An arc fault diagnosis method, system, device and storage medium

Abstract

The application discloses an arc fault diagnosis method, system, device and storage medium, aiming at the alternating current bus current in the multi-load scene such as industrial production line, commercial building power distribution system, carries out current signal collection and standardization processing, and generates a current signal matrix. The current signal matrix is separated to obtain an arc core characteristic component and a load disturbance component, and a cross-scene common feature is extracted based on the arc core characteristic component and the load disturbance component; the arc core characteristic component and the cross-scene common feature are subjected to multi-domain fusion, and a multi-domain arc feature space is constructed. Based on the multi-domain arc feature space, a fault diagnosis classifier is trained, an arc fault diagnosis model is generated, and arc fault diagnosis is carried out using the arc fault diagnosis model. In this way, the adaptability of arc fault diagnosis in the cross-load scene is improved, and the diagnosis accuracy of arc fault diagnosis in the cross-load scene is improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to an arc fault diagnosis method, system, device and storage medium. Background Technology

[0002] With the rapid development of smart grid technology, arc fault diagnosis is playing an increasingly crucial role in power system safety protection. As the most hidden and frequent safety hazard in power lines, the accurate diagnosis of arc faults directly affects power supply reliability and equipment safety. Especially in complex multi-load scenarios (such as industrial production lines and commercial building power distribution systems), the cross-scenario generalization ability of arc characteristics has become a core challenge in ensuring diagnostic accuracy.

[0003] However, existing arc fault diagnosis technologies have the following problems: load noise and arc characteristics cannot be adaptively decoupled, resulting in a high false alarm rate in nonlinear scenarios; feature extraction lacks a dynamic adaptation mechanism across scenarios, causing a sharp drop in diagnostic accuracy during load switching. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an arc fault diagnosis method, system, device and storage medium to improve the diagnostic accuracy of arc faults in cross-load scenarios.

[0005] In one aspect of the present invention, an arc fault diagnosis method is provided. The method includes the following steps: standardizing the collected AC bus currents from multiple load scenarios to generate a current signal matrix; separating the current signal matrix to obtain arc core feature components and load disturbance components; extracting cross-scenario common features based on the arc core feature components and the load disturbance components; performing multi-domain fusion on the arc core feature components and the cross-scenario common features to construct a multi-domain arc feature space; training a fault diagnosis classifier based on the multi-domain arc feature space to generate an arc fault diagnosis model; and using the arc fault diagnosis model to perform arc fault diagnosis.

[0006] In another aspect of the present invention, an arc fault diagnosis system is provided. The system includes: an information acquisition and standardization module for standardizing the acquired AC bus currents from multiple load scenarios to generate a current signal matrix; a feature decoupling module for separating the current signal matrix to obtain arc core feature components and load disturbance components; a domain adaptation processing module for extracting cross-scenario common features based on the arc core feature components and the load disturbance components; a multi-domain feature fusion module for multi-domain fusion of the arc core feature components and the cross-scenario common features to construct a multi-domain arc feature space; and a model training module for training a fault diagnosis classifier based on the multi-domain arc feature space to generate an arc fault diagnosis model, and using the arc fault diagnosis model to perform arc fault diagnosis.

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

[0009] The beneficial effects of this invention are as follows: For AC bus currents in multi-load scenarios such as industrial production lines and commercial building power distribution systems, current signals are acquired and standardized to generate a current signal matrix. The current signal matrix is ​​separated to obtain the arc core feature component and the load disturbance component. Based on these components, common features across scenarios are extracted. The arc core feature component and the common features across scenarios are then fused across multiple domains to construct a multi-domain arc feature space. A fault diagnosis classifier is trained based on this multi-domain arc feature space to generate an arc fault diagnosis model, which is then used for arc fault diagnosis. This approach increases the adaptability of arc fault diagnosis in multi-load scenarios and improves its diagnostic accuracy. Attached Figure Description

[0010] Figure 1 This is a flowchart of an arc fault diagnosis method according to an embodiment of the present invention;

[0011] Figure 2 This is a comparison diagram of the waveforms after the current signal has been standardized according to an embodiment of the present invention.

[0012] Figure 3 This is a singular value decomposition component diagram of the current signal according to an embodiment of the present invention;

[0013] Figure 4This is a comparison chart of the accuracy of different arc fault diagnosis algorithms according to embodiments of the present invention;

[0014] Figure 5 This is another flowchart of the arc fault diagnosis method according to an embodiment of the present invention;

[0015] Figure 6 This is a structural diagram of the arc fault diagnosis system according to an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] Before detailing the embodiments of this application, some related concepts will first be explained:

[0018] 1. Phase synchronization sampling technology: The current signals of different load branches are sampled synchronously using a unified clock reference to ensure that the signals of each channel are consistent in phase and to eliminate phase deviation caused by asynchronous sampling.

[0019] 2. Adaptive filtering algorithm: An intelligent filtering method that automatically adjusts parameters based on environmental changes, widely used in signal processing, communication and image denoising.

[0020] 3. Hankel matrix: A special matrix in which all elements on each subdiagonal are equal. This matrix describes the dynamic characteristics of a linear system in a block matrix form. In time series prediction models, it is used to construct trajectory matrices to achieve data decomposition and trend extraction.

[0021] 4. Singular value decomposition: An important matrix decomposition method in linear algebra;

[0022] 5. Variational Mode Decomposition Algorithm: A signal decomposition and estimation method. In the process of obtaining the decomposed components, this method determines the frequency center and bandwidth of each component by iteratively searching for the optimal solution of the variational model, thereby adaptively realizing the frequency domain partitioning of the signal and the effective separation of each component.

[0023] 6. Adaptive Mode Decomposition: A signal processing technique used to decompose complex signals into several intrinsic mode functions. This method automatically adjusts parameters through an optimization process, thereby improving the decomposition effect.

[0024] 7. Time-frequency joint wavelet envelope analysis: A signal processing method that combines time and frequency analysis, mainly used to process non-stationary signals. It reveals the frequency variation characteristics of the signal in different time periods through multi-scale decomposition, and realizes time-frequency joint analysis.

[0025] 8. Multiscale Entropy (MSE): Constructs multiscale time series through coarse-grained processing and uses sample entropy to quantify signal complexity;

[0026] 9. Attention Weighting Mechanism: A neural network technology that simulates human selective attention. It effectively solves the problems of information overload and long-distance dependence by dynamically allocating computing resources to focus on key information.

[0027] 10. Wavelet basis: A set of functions generated by scaling and translation transformations of the mother wavelet, forming the mathematical basis of signal decomposition;

[0028] 11. Wavelet filtering: Decompose the signal into wavelet components of different frequencies, process each component independently through thresholding or filter banks (such as noise removal), and finally reconstruct the signal through inverse transform.

[0029] 12. Instantaneous envelope extraction: A signal processing technique used to analyze the intensity change trend of non-stationary signals (such as current, voltage or vibration signals). It is mainly used to capture the energy envelope (i.e., outer contour) of the signal on the time axis to extract the modulation characteristics or instantaneous amplitude of the signal.

[0030] 13. Tensor Outer Product Fusion Mechanism: A multi-dimensional feature fusion method that combines features from multiple different modalities, domains or sources in a higher-order space through the tensor outer product operation, preserving the interaction relationship between each feature dimension, thereby constructing a richer and more detailed joint feature representation.

[0031] 14. Law of Conservation of Arc Energy: During the arc discharge process, the electrical energy input to the system will be released and conserved in various forms, including heat energy, light energy, electromagnetic energy, sound energy and plasma energy. These energy forms can be converted into each other, but the total energy remains conserved within the closed system.

[0032] In existing technologies, with the rapid development of smart grid technology, arc fault diagnosis plays an increasingly crucial role in power system safety protection. As the most hidden and frequent safety hazard in power lines, the accurate diagnosis of arc faults directly affects power supply reliability and equipment safety. However, arc fault diagnosis in multi-load scenarios such as industrial production lines and commercial building power distribution systems lacks a dynamic adaptation mechanism across different scenarios, resulting in a sharp drop in diagnostic accuracy during load switching.

[0033] To solve at least the above-mentioned technical problems, please refer to Figure 1 This invention provides a method for diagnosing electric arc faults, comprising the following steps:

[0034] S101. The collected AC bus currents under multiple load scenarios are standardized to generate a current signal matrix. The current signal matrix is ​​then separated to obtain the arc core feature component and the load disturbance component.

[0035] S102. Extract cross-scene common features based on the core feature components of the electric arc and the load disturbance components;

[0036] S103. Perform multi-domain fusion of the core feature components of the electric arc and the common features across scenarios to construct a multi-domain electric arc feature space;

[0037] S104. Train a fault diagnosis classifier based on the multi-domain arc feature space, generate an arc fault diagnosis model, and use the arc fault diagnosis model to perform arc fault diagnosis.

[0038] Among them, the separated current signal matrix yields the core arc feature component, which refers to the core arc fault signal extracted after physical constraint enhancement, reflecting the time-domain dynamic characteristics of arc discharge (such as current transient response); the separated load disturbance component characterizes the oscillation interference caused by load operation. The cross-scenario common features extracted from the core arc feature component and the load disturbance component are stable features extracted through domain adaptation processing that do not change with the load scenario; the multi-domain arc feature space obtained by fusing the core arc feature component and the cross-scenario common features is a high-dimensional space that integrates time-domain and frequency-domain features, used to improve the generalization ability of the diagnostic model.

[0039] As described above, for AC bus currents in multi-load scenarios such as industrial production lines and commercial building power distribution systems, current signals are acquired and standardized to generate a current signal matrix. The current signal matrix is ​​then separated to obtain the arc core feature component and the load disturbance component. Based on these components, the consistency characteristics of the arc and load across different scenarios are first obtained, and then common features across scenarios are generated. Combining the arc core feature component separated from the current signal matrix with the common features across scenarios, this space simultaneously preserves the temporal dynamic characteristics of the arc and the common features across scenarios, thus constructing a multi-domain arc feature space.

[0040] Subsequently, the feature data in the multi-domain arc feature space were divided into training, validation, and test sets. The training set was used to learn the model parameters, the validation set was used to adjust the model hyperparameters, and the test set was used to evaluate the model's generalization ability. A backpropagation neural network was chosen as the basic classifier architecture, the number of hidden layer nodes was set, and the network weights and biases were initialized. The features from the training set were input into the neural network, and the output was obtained through forward propagation. The error between the predicted value and the true label was calculated based on the loss function, and the network parameters were updated again through the backpropagation algorithm. This process was repeated until the loss function converged or the maximum number of iterations was reached. During training, the model performance was evaluated in real time using the validation set, and hyperparameters such as the learning rate and number of iterations were adjusted to avoid overfitting. The performance of the trained fault diagnosis classifier was validated using the test set, evaluating its accuracy, recall, and other metrics for arc fault detection under cross-scenario conditions, generating an arc fault diagnosis model with cross-scenario diagnostic capabilities.

[0041] In this way, compared with the existing technology, the load disturbance component decoupling mechanism in step S101 solves the problem that load noise and arc characteristics cannot be adaptively separated, reducing the false alarm rate in nonlinear scenarios; the cross-scenario common feature extraction and multi-domain fusion mechanism in steps S102 and S103 enhances the dynamic adaptability of features during load switching, avoids a sharp drop in diagnostic accuracy, increases the adaptability of arc fault diagnosis in cross-load scenarios, and improves the diagnostic accuracy of arc fault diagnosis in cross-load scenarios.

[0042] Furthermore, the collected AC bus currents from multiple load scenarios are standardized to generate a current signal matrix, including:

[0043] Phase-synchronous sampling of AC bus current in multi-load scenarios yields the original current signal sequence;

[0044] By suppressing load fluctuations in the original current signal sequence, a standardized current signal sequence is obtained;

[0045] The standardized current signal sequence is reconstructed into a matrix from the time and frequency dimensions to obtain the current signal matrix.

[0046] As described above, for AC bus currents in multi-load scenarios such as industrial production lines and commercial building power distribution systems, phase synchronization sampling technology is used for signal acquisition to ensure phase consistency of current signals from different loads and locations, obtaining the original current signal sequence. Then, dynamic load fluctuation suppression processing is applied to the original current signal sequence, eliminating interference fluctuations caused by load switching, start-up, and shutdown through filtering and denoising techniques, resulting in a standardized current signal sequence. Next, the standardized current signal sequence is reconstructed using time-frequency matrixing, analyzing and organizing the signal from both time and frequency dimensions to construct a current signal matrix that includes the signal's time-domain dynamic characteristics and frequency-domain feature distribution. This method ensures phase consistency of the acquired current signals from different loads and locations, and the standardized current signal sequence eliminates interference fluctuations caused by load switching, start-up, and shutdown, improving the accuracy of current signal acquisition.

[0047] Furthermore, the current signal matrix is ​​separated to obtain the arc core feature component and the load disturbance component, including:

[0048] The current signal matrix is ​​reconstructed using the Hankel matrix to generate the trajectory matrix;

[0049] The trajectory matrix is ​​subjected to hierarchical singular value decomposition to obtain the dominant mode component and the residual mode component.

[0050] Decouple the load disturbance characteristics of the dominant mode component to generate a load disturbance component;

[0051] Enhance the arc physical constraints of the residual mode components to generate arc core feature components.

[0052] As described above, dynamic Hankel matrix reconstruction is performed on the current signal matrix. Based on the time-series characteristics of the current signal, the one-dimensional signal is rearranged into a multi-dimensional matrix structure according to the sliding window rule, constructing a trajectory matrix that can characterize the dynamic trajectory of the signal. Hierarchical singular value decomposition is then applied to the trajectory matrix. Based on the energy proportion of singular values, the matrix is ​​decomposed into dominant mode components containing the main energy components and residual mode components carrying secondary detailed information. The dominant mode components mainly contain load-related features, while the residual mode components retain more arc characteristics. Load disturbance features are decoupled from the dominant mode components to obtain load disturbance components. Arc physical constraint enhancement processing is applied to the residual mode components. Through filtering, weighting, and other processing, components that conform to arc fault characteristics are strengthened, and noise interference is suppressed, generating arc core feature components containing the core characteristics of the arc. In this way, the efficiency and accuracy of decomposing arc and load features from the current signal are improved.

[0053] Further, the load disturbance characteristics of the dominant mode component are decoupled to generate a load disturbance component, including:

[0054] The dominant mode component is separated into load-specific frequency band signals by variational mode decomposition.

[0055] The oscillation mode of the load-specific frequency band signal is extracted to obtain the load disturbance component.

[0056] As described above, based on the variational mode decomposition algorithm, the dominant mode component is decomposed into multiple independent load-specific frequency band signals in the frequency domain. The power spectral density distribution of each frequency band signal is obtained through spectral analysis and matched against a pre-established load oscillation mode feature library. Load feature components with periodic oscillation characteristics are extracted, and any non-load-related components that may be present are removed to generate a load disturbance component that characterizes the disturbance characteristics of the load during normal operation. This component includes interference features such as current fluctuations caused by the load operation. This method improves the accuracy of the decomposed load signal, resulting in a more stable signal.

[0057] Furthermore, based on the core feature components of the electric arc and the load disturbance components, cross-scene common features are extracted, including:

[0058] Adaptive mode decomposition is performed on the core feature components of the electric arc to generate an intrinsic mode function sequence that is independent of the scene. Time-frequency joint wavelet envelope analysis is performed on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors.

[0059] Multi-scale entropy analysis is performed on the load disturbance components to calculate the common energy spectrum of the scene;

[0060] By fusing the anti-aliasing time-frequency feature vector with the scene common energy spectrum, cross-scene common features are generated.

[0061] As described above, adaptive mode decomposition is performed on the core feature components of the electric arc. By setting a convergence criterion, the decomposition is performed layer by layer, generating a sequence of intrinsic mode functions (IMFs) independent of the scene. Each IMF corresponds to a signal feature component at a different time scale. Time-frequency joint wavelet envelope analysis is performed on the IMF sequence to obtain an anti-aliasing time-frequency feature vector. Subsequently, multi-scale entropy analysis is performed on the load disturbance components to calculate the scene commonality (invariance) energy spectrum reflecting the commonalities of different scenes. An attention-weighted mechanism is used to fuse the anti-aliasing time-frequency feature vector and the scene commonality (invariance) energy spectrum, highlighting the consistency of key features across scenes and generating cross-scene invariant features. In this way, the anti-aliasing time-frequency feature vector and the scene invariant energy spectrum are weighted and fused, strengthening the cross-scene consistency of the core features of the electric arc, suppressing interference caused by scene differences, generating cross-scene invariant features that can adapt to different load environments, and improving the efficiency and accuracy of cross-scene common features.

[0062] Further, time-frequency joint wavelet envelope analysis is performed on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors, including:

[0063] An adaptive wavelet basis is constructed from the intrinsic mode function sequence to generate a wavelet filter bank that is suitable for the scene.

[0064] The wavelet filter bank is used to perform a dual-density wavelet transform on each intrinsic mode function in the intrinsic mode function sequence to extract the scaling coefficients and detail coefficients in the time-frequency domain.

[0065] The scale coefficients are subjected to instantaneous envelope extraction processing to generate arc feature envelope trajectories;

[0066] By fusing the detail coefficients with the arc feature envelope trajectory, an anti-aliasing time-frequency feature matrix is ​​constructed.

[0067] The anti-aliasing time-frequency feature matrix is ​​subjected to entropy stability quantization to generate an anti-aliasing time-frequency feature vector.

[0068] As described above, an adaptive wavelet basis is constructed for the intrinsic mode function (EMF) sequence. By analyzing the distribution characteristics and nonlinearity of different frequency components in the sequence, wavelet functions adapted to the signal characteristics of the current scene are selected or constructed from a pre-defined wavelet basis library, thereby generating a scene-adaptive wavelet filter bank containing multiple sets of different cutoff frequencies. Then, a dual-density wavelet transform is performed on each EMF in the EMF sequence using the wavelet filter bank. By processing two wavelet transforms with different phase shifts in parallel, scaling coefficients representing the low-frequency trend of the signal and detail coefficients reflecting high-frequency details are extracted in the time-frequency domain. A nonlinear transformation is performed on the scaling coefficients to extract their amplitude envelope, generating an arc feature envelope trajectory that reflects the energy fluctuations and periodic characteristics during arc discharge. The arc feature envelope trajectory and detail coefficients are fused to construct an anti-aliasing time-frequency feature matrix containing multi-scale information in the time-frequency domain. Entropy stability quantization is then performed to obtain an anti-frequency aliasing time-frequency feature vector. This method is particularly suitable for the refined analysis of nonlinear and non-stationary signals, improving the accuracy of nonlinear scene diagnosis.

[0069] Furthermore, the core feature components of the electric arc are fused with the common features across different scenarios to construct a multi-domain electric arc feature space, including:

[0070] The temporal physical properties of the core feature components of the electric arc are analyzed to generate the time-varying feature tensor of the electric arc;

[0071] Enhance the frequency domain invariance of the common features across scenarios to generate a frequency domain robust feature tensor;

[0072] The arc time-varying feature tensor and the frequency domain robust feature tensor are fused to construct a higher-order feature space;

[0073] The higher-order feature space is subjected to physical constraint dimensionality reduction to generate a compressed feature subspace;

[0074] The stability of the compressed feature subspace is optimized to construct a multi-domain arc feature space.

[0075] The "multi-domain" in the multi-domain arc feature space refers to the dual dimensions of the time and frequency domains: the time domain analyzes the physical properties of the core arc features (such as zero-intersection distortion), while the frequency domain enhances the robustness of features across different scenarios (such as stable frequency components). The construction principle is to fuse the time-domain feature tensor and the frequency-domain feature tensor through tensor outer product to form a higher-order space, and then generate a compressed subspace through physical constraint dimensionality reduction (such as optimization based on the arc energy conservation law) to ensure the stability of features under different load scenarios.

[0076] As described above, the core feature components of the electric arc undergo time-domain physical attribute analysis. Attributes reflecting the arc's time-domain dynamic process are extracted in a structured manner and transformed into a time-varying feature tensor with multi-dimensional time-series information. Frequency-domain commonality (invariance) enhancement is applied to cross-scene common (invariant) features, extracting the distribution of features across different frequency components. Emphasis is placed on strengthening stable frequency-domain components that do not change with load scenarios, generating a frequency-domain robust feature tensor. Subsequently, the arc time-varying feature tensor and the frequency-domain robust feature tensor are fused using a tensor outer product fusion mechanism to construct a high-order feature space containing multi-dimensional information such as time, frequency, and energy. This space simultaneously preserves the arc's time-domain dynamic characteristics and frequency-domain invariance features. Finally, physical constraint dimensionality reduction is applied to the high-order feature space, removing redundant information based on the physical principles of arc faults to generate a compressed feature subspace that retains key physical features. Based on the law of conservation of arc energy, the stability of the compressed feature subspace is optimized to ensure the consistency of the feature space under different scenarios. This results in the construction of a multi-domain arc feature space that combines time-domain dynamic characteristics, frequency-domain robustness, and physical stability, providing a reliable feature basis for subsequent fault diagnosis.

[0077] In the following, the technical solutions according to this disclosure will be described with reference to specific embodiments and in conjunction with the accompanying drawings.

[0078] Please refer to Figure 5 One embodiment of the present invention provides an arc fault diagnosis method, comprising:

[0079] S201. Acquire and standardize the AC bus current signal in multi-load scenarios to generate a current signal matrix.

[0080] Specifically, phase synchronization sampling technology is used to acquire AC bus currents in multi-load scenarios such as industrial production lines and commercial building power distribution systems. By synchronously sampling the current signals of different load branches using a unified clock reference, the phase consistency of signals in each channel is ensured, eliminating phase deviations caused by asynchronous sampling. This results in obtaining a raw current signal sequence that accurately reflects the collaborative operation of multiple loads.

[0081] Dynamic load fluctuation suppression processing is applied to the original current signal sequence. An adaptive filtering algorithm is used to process the signal, identifying and suppressing abnormal fluctuations caused by dynamic processes such as load start-up, shutdown, and switching, while retaining characteristic signals related to arc faults. The signal is standardized by setting a dynamic threshold to eliminate the influence of current amplitude differences under different load conditions, generating a standardized current signal sequence with uniform amplitude and suppressed noise, such as... Figure 2 As shown, a waveform comparison diagram of the original current signal and the standardized signal is displayed after dynamic load fluctuation suppression processing. That is, the waveform with larger amplitude corresponds to the original current signal, and the waveform with smaller amplitude corresponds to the standardized signal.

[0082] A time-frequency matrix reconstruction is performed on the standardized current signal sequence. Time-frequency analysis methods such as short-time Fourier transform or wavelet transform are used to convert the time-domain signal into a time-frequency domain representation, obtaining the frequency component distribution of the signal at different time points. The time-frequency domain information is then organized into a matrix along the time-frequency dimension to construct a current signal matrix containing the signal's time-domain dynamic characteristics and frequency-domain feature distribution. This transforms the one-dimensional signal into a two-dimensional structure, preserving the signal's dynamic changes in the time dimension and revealing the frequency domain distribution in the frequency dimension. This provides a structured data foundation for subsequent feature decoupling and analysis, making feature decoupling more efficient.

[0083] S202. Perform hierarchical singular value decomposition on the current signal matrix to separate the arc core characteristic component and the load disturbance component.

[0084] Specifically, such as Figure 3The diagram illustrates the singular value separation results of the current signal matrix, comprising "DC component + AC component + noise component". The AC component primarily characterizes the periodic oscillation interference caused by normal load operation (i.e., the dominant mode, such as motor start-stop and inverter harmonics). The secondary part of the AC component plus the transient noise part of the noise component contains high-frequency transient characteristics of arc discharge (i.e., residual modes, such as spark pulses and zero-crossing distortion), which, despite their weak energy, carry crucial fault information. Dynamic Hankel matrix reconstruction is performed on the current signal matrix, constructing a trajectory matrix that conforms to the signal characteristics to characterize the dynamic trajectory of the current signal. Hierarchical singular value decomposition is then applied to the trajectory matrix, extracting the dominant mode component and residual mode component based on the magnitude of the singular values. The dominant mode component mainly contains load-related features, while the residual mode component retains more arc characteristics.

[0085] Subsequently, load disturbance characteristics are decoupled from the dominant mode components, and load-specific frequency band signals are separated through variational mode constraint decomposition and other methods, from which load disturbance components are extracted. Arc physical constraint enhancement processing is then applied to the residual mode components to strengthen features related to arc faults, thereby generating the core arc feature components.

[0086] This involves dynamic Hankel matrix reconstruction of the current signal matrix. Based on the time-series characteristics of the current signal, the one-dimensional signal is rearranged into a multi-dimensional matrix structure according to a sliding window rule, constructing a trajectory matrix that can characterize the dynamic trajectory of the signal. By adaptively adjusting the window length and overlap, the trajectory matrix is ​​ensured to fully preserve the temporal correlation and spatial topological features of the current signal.

[0087] Hierarchical singular value decomposition (SVD) is performed on the trajectory matrix. The eigenvalues ​​of the trajectory matrix are decomposed using the SVD algorithm, and the matrix is ​​decomposed into dominant mode components containing the main energy components and residual mode components carrying secondary detail information, based on the energy proportions of the singular values. The number of layers is determined by setting an energy threshold, achieving multi-scale separation of signal features.

[0088] Load disturbance characteristics are decoupled from the dominant mode components. Using methods such as variational mode decomposition, the dominant mode components are decomposed into multiple independent load-specific frequency band signals in the frequency domain. By comparing the spectrum with the load feature library, the oscillation modes that match the operating characteristics of common loads are extracted to generate load disturbance components that characterize the disturbances during normal load operation.

[0089] Variational mode constraint decomposition is performed on the dominant mode components. Based on the variational mode decomposition algorithm, the dominant mode components are decomposed into a preset number of mode components in the frequency domain. By setting frequency band constraints and combining the typical frequency domain characteristics of the load during operation, the frequency band signal matching common load characteristics is separated to generate a load-specific frequency band signal. This process ensures that the frequency bands of each mode component do not overlap and completely cover the frequency range of the dominant mode by iteratively optimizing the objective function of the variational model.

[0090] Oscillation modes are extracted from the load-specific frequency band signal. The power spectral density distribution of each frequency band signal is obtained through spectrum analysis and matched with a pre-established load oscillation mode feature library. Using methods such as peak detection and pattern recognition, load feature components with periodic oscillation characteristics are extracted. Non-load-related components that may be included are removed to generate a load disturbance component that can characterize the disturbance characteristics of the load during normal operation. This component includes interference characteristics such as current fluctuations caused by the load operation.

[0091] The residual mode components are enhanced by arc physical constraints. Combining the physical characteristics of arc discharge (such as high-temperature ionization and nonlinear volt-ampere characteristics), physical rules such as time-domain waveform constraints and frequency-domain characteristic constraints are applied to the residual mode components. Through filtering and weighting, components conforming to arc fault characteristics are strengthened, noise interference is suppressed, and arc core feature components containing the core arc characteristics are generated. After variational mode decomposition, the load disturbance of the dominant mode components is separated; the residual mode components are strengthened by arc physical constraints (such as nonlinear volt-ampere characteristic enhancement) to conform to fault characteristics and suppress noise, thus obtaining the arc core characteristics. This step utilizes the energy concentration characteristics of arc discharge, making it easier for the residual components to capture fault signals.

[0092] S203. Based on the core feature components of the electric arc and the load disturbance components, cross-scenario domain adaptation processing is performed to generate cross-scenario invariant features.

[0093] In this embodiment, the following formula is used to perform adaptive mode decomposition on the core feature components of the electric arc, generating a scene-independent sequence of intrinsic mode functions:

[0094]

[0095] In the formula, λ n The adaptive convergence criterion for mode decomposition is represented by T, where T represents the observation time window length of the arc signal, and S represents the mode decomposition. arc (t) represents the original arc core characteristic signal, IMF j (t) represents the j-th intrinsic mode function, and n represents the total number of mode functions obtained by the current decomposition;

[0096] Then, time-frequency joint wavelet envelope analysis was performed on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors; multi-scale spectral entropy analysis was performed on the load disturbance components to calculate the scene invariant energy spectrum; through an attention-weighted feature fusion mechanism, the anti-aliasing time-frequency feature vectors and the scene invariant energy spectrum were fused to generate cross-scene invariant features.

[0097] Specifically, adaptive mode decomposition is performed on the core feature components of the electric arc. Using a preset adaptive convergence criterion for mode decomposition as a threshold, the original core feature signal of the electric arc is decomposed layer by layer within the observation time window of the electric arc signal. By iteratively calculating the difference between the local extrema and the mean envelope until the convergence condition is met, a scene-independent sequence of intrinsic mode functions is generated, with each intrinsic mode function corresponding to a signal feature component at a different time scale.

[0098] Since diagnosing arc faults across load scenarios is difficult, it is necessary to extract common features across scenarios through joint time-frequency analysis. Joint time-frequency wavelet envelope analysis of the intrinsic mode function (EMF) sequence can increase the adaptability of arc fault diagnosis across load scenarios and improve diagnostic accuracy. In implementation, a scenario-adaptive wavelet filter bank is adaptively constructed based on the sequence features. This filter bank is used to perform a dual-density wavelet transform on each EMF to separate the scaling coefficients and detail coefficients in the time-frequency domain. The scaling coefficients are then processed using methods such as Hilbert transform to extract the instantaneous envelope, generating an envelope trajectory characterizing the dynamic changes of arc features. The detail coefficients and the envelope trajectory are then fused to construct an anti-aliasing time-frequency feature matrix containing multi-dimensional information in the time-frequency domain. Entropy stability quantization is then applied to generate an anti-aliasing time-frequency feature vector that resists frequency aliasing.

[0099] Multi-scale entropy analysis is performed on the load disturbance components. Multi-scale entropy analysis extracts stable energy characteristics of load disturbances by quantifying the entropy distribution at different frequency resolutions. This reduces the uncertainty caused by scene differences, enhances the robustness of common features across scenes, and thus improves the adaptability of the diagnostic model during load switching. In implementation, spectral analysis is performed on the load disturbance components at different frequency resolutions, calculating the entropy distribution of the spectrum at each scale. The uncertainty of the signal in the frequency domain is quantified by the entropy values, thereby extracting stable energy distribution characteristics that do not change with the scene, generating a scene-invariant energy spectrum to characterize the common features in load disturbances.

[0100] Two types of features are integrated through an attention-weighted feature fusion mechanism. An attention weight model is constructed, which assigns weights to features based on their importance in cross-scene diagnosis. The anti-aliasing time-frequency feature vector and the scene-invariant energy spectrum are weighted and fused to enhance the cross-scene consistency of the core arc features, suppress interference caused by scene differences, and generate cross-scene invariant features that can adapt to different load environments.

[0101] Specifically, joint time-frequency wavelet envelope analysis is performed on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors, including:

[0102] An adaptive wavelet basis construction process is performed on the intrinsic mode function sequence to generate a wavelet filter bank that is suitable for the scene. The dual-density wavelet transform is performed on each intrinsic mode function through the wavelet filter bank to extract the time-frequency domain scaling coefficients and detail coefficients. The instantaneous envelope extraction process is performed on the scaling coefficients to generate the arc feature envelope trajectory. The detail coefficients and the arc feature envelope trajectory are fused to construct an anti-aliasing time-frequency feature matrix.

[0103] The following formula is used to perform entropy stability quantization on the anti-aliasing time-frequency feature matrix to generate an anti-aliasing time-frequency feature vector:

[0104]

[0105] Among them, S k This represents the stability quantization value of the k-th frequency component. This represents the standard deviation of the k-th frequency component. T1 represents the small constant used to prevent division by zero, and F represents the time window length. kt μ represents the eigenvalue of the k-th frequency at time t. k This represents the mean of the k-th frequency component.

[0106] Specifically, an adaptive wavelet basis is constructed based on the time-frequency characteristics of the intrinsic mode function sequence. By analyzing the distribution characteristics and nonlinearity of different frequency components in the sequence, wavelet functions adapted to the signal characteristics of the current scene are selected or constructed from a preset wavelet basis library. This generates a scene-adaptive wavelet filter bank containing multiple sets of different cutoff frequencies to ensure that the filter bank can effectively match the time-frequency localization requirements of the arc signal.

[0107] A dual-density wavelet transform is performed on each intrinsic mode function using a wavelet filter bank. By processing two wavelet transforms with different phase shifts in parallel, scaling coefficients, which characterize the low-frequency trend of the signal, and detail coefficients, which reflect the high-frequency details, are extracted from the time-frequency domain. The scaling coefficients correspond to the overall profile of the signal, while the detail coefficients correspond to the transient characteristics of the signal, thereby achieving time-frequency multi-resolution decomposition of the intrinsic mode functions.

[0108] Instantaneous envelope extraction is performed on the scaling coefficients. Using Hilbert transform or analytic signal method, a nonlinear transformation is applied to the scaling coefficients to extract their amplitude envelopes, generating an arc feature envelope trajectory that intuitively reflects the dynamic changes in arc characteristics. This trajectory can demonstrate the energy fluctuations and periodic characteristics during the arc discharge process.

[0109] Next, the detail coefficients are fused with the arc feature envelope trajectory. The high-frequency feature distribution of the detail coefficients and the energy changes of the envelope trajectory are matrix-reorganized along the time-frequency dimension to construct an anti-aliasing time-frequency feature matrix containing multi-scale information in the time and frequency domains. This matrix simultaneously preserves the signal's detail abrupt changes and energy envelope trends, effectively suppressing frequency aliasing effects.

[0110] Entropy stability quantization is performed on the anti-aliasing time-frequency feature matrix. Based on statistics such as the standard deviation and mean of the frequency components, the feature stability index of each frequency component within the time window is calculated. The features in the matrix are then reduced in dimensionality using the entropy quantization method to remove unstable features affected by scene interference, generating an anti-aliasing time-frequency feature vector that combines time-frequency resolution and anti-aliasing properties, providing reliable feature input for subsequent cross-scene diagnosis.

[0111] In this way, the anti-aliasing time-frequency feature vector and the scene-invariant energy spectrum are weighted and fused to enhance the cross-scene consistency of the core features of the electric arc, suppress interference caused by scene differences, generate cross-scene invariant features that can adapt to different load environments, and improve the efficiency and accuracy of cross-scene common features. Figure 4 As shown, the cross-scene feature extraction algorithm based on asymptotic singular value decomposition-fast Fourier transform (PSVD-FFT) used in the above steps was compared with the accuracy of traditional algorithms (Support Vector Machine (SVM), Principal Component Analysis-Support Vector Machine (PCA-SVM), Empirical Mode Decomposition-Probabilistic Neural Network (EMD-PNN), and Asymptotic Singular Value Decomposition (PSVD)) in mixed load scenarios. The experiment shows that after cross-scene domain adaptation, the extraction accuracy reaches 92.56%, which is more than 15% higher than that of traditional SVM.

[0112] S204. Perform multi-domain fusion processing on the core feature components of the electric arc and the cross-scene invariant features to construct a multi-domain electric arc feature space.

[0113] Specifically, the core feature components of the electric arc are analyzed for their temporal physical properties. Combined with the temporal dynamic characteristics of the arc fault, a tensor representing the time-varying features of the arc is generated. Cross-scene invariant features are enhanced for frequency domain invariance. Through frequency domain analysis, the stability of features under different scenarios is highlighted, generating a frequency domain robust feature tensor.

[0114] Subsequently, a tensor outer product fusion mechanism is used to fuse the arc time-varying feature tensor with the frequency domain robust feature tensor to construct a high-order feature space containing multi-dimensional information. Then, physical constraint dimensionality reduction is applied to the high-order feature space, and redundant information is removed based on the physical principles of arc faults to generate a compressed feature subspace. Stability optimization of the compressed feature subspace is performed based on the arc energy conservation law to ensure consistency of the feature space under different scenarios, thereby constructing a multi-domain arc feature space.

[0115] Specifically, the core characteristic components of the electric arc are analyzed for their time-domain physical properties. By analyzing the dynamic changes of the current signal in the time dimension, such as the transient response when the arc is generated, the periodic fluctuations of the current amplitude, and the zero-intersection distortion, the properties reflecting the time-domain dynamic process of the arc are extracted in a structured manner and transformed into a time-varying characteristic tensor of the arc with multi-dimensional time series information, so as to characterize the evolution law of the arc fault in the time domain.

[0116] Frequency domain invariance enhancement is performed on cross-scenario invariant features. Frequency domain analysis is used to extract the distribution of features across different frequency components, with a focus on strengthening stable frequency components that do not change with load scenarios, such as the energy distribution of characteristic frequency bands in arc discharge. Filtering and normalization are then used to eliminate frequency domain interference caused by scenario differences, generating a robust frequency domain feature tensor that reflects cross-scenario consistency, ensuring the stability of frequency domain features under different load conditions.

[0117] A tensor outer product fusion mechanism is employed to fuse the time-varying feature tensor of the electric arc with the frequency-domain robust feature tensor. Through tensor outer product operations, the time-domain features and frequency-domain features are correlated and combined in a multi-dimensional space to construct a high-order feature space containing multi-dimensional information such as time, frequency, and energy. This space simultaneously preserves the time-domain dynamic characteristics and frequency-domain invariance features of the electric arc, generating a more comprehensive feature representation.

[0118] Subsequently, physical constraint dimensionality reduction is performed on the high-order feature space. Based on the physical principles of arc faults, constraints with clear physical meaning are set, and redundant dimensions and features with low correlation in the high-order feature space are screened and removed. Through dimensionality reduction algorithms such as principal component analysis, a compressed feature subspace that retains key physical features is generated, which reduces computational complexity while avoiding the loss of feature information.

[0119] Stability optimization of the compressed feature subspace is performed based on the law of energy conservation in electric arc discharge. According to the physical law of energy conservation during electric arc discharge, the compressed feature subspace is constrained and optimized to ensure the consistency of energy distribution of features under different scenarios. By adjusting feature weights or introducing energy balance constraints, feature shifts caused by scenario changes are eliminated, constructing a multi-domain electric arc feature space that combines time-domain dynamic characteristics, frequency-domain robustness, and physical stability, providing a reliable feature foundation for subsequent fault diagnosis.

[0120] S205. Train a fault diagnosis classifier based on the multi-domain arc feature space to generate an arc fault diagnosis model.

[0121] Specifically, the feature data in the multi-domain electric arc feature space is divided into a training set, a validation set, and a test set. The training set is used for model parameter learning, the validation set is used for adjusting model hyperparameters, and the test set is used to evaluate the model's generalization ability. A backpropagation neural network is chosen as the basic classifier architecture, and the number of hidden layer nodes is set and the network weights and biases are initialized. The features from the training set are input into the neural network, and the output results are calculated through forward propagation. The error between the predicted value and the true label is calculated according to the loss function, and the network parameters are updated again through the backpropagation algorithm. This process is repeated until the loss function converges or the maximum number of iterations is reached.

[0122] During training, the model performance is evaluated in real time using a validation set, and hyperparameters such as the learning rate and number of iterations are adjusted to avoid overfitting. The trained classifier is validated using a test set, and its accuracy and recall for detecting arc faults under cross-scenario conditions are evaluated, generating an arc fault diagnosis model with cross-scenario diagnostic capabilities.

[0123] The arc fault detection method of this embodiment can achieve the technical effects of improving the diagnostic accuracy of nonlinear scenarios, enhancing the adaptability of cross-load scenarios, and reducing the missed detection rate and response delay of weak arcs during the detection process.

[0124] Please refer to Figure 6 In another embodiment of the present invention, an arc fault detection system 300 is provided, the system 300 comprising:

[0125] The signal acquisition and standardization module 301 is used to acquire and standardize AC bus current signals in multi-load scenarios to generate a current signal matrix.

[0126] The feature decoupling module 302 is used to perform hierarchical singular value decomposition on the current signal matrix to separate the arc core feature component and the load disturbance component.

[0127] The domain adaptation processing module 303 is used to perform cross-scene domain adaptation processing based on the core feature components of the electric arc and the load disturbance components to generate cross-scene invariant features.

[0128] The multi-domain feature fusion module 304 is used to perform multi-domain fusion processing on the core feature components of the electric arc and cross-scene invariant features to construct a multi-domain electric arc feature space.

[0129] The model training module 305 is used to train a fault diagnosis classifier based on the multi-domain arc feature space to generate an arc fault diagnosis model.

[0130] Specifically, the signal acquisition and standardization module 301 uses phase synchronization sampling technology to synchronously acquire AC bus currents under multi-load scenarios such as industrial production lines and commercial buildings, obtaining the original current signal sequence after eliminating phase deviations. Then, through adaptive filtering and dynamic thresholding, fluctuations caused by load start-up and shutdown are suppressed to generate a standardized current signal sequence. Finally, time-frequency analysis is used to reconstruct it into a current signal matrix containing time-frequency information.

[0131] The feature decoupling module 302 performs dynamic Hankel matrix reconstruction on the current signal matrix to generate a matrix representing the signal trajectory. It extracts the dominant mode and residual mode components through hierarchical singular value decomposition, where the dominant mode contains load disturbance features and the residual mode contains arc features. Variational mode constraint decomposition is performed on the dominant mode to separate the load-specific frequency band signal, and then the oscillation mode is extracted to generate the load disturbance component. Arc physical constraints are applied to the residual mode to enhance the arc features and generate the core arc feature component.

[0132] The domain adaptation processing module 303 performs adaptive mode decomposition on the core feature components of the electric arc, generating a scene-independent sequence of intrinsic mode functions (EMFs). A filter bank is constructed using an adaptive wavelet basis, and a dual-density wavelet transform is performed on the EMFs to extract scale and detail coefficients. The envelope of the scale coefficients is extracted and fused with the detail coefficients to construct an anti-aliasing time-frequency feature matrix. This matrix is ​​then quantized using entropy to generate an anti-aliasing time-frequency feature vector. Multi-scale spectral entropy analysis is performed on the load disturbance components to calculate the scene-invariant energy spectrum. The two types of features are fused using an attention-weighted mechanism to generate cross-scene invariant features.

[0133] The multi-domain feature fusion module 304 analyzes the time-domain physical properties of the core features of the electric arc, such as transient current response and zero-intersection distortion, to generate a time-varying feature tensor of the electric arc. It enhances the frequency-domain stability of cross-scene invariant features, generating a frequency-domain robust feature tensor. A higher-order feature space is constructed through tensor outer product fusion, and then dimensionality is reduced based on the physical principles of the electric arc to generate a compressed feature subspace. Stability is optimized based on the law of conservation of energy, constructing a multi-domain electric arc feature space.

[0134] The model training module 305 divides the multi-domain feature space data into training, validation, and test sets, employs a backpropagation neural network architecture, sets the hidden layer nodes, and initializes the parameters. It iterative training is performed through forward and backpropagation, with hyperparameters adjusted using the validation set until the loss function converges. Subsequently, the test set is used to evaluate the model's cross-scenario diagnostic performance, generating an arc fault diagnosis model.

[0135] Domain adaptation processing module 303 is also used for:

[0136] The following formula is used to perform adaptive mode decomposition on the core feature components of the electric arc, generating a scene-independent sequence of intrinsic mode functions:

[0137]

[0138] In the formula, λ n The adaptive convergence criterion for mode decomposition is represented by T, where T represents the observation time window length of the arc signal, and S represents the mode decomposition. arc (t) represents the original arc core characteristic signal, IMF j (t) represents the j-th intrinsic mode function, and n represents the total number of mode functions obtained by the current decomposition;

[0139] Time-frequency joint wavelet envelope analysis was performed on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors;

[0140] Multi-scale spectral entropy analysis of load disturbance components is performed to calculate the scene invariant energy spectrum;

[0141] By using an attention-weighted feature fusion mechanism, the anti-aliasing time-frequency feature vector is fused with the scene-invariant energy spectrum to generate cross-scene invariant features.

[0142] Domain adaptation processing module 303 is also used for:

[0143] Adaptive wavelet basis construction is performed on the intrinsic mode function sequence to generate a wavelet filter bank that is suitable for the scene.

[0144] The time-frequency domain scaling coefficients and detail coefficients are extracted by performing a dual-density wavelet transform on each eigenmode function using a wavelet filter bank.

[0145] Instantaneous envelope extraction is performed on the scaling coefficients to generate the arc feature envelope trajectory.

[0146] By fusing detail coefficients and arc feature envelope trajectories, an anti-aliasing time-frequency feature matrix is ​​constructed.

[0147] The following formula is used to perform entropy stability quantization on the anti-aliasing time-frequency feature matrix to generate an anti-aliasing time-frequency feature vector:

[0148]

[0149] Among them, S k This represents the stability quantization value of the k-th frequency component. This represents the standard deviation of the k-th frequency component. T1 represents the small constant used to prevent division by zero, and F represents the time window length. kt μ represents the eigenvalue of the k-th frequency at time t. k This represents the mean of the k-th frequency component.

[0150] The multi-domain feature fusion module 304 is also used for:

[0151] The core feature components of the electric arc are subjected to time-domain physical property analysis to generate the time-varying feature tensor of the electric arc.

[0152] Frequency domain invariance enhancement processing is applied to cross-scene invariant features to generate a frequency domain robust feature tensor;

[0153] A higher-order feature space is constructed by fusing the arc time-varying feature tensor with the frequency domain robust feature tensor through a tensor outer product fusion mechanism.

[0154] Physical constraint dimensionality reduction is performed on the high-order feature space to generate a compressed feature subspace;

[0155] Based on the law of conservation of arc energy, the stability of the compressed feature subspace is optimized to construct a multi-domain arc feature space.

[0156] Feature decoupling module 302 is also used for:

[0157] Dynamic Hankel matrix reconstruction is performed on the current signal matrix to generate the trajectory matrix;

[0158] Hierarchical singular value decomposition is performed on the trajectory matrix to extract the dominant mode component and the residual mode component;

[0159] The load disturbance characteristics of the dominant mode component are decoupled to generate the load disturbance component.

[0160] The residual mode components are subjected to arc physical constraint enhancement processing to generate arc core feature components.

[0161] Feature decoupling module 302 is also used for:

[0162] Variational mode constraint decomposition is performed on the dominant mode components to separate the load-specific frequency band signal;

[0163] The oscillation mode of the load-specific frequency band signal is extracted and processed to generate load disturbance components.

[0164] The signal acquisition and standardization module 301 is also used for:

[0165] Phase-synchronous sampling processing is performed on the AC bus current in multiple scenarios to obtain the original current signal sequence;

[0166] The original current signal sequence is subjected to dynamic load fluctuation suppression processing to generate a standardized current signal sequence.

[0167] The standardized current signal sequence is reconstructed by time-frequency matrix conversion to generate a current signal matrix.

[0168] In one embodiment, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0169] In one embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0170] 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 method for diagnosing electric arc faults, characterized in that, Including the following steps: The collected AC bus currents under multiple load scenarios are standardized to generate a current signal matrix. The current signal matrix is ​​then separated to obtain the arc core feature component and the load disturbance component. Extracting cross-scene common features based on the arc core feature components and the load disturbance components includes: performing adaptive mode decomposition on the arc core feature components to generate scene-independent intrinsic mode function sequences; performing time-frequency joint wavelet envelope analysis on the intrinsic mode function sequences to extract anti-aliasing time-frequency feature vectors; performing multi-scale entropy analysis on the load disturbance components to calculate scene common energy spectra; and fusing the anti-aliasing time-frequency feature vectors and scene common energy spectra to generate cross-scene common features. The core feature components of the electric arc are fused with the common features across scenarios to construct a multi-domain electric arc feature space. A fault diagnosis classifier is trained based on the multi-domain arc feature space to generate an arc fault diagnosis model, and the arc fault diagnosis model is used to diagnose arc faults.

2. The arc fault diagnosis method according to claim 1, characterized in that, The collected AC bus currents from multiple load scenarios are standardized to generate a current signal matrix, including: Phase-synchronous sampling of AC bus current in multi-load scenarios yields the original current signal sequence; By suppressing load fluctuations in the original current signal sequence, a standardized current signal sequence is obtained; The standardized current signal sequence is reconstructed into a matrix from the time and frequency dimensions to obtain the current signal matrix.

3. The arc fault diagnosis method according to claim 1, characterized in that, Separating the current signal matrix yields the arc core characteristic component and the load disturbance component, including: The current signal matrix is ​​reconstructed using the Hankel matrix to generate the trajectory matrix; The trajectory matrix is ​​subjected to hierarchical singular value decomposition to obtain the dominant mode component and the residual mode component. Decouple the load disturbance characteristics of the dominant mode component to generate a load disturbance component; Enhance the arc physical constraints of the residual mode components to generate arc core feature components.

4. The arc fault diagnosis method according to claim 3, characterized in that, Decouple the load disturbance characteristics of the dominant mode component to generate a load disturbance component, including: The dominant mode component is separated into load-specific frequency band signals by variational mode decomposition. The oscillation mode of the load-specific frequency band signal is extracted to obtain the load disturbance component.

5. The arc fault diagnosis method according to claim 1, characterized in that, Perform time-frequency joint wavelet envelope analysis on the intrinsic mode function sequence to extract anti-aliasing time-frequency feature vectors, including: An adaptive wavelet basis is constructed from the intrinsic mode function sequence to generate a wavelet filter bank that is suitable for the scene. The wavelet filter bank is used to perform a dual-density wavelet transform on each intrinsic mode function in the intrinsic mode function sequence to extract the scaling coefficients and detail coefficients in the time-frequency domain. The scale coefficients are subjected to instantaneous envelope extraction processing to generate arc feature envelope trajectories; By fusing the detail coefficients with the arc feature envelope trajectory, an anti-aliasing time-frequency feature matrix is ​​constructed. The anti-aliasing time-frequency feature matrix is ​​subjected to entropy stability quantization to generate an anti-aliasing time-frequency feature vector.

6. The arc fault diagnosis method according to claim 1, characterized in that, The core feature components of the electric arc are fused with the common features across different scenarios to construct a multi-domain electric arc feature space, including: The temporal physical properties of the core feature components of the electric arc are analyzed to generate the time-varying feature tensor of the electric arc; Enhance the frequency domain invariance of the common features across scenarios to generate a frequency domain robust feature tensor; The arc time-varying feature tensor and the frequency domain robust feature tensor are fused to construct a higher-order feature space; The higher-order feature space is subjected to physical constraint dimensionality reduction to generate a compressed feature subspace; The stability of the compressed feature subspace is optimized to construct a multi-domain arc feature space.

7. An arc fault diagnosis system, characterized in that, Including the following steps: The information acquisition and standardization module is used to standardize the acquired AC bus currents from multiple load scenarios and generate a current signal matrix. The feature decoupling module is used to separate the current signal matrix to obtain the arc core feature component and the load disturbance component; The domain adaptation processing module is used to extract cross-scene common features based on the arc core feature components and the load disturbance components, including: performing adaptive mode decomposition on the arc core feature components to generate scene-independent intrinsic mode function sequences; performing time-frequency joint wavelet envelope analysis on the intrinsic mode function sequences to extract anti-aliasing time-frequency feature vectors; performing multi-scale entropy analysis on the load disturbance components to calculate scene common energy spectra; and fusing the anti-aliasing time-frequency feature vectors and scene common energy spectra to generate cross-scene common features. The multi-domain feature fusion module is used to perform multi-domain fusion of the core feature components of the electric arc and the common features across scenarios to construct a multi-domain electric arc feature space. The model training module is used to train a fault diagnosis classifier based on the multi-domain arc feature space, generate an arc fault diagnosis model, and use the arc fault diagnosis model to perform arc fault diagnosis.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory being configured to store a computer program; the processor being configured to execute the computer program to implement the steps of the arc fault diagnosis method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the arc fault diagnosis method according to any one of claims 1 to 6.

Images