault diagnosis model is used to output the arc fault type according to the arc signal of the scenario. The adversarial training strategy is used to enable the classifier to have discriminative adaptation ability to different sample domains. The dynamic reweighting mechanism is used to give high confidence samples high training weights.

[0075] As described above, manual or semi-automatic labeling is performed based on a subset of periodic features to construct a training sample set with semantic labels. Optionally, for some or every periodic feature vector in the subset of periodic features, the corresponding arc behavior type is manually confirmed by combining the bus current and voltage signal waveforms of its source periodic segment, and its fault attribute is labeled as a series arc fault, parallel arc fault, or normal fault-free state. This labeling process can refer to actual switch operation records, load disturbance triggering information, or waveform feature templates determined by experts to ensure labeling accuracy. The final labeled training set consists of a set of feature vectors and their corresponding arc type labels.

[0076] The labeled training set and the generated transfer feature set are jointly input into the classifier model, and the model is trained primarily using an adversarial training strategy. During training, a dynamic reweighting mechanism is introduced to improve training efficiency and controllability of sample contributions. Through this training strategy, the resulting cross-scenario AC arc fault diagnosis model can be deployed in field control systems or edge intelligent terminals. It receives periodic feature vectors extracted from any new scenario as input and automatically outputs the corresponding arc fault type, demonstrating adaptability and accuracy in scenarios involving equipment changes, voltage level fluctuations, or load switching.

[0077] Please refer to Figure 4 Another embodiment of the present invention provides a cross-scenario AC arc fault diagnosis system. The system includes:

[0078] The data acquisition module is used to acquire AC bus current signals and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain the original scenario signals. The preset classic scenarios cover different load types, power supply environments and equipment models.

[0079] The feature extraction module is used to perform wavelet packet decomposition and singular value decomposition on the original scene signal to obtain a set of periodic feature vectors;

[0080] A periodic feature filtering module is used to filter the periodic feature vector set to satisfy the periodic arc structure features, thereby obtaining a periodic feature subset.

[0081] The training set construction module is used to perform scene transfer based on the periodic feature subset using a generative adversarial network to obtain a transfer feature set, which includes real scene samples and generated samples.

[0082] The model training module is used to train the classifier using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, and to use the cross-scenario AC arc fault diagnosis model to diagnose arc faults.

[0083] In another aspect of the invention, a computer device is provided. The computer device includes a memory configured to store a computer program; and a processor configured to execute the computer program to perform the aforementioned cross-scenario AC arc fault diagnosis method.

[0084] In another aspect of the invention, a computer-readable medium is provided. This medium stores a computer program that is executed by a processor to implement the aforementioned cross-scenario AC arc fault diagnosis method.

[0085] The cross-scenario AC arc fault diagnosis method, system, device, and storage medium of the present invention are described below through specific embodiments:

[0086] Please refer to Figure 1 One embodiment of the present invention provides a cross-scenario AC arc fault diagnosis method. This embodiment illustrates the method by applying it to a terminal. It is understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and can be implemented through the interaction between the terminal and the server. In this embodiment, the method includes steps S101 to S105.

[0087] S101. Acquire AC bus current and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain the original scenario signal. The preset classic scenarios cover different load types, power supply environments and equipment models.

[0088] This study schematically collects current and voltage signal data under AC arc fault and non-fault conditions from multiple classic power supply scenarios to form the original scenario signals. Classic scenarios refer to representative power supply configurations in actual environments, including load-side power grid operating conditions composed of resistive loads such as electric furnaces, inductive loads such as motors, and capacitive loads such as compensation capacitor banks. These scenarios cover common environments such as low-voltage distribution networks, high-voltage power frequency systems, and complex load switching states. AC arc fault conditions typically manifest as voltage spikes, oscillating currents, and periodic disturbances caused by intermittent load contact, while non-arc fault conditions exhibit stable waveforms under normal load connection. To ensure the diagnostic model has good versatility, the sampling signals should cover bus voltage and current signals under various operating conditions. Since arcing generates high-frequency signals, the sampling frequency should be much higher than the fundamental power frequency to capture high-frequency arc characteristics.

[0089] S102. Perform wavelet packet decomposition and singular value decomposition on the original scene signal to obtain a set of periodic feature vectors.

[0090] Indicatively, the original signal set undergoes multi-scale processing and feature compression to construct learnable periodic feature vectors. Specifically, wavelet packet decomposition (WPD) is used to decompose signal segments of each period length. As a generalization of wavelet analysis, wavelet packet decomposition can completely decompose the signal in the time-frequency domain, obtaining more detailed high-frequency and low-frequency sub-band coefficients, avoiding the problem of insufficient high-frequency resolution in ordinary wavelets. For example, during the decomposition process, mother wavelets such as Daubechies or Symlets, which have tight support and good orthogonality, are selected. After setting the number of decomposition levels, the signal is divided into frequency bands, each corresponding to a set of energy coefficients within a local frequency range. Further, the sub-band coefficient sets corresponding to each period segment are used to construct a two-dimensional sub-band energy matrix. This matrix represents the frequency band number horizontally and the sample time window vertically. To further reduce dimensionality and preserve the main structural features, Singular Value Decomposition (SVD) is performed on the energy matrix of this sub-band. The decomposition yields singular value vectors, which reflect the energy distribution pattern of the sample along each principal axis. The earlier singular values ​​are retained as compressed vectors representing the signal energy structure, i.e., the periodic feature vectors constituting this periodic segment, ultimately yielding the feature vector set corresponding to the complete set of periodic segments.

[0091] S103. The periodic feature vector set is filtered to meet the periodic arc structure characteristics to obtain a periodic feature subset.

[0092] This study illustrates how a subset of periodic arc structures with typical periodic arc structures is selected from a set of periodic feature vectors to provide more representative fault samples for subsequent migration and discrimination. Periodic arcs typically exhibit energy concentration and frequency band variation patterns that repeat across multiple cycles, demonstrating strong temporal structural consistency. Specifically, to capture this feature, a multi-cycle sliding window is constructed, aggregating sub-frequency band energy matrices from multiple adjacent cycle segments within the window, and tensorizing them to preserve the time dimension. Using Progressive Singular Value Decomposition (PSVD), the tensor within the window is expanded into a two-dimensional matrix, and PSVD is performed to extract the principal components of cross-cycle variations, thereby measuring its periodic consistency. Furthermore, based on the average trend of energy evolution across cycles in each frequency band, a normalized frequency-band power allocation (FPA) index is calculated, which quantifies the stability of arc energy distribution in the frequency domain. Under the conditions of satisfying singular value stability and the FPA change amplitude being less than a set threshold, periodic samples within the window are considered to have periodic arc structure characteristics and are included in the periodic feature subset.

[0093] S104. Using a generative adversarial network, scene transfer is performed based on the periodic feature subset to obtain a transfer feature set, which includes real scene samples and generated samples.

[0094] To illustrate, Generative Adversarial Networks (GANs) are introduced to transfer periodic fault features from the source scene to the target scene, addressing the issues of sample scarcity and training bias. For example, the GAN structure is based on a modified CycleGAN design, comprising two sub-modules: a generator and a discriminator. The generator aims to learn to map periodic features from the source domain to a feature distribution characteristic of the target domain. Its training objectives include not only traditional adversarial loss but also a structure-preserving loss function to ensure that the mapped samples retain key arc characteristics in terms of singular value structure and frequency band energy distribution. Specifically, the structure-preserving loss measures the difference between the source and generated samples in a predefined structure space using Euclidean distance or subspace projection error. The discriminator is responsible for identifying the distributional differences between the generated samples and real samples in the target scene. Its optimization objective is to improve the recognition rate of real samples while suppressing the deception success rate of generated samples. Through alternating training using the aforementioned adversarial game mechanism, a set of transfer samples with a target domain distribution style but retaining arc structure characteristics is eventually formed, called the transfer feature set. The transfer feature set includes a portion of real target samples and a portion of pseudo target samples mapped by the generator.

[0095] Specifically, step S104 includes:

[0096] S1041. Based on minimizing the generator loss objective, feature mapping is performed on the periodic feature subset to generate target scene features; the generator loss includes adversarial loss and structure preservation loss; the structure preservation loss is used to constrain the target scene samples to retain the arc feature structure in the periodic feature subset.

[0097] The generator loss function can be obtained using the following formula:

[0098]

[0099]

[0100] in It is a periodic feature subset; A single periodic feature within a subset of periodic features; The generated target scene features; This represents the probability of the discriminator identifying the authenticity of the target scene features. To counteract the loss; Weighted hyperparameters for preserving the loss in the structure; To preserve the structure loss function, This represents the transformation function used for structure extraction.

[0101] First, construct the generator network model. periodic feature subsets As input, the output is the transfer feature of the target scene style, i.e., the target scene feature. During the generator training process, its optimization objective includes not only making the generated results as realistic as possible in the target domain, but also achieving distribution mapping while maintaining the periodic arc behavior structure, i.e., minimizing the generator loss function. The generator loss function consists of two parts: adversarial loss and structure-preserving loss. The adversarial loss measures whether the generated sample can fool the discriminator; the structure-preserving loss constrains the generated sample of the target scene to maintain a similar physical structure to the source sample in the high-order feature space. The transformation function used for structure extraction can be a shallow autoencoder or a projector. This loss design ensures that the generator not only focuses on the similarity of sample distributions, but also preserves the characterization properties of the original arc, such as its frequency band energy structure and singular principal directions, during the generation process.

[0102] S1042. Based on maximizing the discriminator loss target, determine the authenticity probability of real samples in the target scene and the features of the target scene to obtain the sample type; the sample type includes real samples and generated samples.

[0103] The discriminator loss function can be obtained using the following formula:

[0104]

[0105] in, The set of real samples for the target scene; A single real sample within the set of real samples in the target scene; This represents the probability that the discriminator classifies a real sample as genuine. The target scene feature set generated by the generator based on a periodic feature subset; The loss term is intended to help the discriminator correctly identify real samples; The goal is to enable the discriminator to correctly identify the loss term of the generated samples.

[0106] Training the discriminator model This is used to distinguish between the target scene features output by the generator and real target scene samples. The discriminator receives two input sources: a set of periodic samples from the real target scene. Each sample It should be judged as "true"; the set of target scene samples output by the generator. This should be classified as "fake". The training objective of the discriminator is to maximize its ability to distinguish between real and fake samples, that is, to maximize the discriminator loss function:

[0107] First item This indicates the desire for the discriminator to correctly determine the authenticity of the real target sample;

[0108] Second item This indicates a desire for it to be able to identify the fake attributes of the generated samples.

[0109] The discriminator and generator are optimized alternately during training, approaching the target distribution boundary through continuous game, pushing the generator to output more realistic samples, while enhancing the discriminator's sensitivity to small distribution differences.

[0110] S1043. When the generated target scene features are the real samples, the target scene features and the real samples of the target scene are determined as a transfer feature set.

[0111] Indicatively, after the discriminator training converges, the target scene features output by the generator are fused with real target samples to construct the final transfer feature set for cross-scene training. Specifically, if the discriminator's output probability for a generated sample is higher than a set threshold, the sample is considered a high-realism sample and is determined as a transfer feature; if it is a real target scene sample, it is directly included in the transfer feature set. Through this process, the highly structured periodic features of the source scene are effectively fused with the actual distribution features of the target scene, thereby significantly improving the domain diversity and representativeness of the samples and providing high-quality support for the subsequent learning of the classifier in complex and inconsistent environments.

[0112] S105. The classifier is trained using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, and the cross-scenario AC arc fault diagnosis model is used to diagnose arc faults.

[0113] This paper illustrates a fault identification model constructed based on a periodic feature subset and a transfer feature set. To improve the model's adaptability to samples from different sources, a joint discriminant classifier is constructed. This classifier can be a multilayer perceptron (MLP) or a convolutional long short-term memory network (CNN-LSTM) structure to simultaneously model frequency domain spatial features and time series information. During training, an auxiliary discriminator is introduced to identify whether the input sample is generated by the generator. Combined with an adversarial training strategy, this encourages the classifier to output stable classification results without relying on input domain information. Furthermore, to suppress the negative impact of low-quality generated samples on training, a dynamic reweighting mechanism is adopted. Based on the discriminator's judgment of sample authenticity, dynamic training weights are assigned to each sample, allowing high-confidence samples to contribute more to training. Finally, by fusing real target samples and generated transfer samples, the classifier parameters are optimized under cross-domain training conditions to form an AC arc fault diagnosis model with cross-scenario generalization capabilities.

[0114] In the aforementioned method for constructing a cross-scenario AC arc fault diagnosis model based on GAN, bus current and voltage data are collected under multiple typical power supply environments, and a unified periodic feature vector set is constructed. This effectively alleviates the data distribution offset problem caused by changes in power grid conditions, providing structured input for subsequent unified modeling. By introducing a periodic structure recognition and filtering mechanism into the periodic feature vector set, only highly reliable features with time-frequency consistency and frequency band energy concentration trends are retained, effectively eliminating non-periodic disturbances and improving the anti-interference capability of the diagnostic model. By constructing a generative adversarial network that combines adversarial loss and structure preservation loss, adaptive mapping of source scenario samples to target scenario styles is achieved, ensuring the physical authenticity of generated samples in terms of structural features, thereby expanding the model's applicability. By constructing a hybrid training set combining real target samples and transferred generated samples, a fault recognition model with controllable performance and clear classification boundaries can still be efficiently trained even when labeled data is lacking in the target scenario. The final cross-scenario diagnostic model has input adaptation and real-time recognition capabilities, and can be embedded and deployed in different power supply equipment to meet the multi-scenario needs of fault monitoring in industrial sites.

[0115] Specifically, step S105 includes:

[0116] S1051. The periodic features in the periodic feature subset are labeled with arc fault types to obtain a labeled training set; the labeled data in the labeled training set includes periodic features and their corresponding arc fault types; the arc fault types include series arc faults, parallel arc faults and normal.

[0117] In a schematic manner, manual or semi-automatic labeling is performed based on a subset of periodic features to construct a training sample set with semantic labels. Optionally, for some or every periodic feature vector in the subset of periodic features, the corresponding arc behavior type is manually confirmed by combining the bus current and voltage signal waveforms of its source periodic segment, and its fault attribute is labeled as a series arc fault, parallel arc fault, or normal fault-free state. This labeling process can refer to actual switch operation records, load disturbance triggering information, or waveform feature templates determined by experts to ensure labeling accuracy. The final labeled training set consists of a set of feature vectors and their corresponding arc type labels.

[0118] S1052. Based on the adversarial training strategy and the dynamic reweighting mechanism, the classifier's model parameters are adjusted by combining the labeled training set and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model. The cross-scenario AC arc fault diagnosis model is used to output the arc fault type according to the scene arc signal. The adversarial training strategy is used to enable the classifier to have discriminative adaptation ability to different sample domains. The dynamic reweighting mechanism is used to give high-confidence samples high training weights.

[0119] The labeled training set and the generated transfer feature set are jointly input into the classifier model, and the model is trained primarily using an adversarial training strategy. The classifier itself can be an MLP, a convolutional neural network (CNN), or a CNN-LSTM structure, with the architecture selected based on the feature dimension and sequence characteristics. MLPs are suitable for vectors with fixed and unordered feature dimensions. CNN-LSTMs are suitable for sub-band sequences with clear local temporal relationships. To avoid the model overfitting to the labeled sample domain and ignoring the distributional differences in the transfer sample domain, a domain discriminator is introduced during training. As an auxiliary module, it is used to determine whether the input sample comes from the generated domain or the true labeled domain. During training, the classifier needs to consider two tasks: accurately classifying arc fault types and maintaining output stability without relying on sample domain information, thereby achieving stronger cross-domain discriminative adaptability. For example, the total loss for classifier training is... ,in, ,in, The loss is calculated by comparing the generated arc fault types with the labeled arc fault types in the model. Domain discriminator The losses in the fight, For all training data input, The training set is labeled.

[0120] During training, a dynamic reweighting mechanism is introduced to improve training efficiency and controllability of sample contributions. Specifically, the discriminator output confidence score is calculated for all input samples. Samples with higher confidence scores are assigned larger training weights, while transfer samples with uncertain discrimination results or low confidence scores are assigned smaller weights. The weight factors can be dynamically generated using a sigmoid function or a softmax normalization strategy to ensure the dominance of high-quality samples in the parameter update process. This strategy effectively alleviates the imbalance in the quality distribution of transfer samples, improving the convergence speed and accuracy of the classifier in recognizing arc pattern structures. For example, to prevent the proportion of generated samples from being too large or too small from interfering with classification, the sample loss weights are dynamically adjusted with the number of training epochs or the discrimination confidence score.

[0121]

[0122] Target scene features that approximate the distribution of real samples are given higher training weights to improve the generalization ability of the classifier.

[0123] Through the above training strategy, the resulting cross-scenario AC arc fault diagnosis model can be deployed in field control systems or edge intelligent terminals. It can receive periodic feature vectors extracted from any new scenario as input and automatically output the corresponding arc fault type. It has adaptability and accuracy in scenarios such as equipment changes, voltage level fluctuations, or load switching.

[0124] Please refer to Figure 2 An embodiment of the present invention provides a method for generating a set of periodic feature vectors, the method comprising steps S201 to S205.

[0125] S201. The original scene signal is periodically segmented to obtain a periodic segment sequence, which includes multiple periodic segments.

[0126] Specifically, the original scene data is periodically segmented to construct a sequence of periodic segments with temporal boundary characteristics. For example, the original scene signal is... The sampling frequency is The number of sampling points corresponding to each cycle is calculated by detecting the zero point of the AC signal or based on the power frequency cycle. Based on this, the original bus current and voltage signals are divided into continuous, non-overlapping periodic segments. Each periodic segment in this periodic segment sequence All of them are signal segments of consistent length that contain characteristic changes within a complete period, and have good local time series comparability.

[0127] S202. Perform wavelet packet decomposition on the periodic segment to obtain a sub-frequency band coefficient group. The sub-frequency band coefficient group includes multiple sub-frequency band nodes, and each sub-frequency band node includes signal coefficients within a corresponding local frequency range.

[0128] Indicatively, wavelet packet decomposition is performed on each periodic segment. Wavelet packet decomposition is a generalization of the standard wavelet transform, which not only further decomposes the low-frequency components but also further subdivides the high-frequency components within the subbands, thereby obtaining the multi-scale energy distribution of the signal across the entire frequency band. Specifically, for each periodic segment... Choose appropriate wavelet basis functions, such as Daubechies or Symlet, and set the number of decomposition levels. The periodic signal is subdivided into Each sub-band node corresponds to the signal's response, or signal coefficients, within a specific local frequency range. ,in, , indicating the first Each sub-band has its coefficients representing the instantaneous energy or fluctuation characteristics within that band. Ultimately, a sub-band coefficient set is formed for each period segment, containing multiple sub-band nodes and their corresponding signal coefficients.

[0129] S203. Assemble the sub-band coefficient groups corresponding to each periodic segment into a two-dimensional matrix to obtain the sub-band energy matrix.

[0130] Indicatively, the sub-band coefficients corresponding to each periodic segment are assembled into a two-dimensional matrix. This involves constructing the sub-band energy matrix for the periodic segment. The rows of this matrix represent the sub-band numbers, and the columns represent the discrete sampling points or statistical descriptions of the sub-band coefficients in the time domain, such as average, variance, and maximum values, thus representing the periodic segment. Energy distribution across each frequency band dimension. The sub-band energy matrix preserves the full-frequency domain distribution structure of the periodic segments and serves as an important intermediate representation for subsequent compression and feature extraction.

[0131] S204. Perform singular value decomposition on the sub-band energy matrix and select singular values ​​of a set order to obtain the time-frequency feature matrix.

[0132] This illustration demonstrates singular value decomposition (SVD) based on the sub-band energy matrix. SVD, a classic matrix factorization method, maps high-dimensional data to a low-dimensional orthogonal space, extracting the most significant structural features. Specifically, the sub-band energy matrix is ​​decomposed into three parts: the product of the left singular vector, the singular value diagonal matrix, and the right singular vector. The singular values ​​represent the energy concentration of that period along the corresponding principal component direction. For example, the sub-band energy matrix... Perform singular value decomposition to obtain , where the singular value vector is As a compact representation of the principal component energy distribution, to achieve information compression and compact feature representation, only the first few elements are selected. Singular values ​​of order form a one-dimensional vector This is used to express the main time-frequency structure of the current periodic segment, that is, to obtain the time-frequency feature vector of the periodic segment.

[0133] S205. Construct a set of periodic feature vectors based on the time-frequency feature matrices corresponding to all periodic segments in the periodic segment sequence.

[0134] To illustrate, the time-frequency feature vectors corresponding to all periodic segments in the periodic segment sequence are aggregated to construct a periodic feature vector set. This vector set, serving as a high-level semantic representation of the periodic behavior in the original scene, not only preserves the temporal structure and frequency domain energy characteristics of the original signal but also significantly reduces the feature dimension, facilitating subsequent identification and transfer modeling of periodic arc behavior.

[0135] Please refer to Figure 3 An embodiment of the present invention provides a method for generating a periodic feature subset, the method comprising steps S301 to S305.

[0136] S301. A sliding window is used to construct a multi-period input on the periodic segment sequence to obtain a periodic sequence window.

[0137] Schematic illustration of using a sliding window mechanism in a periodic segment sequence Construct a multi-period input window to form a set of periodic sequence windows. The window uses fixed-length P period segments as units and moves across the entire period sequence with a set sliding step size of 1. This allows each window to contain periodic feature data from multiple consecutive periods, which helps to capture the structural change patterns across periods. It is particularly suitable for analyzing the continuous and oscillating pattern characteristics of arc-type faults over multiple periods.

[0138] S302. Stack the corresponding sub-band energy matrices within the periodic sequence window to obtain a three-dimensional tensor.

[0139] Indicatively, for each periodic sequence window, the window is... Sub-band energy matrix corresponding to each periodic segment Stack them to form a three-dimensional tensor The three dimensions of this tensor correspond to the frequency band number, time characteristics, and periodic order, respectively. By preserving information in the periodic dimension, this three-dimensional tensor effectively encapsulates the frequency energy evolution path across periods, enabling subsequent modeling of periodic consistency and perturbation modes.

[0140] S303. Expand the three-dimensional tensor into a two-dimensional matrix according to the periodic dimension, and perform progressive singular value decomposition to obtain the windowed composite singular value vector.

[0141] To illustrate, a three-dimensional tensor is flattened along its periodic dimension, that is, the periodic dimension is expanded and merged with the time dimension to obtain a two-dimensional matrix. This facilitates processing within traditional matrix decomposition methods. PSVD is applied to this two-dimensional matrix to extract the principal transformation direction within the entire window, i.e., the window-combined singular value vector. Unlike singular value compression of single-frame signals, PSVD emphasizes preserving the consistency and main structural trend of periodic variations. The generated composite singular value vector can be regarded as the overall characteristic fingerprint of the periodic window in terms of energy distribution and structural stability. If multiple periods within the window exhibit similar singular value distributions in the main direction, it often means that there is a strong periodic arc disturbance mode.

[0142] S304. Calculate the frequency band energy distribution index for the corresponding sub-band energy matrix within the periodic sequence window to obtain the frequency band energy normalization index.

[0143] Furthermore, to determine the stability and concentration of frequency domain energy within the window, a frequency band energy distribution index is calculated for the sub-band energy matrix of each period within the periodic sequence window. The frequency band energy distribution index is obtained by statistically analyzing the average energy of each frequency band within the periodic window and normalizing the energy values ​​of each band to form a normalized frequency band energy index vector. This index reflects the degree of shift and concentration trend of arc energy in the frequency domain distribution. When the energy of certain frequency bands maintains a consistently high proportion over multiple periods, it usually indicates that the arc characteristics possess structural stability within that frequency band.

[0144] Specifically, step S304 includes:

[0145] S3041. Perform frequency band energy statistics based on the corresponding sub-band energy matrix within the periodic sequence window to obtain the average energy distribution vector.

[0146] Indicatively, frequency band energy statistics are performed on the sub-band energy matrix corresponding to each periodic segment within a periodic sequence window to extract the average energy level of each frequency band over multiple periods. Specifically, a periodic sequence window contains P periodic segments, and the sub-band energy matrix formed by wavelet packet decomposition of each segment has j frequency band nodes, i.e., the energy of the j-th frequency band in the i-th period is:

[0147]

[0148] Then each periodic sequence window can be used The average energy distribution vector is obtained by summing the energy values ​​of all cycles in each frequency band:

[0149]

[0150] S3042. The frequency band energy normalization index is calculated based on the energy distribution of each frequency band within the period and the average energy distribution vector.

[0151] Specifically, to ensure the comparability of energy distributions across different frequency bands and to eliminate the impact of overall energy amplitude variations on feature determination, the aforementioned average energy distribution vector is normalized to form a frequency band energy normalization index:

[0152]

[0153] This index describes the relative distribution of energy across frequency bands, and its shape and distribution pattern can be used to identify the unique frequency domain behavior patterns of arc faults. In particular, when a certain type of arc fault repeatedly accumulates energy in several frequency bands, the proportion of that frequency band in the normalized index will be significantly higher than in other frequency bands, thus forming a recognizable structural template. This index can also effectively eliminate interference caused by differences in overall signal strength due to voltage levels, load variations, etc., allowing subsequent feature selection to focus more on the structural essence rather than signal amplitude.

[0154] S305. Based on the window comprehensive singular value vector and the frequency band energy normalization index of each periodic sequence window, the feature vectors corresponding to the periodic feature vector set are filtered to obtain a periodic feature subset.

[0155] In a schematic manner, by comprehensively considering the singular value vector and frequency band energy normalization index corresponding to each periodic sequence window, and by setting multiple screening thresholds or clustering methods, the original periodic feature vectors in the periodic feature vector set are screened. Specifically, if a periodic sequence window has high stability in the singular value principal component direction and its frequency band energy distribution concentration meets the set requirements, then the window is determined to be a region with a significant periodic arc structure, and the feature vectors corresponding to the periodic segments contained therein are retained and included in the periodic feature subset. This screening process eliminates interference data or unstable arc signals that do not have periodic regularity, effectively improving the feature quality and sample representativeness of subsequent transfer modeling and classification training.

[0156] Specifically, step S305 includes:

[0157] S3051. The stability of the window comprehensive singular value vector and the frequency band energy normalization index of each periodic sequence window is analyzed by KL divergence analysis to obtain the stability analysis results of the periodic sequence window. The stability analysis results include stable arc periodic behavior segments and unstable arc periodic behavior segments.

[0158] The KL divergence analysis involves the following steps:

[0159] If the window's integrated singular value vector satisfies the condition of periodic change trend, and the frequency band energy normalization index satisfies the condition of fixed frequency band energy concentration trend, then the corresponding periodic sequence window is determined to be a stable arc periodic behavior segment.

[0160] If the window's integrated singular value vector does not meet the periodic change trend condition, or the frequency band energy normalization index does not meet the fixed frequency band energy concentration trend condition, then the corresponding periodic sequence window is determined to be an unstable arc periodic behavior segment.

[0161] To illustrate, Kullback-Leibler divergence (KL divergence) is introduced as the core criterion to analyze the distribution stability of the window-combined singular value vector and the frequency band energy normalization index in each periodic sequence window over time. KL divergence essentially measures the degree of information difference between two probability distributions; a smaller value indicates greater similarity, while a larger value indicates more significant differences. Specifically, the combined singular value vector of each periodic sequence window is normalized and treated as a probability density distribution in a high-dimensional space. A KL divergence sequence between adjacent windows is constructed to capture the changing trend of periodic behavior. Similarly, the frequency band energy normalization index is used as a probability vector describing the frequency domain structure, and its cross-period KL divergence trend is used to determine whether energy is concentrated in a fixed frequency band. When the singular value vector of a periodic sequence window exhibits a clear periodic structure—that is, the KL divergence changes between adjacent windows show low amplitude and high repetition characteristics, and the corresponding frequency band energy normalization index consistently occupies a dominant proportion in a specific frequency band—it can be considered that stable and periodically prominent arc behavior exists within that window. At this point, the window is determined to be a stable arc cycle behavior segment; conversely, if any indicator shows violent fluctuations or lacks a concentrated trend, the window is determined to be an unstable arc cycle behavior segment. Optionally, the stability analysis results are recorded in the form of labels for the structural characteristics of each cycle sequence window for subsequent screening.

[0162] S3052. If the stability analysis result of the periodic sequence window is a stable arc periodic behavior segment, then the feature vector corresponding to the periodic feature vector set is selected as the periodic feature.

[0163] In a schematic manner, based on the stability analysis results, feature vector filtering is performed on the periodic sequence windows identified as stable arc periodic behavior segments. Specifically, the periodic segments corresponding to each periodic sequence window have been mapped to several periodic feature vectors. Under the premise of confirming structural stability within the window, all periodic segments contained in the window are selected. The corresponding feature vectors are labeled as periodic features. This ensures that the selected samples not only have energy structure characteristics within a single period, but also have pattern consistency and frequency band energy commonality across period sequences, thus improving the physical semantic reliability of the samples.

[0164] S3053. The periodic feature subset is constructed by constructing the periodic features corresponding to the periodic sequence windows of the multiple stable arc periodic behavior segments.

[0165] Specifically, the periodic features contained in all periodic sequence windows that are determined to be stable arc periodic behavior segments are summarized to construct a periodic feature subset. Compared to the initial periodic feature vector set, this subset eliminates aperiodic disturbances, transient anomalies, and invalid segments, while retaining periodic segment features that remain stable and representative under multidimensional structural analysis.

[0166] In one exemplary embodiment, such as Figure 4 As shown, a cross-scenario AC arc fault diagnosis system is provided, the system comprising:

[0167] Data acquisition module 401 is used to acquire AC bus current signals and bus voltage signals under arc fault and non-fault states in multiple preset classic scenarios to obtain original scenario signals. The preset classic scenarios cover different load types, power supply environments and equipment models.

[0168] The feature extraction module 402 is used to perform wavelet packet decomposition and singular value decomposition on the original scene signal to obtain a set of periodic feature vectors;

[0169] The periodic feature filtering module 403 is used to filter the periodic feature vector set to satisfy the periodic arc structure features, thereby obtaining a periodic feature subset.

[0170] The training set construction module 404 is used to perform scene transfer based on the periodic feature subset using a generative adversarial network to obtain a transfer feature set, wherein the transfer feature set includes real scene samples and generated samples;

[0171] The model training module 405 is used to train the classifier using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, and to use the cross-scenario AC arc fault diagnosis model to perform arc fault diagnosis.

[0172] In one embodiment, it further includes:

[0173] The periodic segmentation module is used to periodically segment the original scene data to obtain a periodic segment sequence; the periodic segment sequence includes multiple periodic segments.

[0174] The wavelet packet decomposition module is used to perform wavelet packet decomposition on periodic segments to obtain sub-band coefficient groups; the sub-band coefficient groups include multiple sub-band nodes; the sub-band nodes include signal coefficients within the corresponding local frequency range;

[0175] The feature extraction module 402 is also used to assemble the sub-band coefficient groups corresponding to each period segment into a two-dimensional matrix to obtain the sub-band energy matrix;

[0176] The singular value decomposition module is used to perform singular value decomposition on the sub-band energy matrix and select singular values ​​of a set order to obtain the time-frequency feature matrix.

[0177] The feature extraction module 402 is also used to construct a set of periodic feature vectors based on the time-frequency feature matrices corresponding to all periodic segments in the periodic segment sequence.

[0178] In one embodiment, it further includes:

[0179] The window segmentation module is used to construct a multi-period input on a periodic segment sequence using a sliding window to obtain a periodic sequence window;

[0180] The data processing module is used to stack the corresponding sub-band energy matrices within the periodic sequence window to obtain a three-dimensional tensor.

[0181] The singular value decomposition module is also used to expand a three-dimensional tensor into a two-dimensional matrix along the periodic dimension and perform progressive singular value decomposition to obtain a windowed composite singular value vector.

[0182] The energy distribution module is used to calculate the frequency band energy distribution index of the corresponding sub-frequency band energy matrix within the periodic sequence window, and obtain the frequency band energy normalization index.

[0183] The periodic feature filtering module 403 is also used to filter the feature vectors corresponding to the periodic feature vector set based on the window comprehensive singular value vector and the frequency band energy normalization index of each periodic sequence window, so as to obtain a periodic feature subset.

[0184] In one embodiment, the energy distribution module is further configured to perform frequency band energy statistics based on the corresponding sub-band energy matrix within the periodic sequence window to obtain an average energy distribution vector;

[0185] The energy distribution module is also used to calculate the frequency band energy normalization index based on the energy distribution of each frequency band and the average energy distribution vector within the period.

[0186] In one embodiment, it further includes:

[0187] The stability analysis module is used to analyze the stability of the window-combined singular value vector and frequency band energy normalization index of each periodic sequence window using KL divergence analysis, and obtain the stability analysis results of the periodic sequence window; the stability analysis results include stable arc periodic behavior segments and unstable arc periodic behavior segments;

[0188] The periodic feature filtering module 403 is also used to select the feature vector corresponding to the periodic feature vector set as the periodic feature if the stability analysis result of the periodic sequence window is a stable arc periodic behavior segment.

[0189] The periodic feature filtering module 403 is also used to construct a periodic feature subset from the periodic features corresponding to multiple periodic sequence windows that are stable arc periodic behavior segments.

[0190] In one embodiment, it further includes:

[0191] The generator module is used to generate target scene features based on feature mapping of a periodic feature subset, with the objective of minimizing the generator loss. The generator loss includes adversarial loss and structure preservation loss. The structure preservation loss is used to constrain the target scene samples to retain the arc feature structure in the periodic feature subset.

[0192] The discriminator module is used to determine the authenticity probability of real samples and features of the target scene based on maximizing the discriminator loss target, and to obtain the sample type; the sample type includes real samples and generated samples;

[0193] The Generative Adversarial Network (GAN) module is used to determine the target scene features and the target scene real samples as the transfer feature set when the generated target scene features are real samples.

[0194] In one embodiment, it further includes:

[0195] The training set annotation module is used to annotate the periodic features in the periodic feature subset with arc fault types to obtain an annotated training set; the annotated data includes periodic features and their corresponding arc fault types; the arc fault types include series arc faults, parallel arc faults, and normal.

[0196] The joint training module is used to adjust the model parameters of the classifier by jointly labeling the training set and the transfer feature set based on the adversarial training strategy and the dynamic reweighting mechanism, so as to obtain a cross-scenario AC arc fault diagnosis model. The cross-scenario AC arc fault diagnosis model is used to output the arc fault type according to the arc signal of the scenario. The adversarial training strategy is used to enable the classifier to have the ability to discriminate and adapt to different sample domains. The dynamic reweighting mechanism is used to give high confidence samples high training weights.

[0197] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0198] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0199] 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 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 cross-scenario AC arc fault diagnosis method, characterized in that, The method includes: Acquire AC bus current and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain the original scenario signal. The preset classic scenarios cover different load types, power supply environments and equipment models. The original scene signal is subjected to wavelet packet decomposition and singular value decomposition to obtain a periodic feature vector set, including: periodically segmenting the original scene signal to obtain a periodic segment sequence, the periodic segment sequence including multiple periodic segments; performing wavelet packet decomposition on the periodic segments to obtain sub-frequency band coefficient groups, the sub-frequency band coefficient groups including multiple sub-frequency band nodes, the sub-frequency band nodes including signal coefficients within a corresponding local frequency range; assembling the sub-frequency band coefficient groups corresponding to each periodic segment into a two-dimensional matrix to obtain a sub-frequency band energy matrix; performing singular value decomposition on the sub-frequency band energy matrix and selecting singular values ​​of a set order to obtain a time-frequency feature matrix; and constructing a periodic feature vector set based on the time-frequency feature matrices corresponding to all periodic segments in the periodic segment sequence. The periodic feature vector set is filtered to satisfy the periodic arc structure characteristics to obtain a periodic feature subset, including: constructing a multi-period input on the periodic segment sequence using a sliding window to obtain a periodic sequence window; stacking the corresponding sub-band energy matrices within the periodic sequence window to obtain a three-dimensional tensor; expanding the three-dimensional tensor into a two-dimensional matrix along the periodic dimension and performing progressive singular value decomposition to obtain a window-comprehensive singular value vector; calculating the frequency band energy distribution index of the corresponding sub-band energy matrix within the periodic sequence window to obtain a frequency band energy normalization index; and filtering the feature vectors corresponding to the periodic feature vector set based on the window-comprehensive singular value vector and the frequency band energy normalization index of each periodic sequence window to obtain a periodic feature subset. A scene transfer is performed using a generative adversarial network based on the periodic feature subset to obtain a transfer feature set, which includes real scene samples and generated samples. The classifier is trained using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, which is then used for arc fault diagnosis.

2. The method according to claim 1, characterized in that, The frequency band energy distribution index is calculated for the corresponding sub-band energy matrix within the periodic sequence window to obtain the frequency band energy normalization index, including: Based on the corresponding sub-band energy matrix within the periodic sequence window, frequency band energy statistics are performed to obtain the average energy distribution vector. The frequency band energy normalization index is calculated based on the energy distribution of each frequency band within the period and the average energy distribution vector.

3. The method according to claim 1, characterized in that, Based on the window-comprehensive singular value vector of each periodic sequence window and the frequency band energy normalization index, the feature vectors corresponding to the periodic feature vector set are filtered to obtain a periodic feature subset, including: The stability of the window-integrated singular value vector and the frequency band energy normalization index of each periodic sequence window is analyzed by KL divergence analysis to obtain the stability analysis results of the periodic sequence window. The stability analysis results include stable arc periodic behavior segments and unstable arc periodic behavior segments. If the stability analysis result of the periodic sequence window is a stable arc periodic behavior segment, then the feature vector corresponding to the periodic feature vector set is selected as the periodic feature. The periodic feature subset is constructed by constructing the periodic features corresponding to the periodic sequence windows of the multiple stable arc periodic behavior segments; The KL divergence analysis corresponds to the following steps: If the window's integrated singular value vector satisfies the condition of periodic change trend, and the frequency band energy normalization index satisfies the condition of fixed frequency band energy concentration trend, then the corresponding periodic sequence window is determined to be a stable arc periodic behavior segment. If the window's integrated singular value vector does not meet the periodic change trend condition, or the frequency band energy normalization index does not meet the fixed frequency band energy concentration trend condition, then the corresponding periodic sequence window is determined to be an unstable arc periodic behavior segment.

4. The method according to claim 1, characterized in that, A scene transfer is performed using a generative adversarial network based on the periodic feature subset, resulting in a transfer feature set, including: Based on minimizing the generator loss objective, feature mapping is performed on the periodic feature subset to generate target scene features; the generator loss includes adversarial loss and structure preservation loss; the structure preservation loss is used to constrain the target scene samples to retain the arc feature structure in the periodic feature subset; Based on maximizing the discriminator loss target, the probability of authenticity of real samples and features of the target scene is determined to obtain the sample type; the sample type includes real samples and generated samples. When the generated target scene features are the real samples, the target scene features and the real samples of the target scene are determined as a transfer feature set; The generator loss function can be obtained using the following formula: in, It is a periodic feature subset; A single periodic feature within a subset of periodic features; The generated target scene features; This represents the probability of the discriminator identifying the authenticity of the target scene features. To counteract the loss; Weighted hyperparameters for preserving the loss in the structure; To preserve the structure loss function, This represents the transformation function used for structure extraction; The discriminator loss function can be obtained using the following formula: in, The set of real samples for the target scene; A single real sample within the set of real samples in the target scene; This represents the probability that the discriminator classifies a real sample as genuine. The target scene feature set generated by the generator based on a periodic feature subset; The loss term is intended to help the discriminator correctly identify real samples; The goal is to enable the discriminator to correctly identify the loss term of the generated samples.

5. The method according to claim 4, characterized in that, The classifier is trained using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, including: The periodic features in the periodic feature subset are labeled with arc fault types to obtain a labeled training set; the labeled data in the labeled training set includes periodic features and their corresponding arc fault types; the arc fault types include series arc faults, parallel arc faults, and normal. Based on the adversarial training strategy and the dynamic reweighting mechanism, the classifier's model parameters are adjusted by combining the labeled training set and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model. The cross-scenario AC arc fault diagnosis model is used to output the arc fault type according to the arc signal of the scenario. The adversarial training strategy is used to enable the classifier to have discriminative adaptation ability to different sample domains. The dynamic reweighting mechanism is used to give high confidence samples high training weights.

6. A cross-scenario AC arc fault diagnosis system, characterized in that, The system includes: The data acquisition module is used to acquire AC bus current signals and bus voltage signals under arc fault and non-fault conditions in multiple preset classic scenarios to obtain the original scenario signals. The preset classic scenarios cover different load types, power supply environments and equipment models. The feature extraction module is used to perform wavelet packet decomposition and singular value decomposition on the original scene signal to obtain a periodic feature vector set, including: periodically segmenting the original scene signal to obtain a periodic segment sequence, the periodic segment sequence including multiple periodic segments; performing wavelet packet decomposition on the periodic segments to obtain sub-frequency band coefficient groups, the sub-frequency band coefficient groups including multiple sub-frequency band nodes, the sub-frequency band nodes including signal coefficients within a corresponding local frequency range; assembling the sub-frequency band coefficient groups corresponding to each periodic segment into a two-dimensional matrix to obtain a sub-frequency band energy matrix; performing singular value decomposition on the sub-frequency band energy matrix and selecting singular values ​​of a set order to obtain a time-frequency feature matrix; and constructing a periodic feature vector set based on the time-frequency feature matrices corresponding to all periodic segments in the periodic segment sequence. A periodic feature filtering module is used to filter the periodic feature vector set to satisfy the periodic arc structure characteristics to obtain a periodic feature subset. This includes: constructing a multi-period input using a sliding window on the periodic segment sequence to obtain a periodic sequence window; stacking the corresponding sub-band energy matrices within the periodic sequence window to obtain a three-dimensional tensor; expanding the three-dimensional tensor into a two-dimensional matrix along the periodic dimension and performing progressive singular value decomposition to obtain a window-comprehensive singular value vector; calculating the frequency band energy distribution index of the corresponding sub-band energy matrix within the periodic sequence window to obtain a frequency band energy normalization index; and filtering the feature vectors corresponding to the periodic feature vector set based on the window-comprehensive singular value vector and the frequency band energy normalization index of each periodic sequence window to obtain a periodic feature subset. The training set construction module is used to perform scene transfer based on the periodic feature subset using a generative adversarial network to obtain a transfer feature set, which includes real scene samples and generated samples; The model training module is used to train the classifier using the periodic feature subset and the transfer feature set to obtain a cross-scenario AC arc fault diagnosis model, and to use the cross-scenario AC arc fault diagnosis model to diagnose arc faults.

7. 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 according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images