Non-intrusive alternating current series arc fault detection generalization enhancement and strange scene adaptation method, system and device and medium

Abstract

The invention discloses a non-intrusive alternating-current series arc fault detection generalization enhancement and strange scene adaptation method, system and device and a medium, and belongs to the technical field of low-voltage user fault arc detection, and the method comprises the steps: obtaining first scene electrical measurement data, generating a first feature data set, building a processing reference through a key state point, and generating a second feature data set; training a basic diagnosis model based on the multi-dimensional features; generating a fault feature resource library through a section selection algorithm; acquiring normal operation data of a second scene to generate a second feature data set; combining the fault feature resource library with the second feature data set to generate a synthesized fault feature data set, and adjusting the basic diagnosis model to obtain an optimized diagnosis model; and performing arc fault detection by using the optimized diagnosis model. According to the method, weak arc characteristics are amplified through a current difference strategy, time-frequency information is fused by adopting a double-branch model, unfamiliar scene adaptability is constructed based on interpretability analysis, and the problems of characteristic covering and scene generalization in arc detection are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of low-voltage user fault arc detection technology, specifically to a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method, system, device and medium. Background Technology

[0002] As the final stage of power distribution systems used for production and daily life, the safety of low-voltage power distribution lines is closely linked to people's lives. Statistics show that line faults are one of the main causes of electrical fires, often related to arc faults. Arc faults are typically caused by insulation aging or damage, cable core breakage, or loose electrical connections. When an arc fault occurs, it generates a large amount of heat, potentially reaching thousands of degrees Celsius, thus causing an electrical fire. Arc faults can be divided into parallel arc faults and series arc faults. Parallel arc faults generate large fault currents, which can be effectively detected by conventional protection devices. However, series arc faults are characterized by their concealment and randomness, and their characteristics are related to the downstream loads connected in series. With the increasing types and number of loads, the dynamic characteristics of arcs become increasingly complex, and traditional arc fault protection devices struggle to meet the high-performance requirements of field applications under complex conditions. Therefore, to prevent fires caused by arc faults, it is urgent to research reliable methods for detecting series arc faults.

[0003] Because electric current contains rich characteristics of arc faults and load information, it provides a basis for arc fault detection. Therefore, many researchers have studied detection methods based on high-frequency current data, mainly involving feature extraction of current signals from the time domain, frequency domain, and time-frequency domain, and the use of detection algorithms to diagnose arc faults. Initially, due to its low computational cost and ease of embedding, threshold judgment criteria composed of single or multiple features were used as SAF detection methods. Once the arc feature value exceeds a preset threshold, an arc fault is considered to have occurred. However, when the number of features is increased to reduce feature aliasing, manually setting a threshold with strong generalization is very difficult. In recent years, artificial intelligence algorithms have been widely used in the field of fault diagnosis. Unlike manually setting thresholds, artificial intelligence algorithms such as random forests, decision trees, K-Nearest Neighbors, fully connected neural networks, and convolutional neural networks can automatically determine the detection boundary based on human-designed features through training with feature samples and their labels. However, these SAF diagnostic methods heavily rely on expert-designed features, making it difficult to train models with strong generalization. To address this issue, some researchers have attempted to use deep learning algorithms to automatically extract features directly from the original current signal or the matrix / image after current transformation, fully exploring the information distribution of SAF samples. Its excellent fitting ability leads to higher detection accuracy, and it does not require expert features, thus eliminating the computational cost of feature extraction in the embedded system.

[0004] However, these methods typically reproduce electric arcs using arc generators and rely on a large number of known samples and labels for threshold setting or model training, followed by testing on samples with the same distribution to achieve high detection rates. However, in real-world deployment scenarios with unknown loads or unknown topology circuits, the sample distribution may shift from the training set, leading to a high risk of failure for the trained fault detection model. Therefore, the effectiveness of these methods in real-world home environments is unknown. Another key issue is the obscuring of arc characteristic information in the main circuit current. Residential circuits are complex, often with multiple branches operating simultaneously. The loads in each branch affect the main circuit current. When a series arc occurs in a branch, the main circuit current includes both the arc fault current signal from the faulty branch and the normal operating current signals from all other normal branches. In real-world homes, many loads operate simultaneously, and the normal load signals can mask the arc characteristics in the main circuit, making arc fault detection even more challenging. To solve this critical problem, in addition to extracting the general characteristics of the submerged arc in the main circuit as much as possible, it is also necessary to fine-tune the detection model based on the arc data in the actual scenario to adapt to the new sample distribution. However, in real-world scenarios, it is impossible to collect arc data in the current scenario, which requires manually installing the arc generator into the household circuit to generate the arc and collect data. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: how to improve the generalization performance and adaptability of an arc detection model trained in a controlled environment such as a laboratory when deployed in real-world scenarios with unknown load types and operating conditions.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method, comprising the following steps: acquiring electrical measurement data of a first scenario; applying a feature extraction strategy to the electrical measurement data of the first scenario to generate a first feature data set; wherein, the feature extraction strategy utilizes key state points in the electrical measurement data of the first scenario to establish a processing benchmark, and trains a basic diagnostic model based on multi-dimensional features in the first feature data set; according to the basic diagnostic model, applying a segment selection algorithm based on model parameters to fault indication samples in the first feature data set to generate a fault feature resource library; acquiring normal operating electrical data of a second scenario, and applying the feature extraction strategy to generate a second feature data set; combining the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, and adjusting the basic diagnostic model using the second feature data set and the synthetic fault feature data set to obtain an optimized diagnostic model suitable for the second scenario; and performing arc fault detection in the second scenario using the optimized diagnostic model.

[0008] As a preferred embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method of the present invention, the feature extraction strategy is to analyze the changes of predetermined characteristic parameters in the electrical measurement data, determine the key state points, and establish the processing benchmark based on the key state points.

[0009] As a preferred embodiment of the non-invasive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method described in this invention, the basic diagnostic model includes multiple feature learning pathways.

[0010] In this process, multiple feature learning pathways extract and process feature information of different dimensions from the first feature data set, and the basic diagnostic model integrates the processing results of the multiple feature learning pathways to perform fault diagnosis.

[0011] The beneficial effects of this preferred technical solution are that by setting up multiple feature learning paths to extract and process feature information of different dimensions, it is possible to simultaneously mine the feature manifestations of arc faults from both time and frequency domain perspectives. The time domain feature learning path captures the transient change features when the arc occurs, while the frequency domain feature learning path extracts the harmonic distribution changes caused by the arc fault. The collaborative work of the two paths improves the ability to identify weak arc features. Compared with single-dimensional feature extraction methods, the detection accuracy is significantly improved, especially the arc detection performance is more stable and reliable in multi-branch complex load environments.

[0012] As a preferred embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method of the present invention, the segment selection algorithm based on model parameters includes: evaluating the correlation between each data segment in the fault indication sample in the first feature data set and the preset fault category based on the internal state parameters of the basic diagnostic model; and selecting the data segment whose correlation satisfies the preset condition as a resource in the fault feature resource library.

[0013] As a preferred embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method of the present invention, the step of combining the resources in the fault feature resource library with the data of the second feature data set to generate the synthetic fault feature data set includes: superimposing or embedding data fragments in the fault feature resource library into the background data of the second feature data set according to a preset fusion rule.

[0014] As a preferred embodiment of the non-invasive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method of the present invention, the basic diagnostic model includes: a convolutional module for extracting local features from feature information of different dimensions, a self-attention module for capturing long-range dependencies between local features or between features of different dimensions, and a post-processing module for making classification decisions based on the information processed by the convolutional module and the self-attention module; adjusting the basic diagnostic model using the second feature data set and the synthetic fault feature data set includes: keeping the parameters of some of the convolutional modules in the basic diagnostic model unchanged; and optimizing and updating the parameters of the self-attention module, the post-processing module, and the remaining trainable modules in the basic diagnostic model using the second feature data set and the synthetic fault feature data set.

[0015] The beneficial effects of this preferred technical solution are that by selectively fine-tuning the parameters of some convolutional modules in the basic diagnostic model while keeping them unchanged, the model's original ability to identify arc fault features is maintained, avoiding catastrophic forgetting during the adaptation to new scenarios. At the same time, by optimizing and updating the parameters of the self-attention module and the subsequent processing module, the model can quickly adapt to the data distribution characteristics of new scenarios. This progressive model adjustment method significantly reduces computational costs and training time compared to full parameter retraining, and significantly improves the detection accuracy in new scenarios compared to the method of completely freezing parameters.

[0016] As a preferred embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method described in this invention, the first scenario is a controlled experimental environment containing multiple known types of loads, and the electrical measurement data including normal operating conditions and series arc fault operating conditions are generated and collected in a controllable manner; the second scenario is an actual user power environment whose load composition and operating status differ from the first scenario.

[0017] The beneficial effects of this preferred technical solution are that the segment selection algorithm based on model parameters can automatically identify and extract key waveform segments that are strongly correlated with arc faults, avoiding the subjectivity and uncertainty of feature selection based on expert experience in traditional methods. By analyzing the internal state parameters of the model to evaluate the importance of each data segment, it ensures that the extracted fault feature resources have high representativeness and effectiveness. The constructed fault feature resource library provides high-quality seed data for the generation of synthetic data in unfamiliar scenarios, significantly improving the authenticity of synthetic fault data and the model training effect, and solving the key technical problem of difficulty in obtaining real fault data in practical applications.

[0018] Another objective of this invention is to provide a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation system.

[0019] To address the aforementioned technical problems, this invention provides the following technical solution: a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation system, comprising: an initial model training module, used to acquire electrical measurement data of a first scenario, apply a feature extraction strategy to the electrical measurement data of the first scenario, and generate a first feature data set; a feature resource generation module, used to apply a segment selection algorithm based on model parameters to fault indication samples in the first feature data set according to the basic diagnostic model, and generate a fault feature resource library; a target scenario data processing module, used to acquire normal operating electrical data of a second scenario, and apply the feature extraction strategy to generate a second feature data set; a model adaptation module, used to combine the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, and use the second feature data set and the synthetic fault feature data set to adjust the basic diagnostic model to obtain an optimized diagnostic model suitable for the second scenario; and a fault detection module, used to perform arc fault detection in the second scenario using the optimized diagnostic model.

[0020] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method.

[0021] The present invention provides 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 non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method.

[0022] The beneficial effects of this invention are as follows: This invention proposes a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method, integrating a current differential strategy guided by key anchor points, an arc fault detection model driven by time-frequency information fusion, and a model localization method for unfamiliar scenes to improve the detection performance of the prior model in actual unfamiliar deployment scenarios. On the one hand, the current differential strategy amplifies the weak arc fault features masked by branch loads, and using differential current as input can avoid the deep learning model learning irrelevant background load features. On the other hand, by constructing a dual-branch CNN-Transformer model to fuse the time-domain and frequency-domain information of the differential current, the model can simultaneously capture dynamic features such as abrupt changes in the differential current signal and the distribution and intensity of different frequency components. These two aspects improve the model's ability to handle unfamiliar samples from the perspective of feature extraction. Meanwhile, addressing the issue of not being able to obtain real arc samples in unfamiliar scenarios, the interpretable time-domain waveform key segment selection algorithm can automatically extract segments strongly correlated with arc features from the prior database. This allows for the construction of a pseudo-arc difference database for unfamiliar scenarios, enabling the localization of the prior model in these scenarios. This reduces the model's false detection rate while preventing the model from forgetting arc features, further improving the model's detection performance in unfamiliar scenarios. Furthermore, the current difference strategy and unfamiliar scenario model localization method proposed in this invention can be integrated with previous research to enhance the ability of existing deep learning arc detection methods to handle unfamiliar scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 The overall flowchart of a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method provided in an embodiment of the present invention is shown below.

[0025] Figure 2 A flowchart of prior laboratory scenario analysis for a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method provided in an embodiment of the present invention.

[0026] Figure 3A flowchart illustrating the unfamiliar application scenario analysis of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method provided in an embodiment of the present invention.

[0027] Figure 4 A schematic diagram of a key anchor point-guided current differential strategy for a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method provided in an embodiment of the present invention.

[0028] Figure 5 The original current waveforms under different load normal and arc states in a single branch of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method provided in an embodiment of the present invention are shown.

[0029] Figure 6 The original current waveforms of different load normal and arc states under multiple branches are provided in an embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method according to an embodiment of the present invention.

[0030] Figure 7 The differential current waveforms of different load normal and arc states under multiple branches are provided in an embodiment of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method according to an embodiment of the present invention.

[0031] Figure 8 The importance evaluation result of the spectral amplitude feature based on the random forest algorithm for a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method provided in an embodiment of the present invention;

[0032] Figure 9 The diagram shows the structure of a dual-branch CNN-Transformer model driven by time-frequency information fusion, which is a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method provided in an embodiment of the present invention.

[0033] Figure 10 This is a schematic diagram of the unfamiliar scene model localization method based on dCAM, which is a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0035] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method, including:

[0036] S100: Obtain electrical measurement data of the first scenario, apply a feature extraction strategy to the electrical measurement data of the first scenario, and generate a first feature data set; wherein, the feature extraction strategy uses key state points in the electrical measurement data of the first scenario to establish a processing benchmark, and trains a basic diagnostic model based on the multi-dimensional features in the first feature data set.

[0037] S200: Based on the basic diagnostic model, apply a segment selection algorithm based on model parameters to the fault indication samples in the first feature data set to generate a fault feature resource library;

[0038] S300: Acquire normal operating electrical data of the second scenario and apply a feature extraction strategy to generate a second feature data set;

[0039] S400: Combine the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, and use the second feature data set and the synthetic fault feature data set to adjust the basic diagnostic model to obtain an optimized diagnostic model suitable for the second scenario;

[0040] S500: Utilizes an optimized diagnostic model to perform arc fault detection in the second scenario.

[0041] It should be noted that in low-voltage power distribution systems, series arc faults are characterized by their concealment and randomness, and their characteristics become complex due to the influence of downstream loads, making it difficult for traditional protection devices to meet high-performance requirements. Especially in multi-branch circuits, the normal load operating current can mask and overwhelm the weak arc characteristics in the main circuit, making detection even more challenging. Furthermore, existing artificial intelligence detection methods are typically based on models trained using samples from prior environments such as laboratories. When deployed to unfamiliar real-world scenarios with unknown loads, the model faces a high risk of failure due to shifts in sample distribution. Collecting arc data from the current scenario in real-world conditions for model adjustment is also impractical.

[0042] Therefore, to address the aforementioned issues of arc feature obscuring and poor model scene adaptability, steps S100-S500 are implemented. In S100 and S300, feature extraction strategies are applied to establish processing benchmarks using key state points, and the original electrical measurement data is processed to address the problem of arc feature obscuring. In S100, a basic diagnostic model is trained based on multi-dimensional features, providing a foundation for subsequent fault detection and model optimization. Simultaneously, a fault feature resource library is generated in S200, and in S400, the resources are combined with a second feature data set to generate a synthetic fault feature data set. This provides a way to adjust the basic diagnostic model without collecting real fault data in the second scene, ultimately obtaining an optimized diagnostic model that can adapt to the second scene. Finally, in S500, the optimized diagnostic model is used to perform arc fault detection in the second scene, solving the problem of poor model detection performance in unfamiliar scenes.

[0043] Example 2, refer to Figures 1-10 This is the second embodiment of the present invention, which provides a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method.

[0044] In this embodiment of the invention, step S100 involves acquiring electrical measurement data of the first scenario and constructing a first feature data set by acquiring basic data such as electrical parameters of the experimental environment and processing them using a feature extraction strategy.

[0045] In an alternative implementation, the first scenario in step S100 can be a simplified experimental environment containing a single type of load, in which the basic characteristics of arc faults are studied by controlling the operating state of a single load.

[0046] In another alternative implementation, the first scenario in step S100 can be a virtual experimental environment based on digital simulation, in which different load types and arc fault models are constructed using circuit simulation software.

[0047] In this embodiment of the invention, the first scenario in step S100 is a controlled experimental environment containing various known types of loads. Electrical measurement data, including normal operating conditions and series arc fault conditions, are generated and collected in a controllable manner. Specifically, an arc test platform is built in a laboratory setting to collect the main circuit current and terminal voltage data of multi-branch circuits, covering the electrical appliance start-up and shutdown process, the electrical appliance start-up transient process, the electrical appliance steady-state operation process, and the electrical appliance arc fault state. A priori database is constructed, and the collected data is analyzed offline.

[0048] In an optional implementation, the feature extraction strategy in step S100 can employ a wavelet transform-based multi-resolution analysis method to extract multi-level features of electrical signals through wavelet basis functions of different scales.

[0049] In another optional implementation, the feature extraction strategy in step S100 can adopt an adaptive signal decomposition method based on empirical mode decomposition (EMD) to decompose the complex electrical signal into several intrinsic mode functions.

[0050] In this embodiment of the invention, the feature extraction strategy in step S100 involves analyzing the changes in predetermined characteristic parameters in electrical measurement data, determining key state points, and establishing a processing benchmark based on these key state points. It should be noted that, in order to effectively extract weak arc features masked by multi-branch loads, a signal preprocessing method based on load state identification is needed. This invention employs a current differential strategy guided by key anchor points to implement this feature extraction strategy.

[0051] For example, in prior laboratory scenario analysis, this feature extraction strategy includes the following steps:

[0052] Step 102: Construct a critical load change point detector and combine it with a sliding time window to detect critical load change points in the current sequence within the window. If a critical change point is detected, proceed to step 103.

[0053] Step 103: Construct a critical load steady-state point detector and combine it with a sliding time window to detect the critical load steady-state point of the current sequence within the window. If a critical steady-state point is detected, save the first cycle data after the critical load steady-state point as the differential current reference and proceed to step 104.

[0054] Step 104: Align each cycle after the differential current reference with the differential current reference according to the voltage phase, and perform differential analysis. At the same time, perform key load change point detection. If a key change point is detected, proceed to step 103. Continue until all the collected data has been differentially analyzed to obtain the prior differential database, and then proceed to step 105.

[0055] It should be noted that the process of the non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method is as follows in the prior laboratory scenario analysis: Figure 2 As shown, it mainly includes steps 101 to 107. Before step 102, it also includes step 101: build an arc test platform in a laboratory setting, collect the main circuit current and terminal voltage data of multi-branch circuits, covering the electrical appliance start-up and shutdown process, the electrical appliance start-up transient process, the electrical appliance steady-state operation process, and the electrical appliance arc fault state, construct a priori database and perform offline analysis on the collected data.

[0056] Step 105: Perform first-order and second-order derivative and FFT transformation on the differential current waveforms in the prior differential database to obtain the three-dimensional time series (differential current waveform, first-order derivative of differential current, second-order derivative of differential current) and the amplitude of each spectrum. Then, use the random forest algorithm to evaluate and filter the importance of the spectrum amplitude features.

[0057] Step 106: Construct a deep learning model based on time-frequency information fusion, and train it using the three-dimensional time series of the prior difference database and the selected spectral amplitude features to obtain the trained prior scenario arc fault diagnosis model; wherein, the deep learning model based on time-frequency information fusion is a neural network model with a two-branch CNN-Transformer and an attention mechanism.

[0058] Step 107: Using the proposed interpretation-wise classactivation map (dCAM)-based algorithm for selecting key segments of time-domain waveforms, combined with the parameters of the arc fault diagnosis model, waveform segments strongly associated with the arc are automatically extracted from the prior difference database to form a prior arc-associated waveform segment library.

[0059] In an optional implementation, the key state points in step S100 can be abnormal change points of electrical parameters determined based on statistical analysis, which are identified by monitoring changes in the statistical characteristics of the parameters.

[0060] In another alternative implementation, the key state points in step S100 can be system state transition points identified based on machine learning anomaly detection algorithms, such as anomaly moments detected using the Isolation Forest algorithm.

[0061] In this embodiment of the invention, the key state points in step S100 include key load change points and key load steady-state points. Specifically, a key load change point refers to a current change point caused by the switching on and off of electrical appliances. The specific implementation method for constructing a key load change point detector is as follows: a rule-based key load change point detector is constructed using the concept of mean change.

[0062] For example, for the total port current RMS time series I = [i1, i2, ..., i...] N Assume W m It is a window of length m, W d Given a window of length d, calculate the mean of the valid values ​​within the two windows:

[0063]

[0064] Where, μ m For window W m The mean effective value of the current within, μ d For window W d The average effective value of the current within.

[0065] Define the cumulative average of the load change points caused by the switching on and off of electrical appliances as follows:

[0066]

[0067] in, and These are the cumulative sums of the average values ​​corresponding to the load change points when the appliances are turned on and when they are turned off, respectively, at the k-th point. and These correspond to their initial values, respectively, with β representing noise introduced from the outside.

[0068] To prevent misjudgment of point changes caused by arc pulses, when or When the cumulative sum of the means is defined as:

[0069]

[0070] when or This indicates that there is a possibility of a change point occurring at this time, but the cumulative sum has not yet reached the level to confirm the occurrence of an event. or At that time, it was confirmed that a critical load change point was detected, and the change point time was [time missing]. or The moment of increasing from 0.

[0071] Furthermore, the critical load steady-state point refers to the first current steady-state point after the critical load changes. The specific implementation method for constructing the critical load steady-state point detector is as follows: based on the critical load change point detector in step 102, a variance calculation window W is added. v Given a window length of v, calculate the mean and variance of the effective values ​​within the window as follows:

[0072]

[0073] Where, μ v For window W v The mean effective value of the current within, σ v For window W v The variance of the effective value of the current within the range.

[0074] when And σ v Less than the threshold h v When the critical load steady-state point is detected, it is considered that the critical load steady-state point is the kth point.

[0075] In an optional implementation, the processing reference in step S100 can be a fixed reference value determined based on the statistical characteristics of historical data, such as the long-term average current value as a reference.

[0076] In another alternative implementation, the processing benchmark in step S100 can be an adaptive benchmark value that is dynamically updated based on a sliding window, and the benchmark parameters are adjusted in real time according to the characteristics of recent data.

[0077] In this embodiment of the invention, the processing reference in step S100 is a differential current reference, specifically, saving the first cycle data after the steady-state point of the important load as the differential current reference. Each cycle I after the differential current reference... s =[i s1 i s2 ,...,i sn [With differential current reference I] b =[i b1 i b2 ,...,i bn The alignment method is to align them according to the positive zero-crossing points of their respective voltage phases, and the differential current I obtained after differential division is... d for:

[0078] I d =I b -I s ;

[0079] In one optional implementation, the multidimensional features in step S100 may include time-domain statistical features (such as mean, variance, skewness, kurtosis) and frequency-domain energy features (such as power spectral density of each frequency band).

[0080] In another alternative implementation, the multidimensional features in step S100 may include time-frequency features based on wavelet transform and instantaneous frequency features based on Hilbert-Huang transform.

[0081] In this embodiment of the invention, the multidimensional features in step S100 are a three-dimensional time series (differential current waveform, first derivative of differential current, second derivative of differential current) and various spectral amplitudes, specifically referring to the filtered spectral amplitude features.

[0082] Specifically, in step 105, the differential current waveforms in the prior differential database are subjected to first-order and second-order derivative and FFT transformations to obtain a three-dimensional time series (differential current waveform, differential current first-order derivative, differential current second-order derivative) and each spectral amplitude. The random forest algorithm is then used to evaluate and filter the importance of the spectral amplitude features.

[0083] The calculation methods for the first and second derivatives of the differential current waveform are as follows:

[0084] I d1 =diff(I d )

[0085] I d2 =diff(I d1 )

[0086] Among them, I d1 For the first derivative of the differential current waveform, I d2 The second derivative of the differential current waveform is maintained by filling the initial point to keep it consistent with the differential current I. d Keep the same length n. diff represents calculating the difference between adjacent elements in the sequence.

[0087] It should be noted that, before using the random forest algorithm to evaluate the importance of the spectral amplitude features, in order to reduce the computation time, the sampling frequency of the time series was reduced to 12.5kHz before performing FFT transformation and spectral amplitude feature extraction.

[0088] The importance of spectral amplitude features is evaluated using the random forest algorithm by calculating the Gini index of the random forest to determine the importance of each spectral amplitude. The specific implementation method is as follows:

[0089] In the j-th decision tree, the Gini index of node t is:

[0090]

[0091] In the formula, C represents the number of categories at feature node t. In the arc detection direction, C = 2, i.e., normal or arc; P tc This represents the proportion of category c in node t.

[0092] The expression for the change in the Gini exponent before and after branching at node t is:

[0093] ΔG=GI t -GI l -GI r ;

[0094] In the formula, GI l Gini index and GI before splitting r The Gini index is the result of the split. The Gini index of each node is weighted, with the weight being the number of node samples, n. t Divide by the total number of samples N in the prior difference database sample :

[0095]

[0096] Feature F q The importance of node t in the j-th tree is:

[0097]

[0098] Feature F q If a node appearing in decision tree j is in set T, then F q The importance of the j-th tree is:

[0099]

[0100] If there are N in the random forest tree Trees, then F q Its importance in random forests is:

[0101]

[0102] After normalization, the final feature Gini index, i.e., feature importance, is obtained:

[0103]

[0104] In the formula, N feature The number of spectral features depends on the signal sampling rate f. s In this embodiment, f s =12500.

[0105] Figure 8 The results of the importance assessment of spectral amplitude features based on the random forest algorithm in this embodiment show that many low-frequency harmonics have high importance scores, while only a few high-frequency harmonics have importance scores of around 0.5.

[0106] The random forest algorithm is used to filter spectral amplitude features by selecting the top N features in order of importance. s The selection criteria were based on optimizing the model's detection performance. In this embodiment, the top 32 features in terms of feature importance were ultimately selected as the filtered spectral amplitude features.

[0107] In this embodiment of the invention, the basic diagnostic model in step S100 includes multiple feature learning pathways; wherein, the multiple feature learning pathways extract and process feature information of different dimensions from the first feature data set respectively, and the basic diagnostic model integrates the processing results of the multiple feature learning pathways to perform fault diagnosis.

[0108] In an alternative implementation, the basic diagnostic model in step S100 can use traditional machine learning methods, such as shallow models like support vector machines (SVM) or random forests, for fault classification.

[0109] In another alternative implementation, the basic diagnostic model in step S100 can employ a single deep neural network architecture, such as a pure convolutional neural network or a pure recurrent neural network, for end-to-end fault detection.

[0110] In this embodiment of the invention, the basic diagnostic model in step S100 is a priori scenario arc fault diagnostic model, specifically a dual-branch CNN-Transformer model driven by time-frequency information fusion.

[0111] In one alternative implementation, the multiple feature learning pathways in step S100 may include convolutional branches based on different window sizes, each capturing local features at different time scales.

[0112] In another alternative implementation, the multiple feature learning pathways in step S100 may include parallel branches based on different pooling strategies, such as max pooling branches and average pooling branches.

[0113] In this embodiment of the invention, the multiple feature learning pathways in step S100 form a dual-branch structure for the basic diagnostic model, namely, a time-domain feature extraction branch and a frequency-domain feature extraction branch.

[0114] Specifically, step 106: Construct a deep learning model based on time-frequency information fusion, and train it using the three-dimensional time series data from the prior difference database and the selected spectral amplitude features to obtain the trained prior scenario arc fault diagnosis model. The deep learning model based on time-frequency information fusion is a neural network model with a two-branch CNN-Transformer and an attention mechanism.

[0115] Deep learning models based on time-frequency information fusion include a time-domain feature extraction branch, a frequency-domain feature extraction branch, a Transformer encoder, a cross-attention mechanism, an adaptive average pooling layer, and a fully connected layer.

[0116] The input to the temporal feature extraction branch is a three-dimensional time series (differential current waveform, first derivative of differential current, and second derivative of differential current). This branch consists of one residual block, one convolutional block, and one residual block stacked sequentially. Each block uses three convolutional layers for feature extraction. Each convolutional layer uses padding to ensure that the size of the output feature map is consistent with the input. After each convolutional layer, a ReLU activation function and a batch normalization layer (BatchNorm) are applied sequentially. In each residual block, if the input and output channels do not match, channel alignment is performed using a 1×1 convolution, followed by skip connections. In this embodiment, the number of channels in the convolutional layers of the three different blocks are 32, 32, and 64, respectively. The kernel sizes of the three convolutional layers in each block are 100, 50, and 20, respectively.

[0117] The frequency domain feature extraction branch takes as input the filtered spectral amplitude features. This branch consists of three stacked convolutional blocks, each containing two convolutional layers and one max-pooling layer. Each convolutional layer uses padding to ensure the output feature map size matches the input, and a ReLU activation function and a batch normalization layer are applied sequentially after each max-pooling layer. In this embodiment, the number of channels in the convolutional layers of the three convolutional blocks are 16, 32, and 64, respectively. The convolutional kernel size in each block is 5, and the pooling kernel size is 3.

[0118] In an optional implementation, the convolution module in step S100 can adopt a depthwise separable convolution structure, which reduces the number of parameters and computational complexity by separating depthwise convolution and pointwise convolution.

[0119] In another optional implementation, the self-attention module in step S100 can adopt a local attention mechanism, which calculates attention weights only in the local neighborhood, thereby reducing computational complexity.

[0120] In this embodiment of the invention, the basic diagnostic model in step S100 includes a convolutional module for extracting local features from feature information of different dimensions, a self-attention module for capturing long-range dependencies between local features or between features of different dimensions, and a post-processing module for making classification decisions based on the information processed by the convolutional module and the self-attention module. The convolutional module is the CNN part (composed of convolutional layers, etc.) of the basic diagnostic model, and the self-attention module is the Transformer part (composed of self-attention layers, etc.) of the basic diagnostic model.

[0121] The Transformer encoder takes as input the temporal and frequency domain features extracted by the temporal and frequency domain feature extraction branches, respectively. The Transformer encoder consists of multiple stacked self-attention layers. Each self-attention layer is followed by a feedforward neural network, which consists of two fully connected layers connected by the ReLU activation function. Residual connections and layer normalization operations follow the self-attention layers and the feedforward neural network layers.

[0122] The cross-attention mechanism achieves the fusion of time-domain and frequency-domain features after the Transformer encoder. This attention mechanism integrates the frequency-domain features F after the Transformer encoder. f As the query vector, the time-domain feature F encoded by the Transformer t The key vector and value vector are used as the input vectors. The attention weights are calculated and multiplied by the value vectors, and the sum of these results yields the encoder's adjusted output. The formula for calculating the cross-attention output is as follows:

[0123]

[0124] In the formula, d k F represents t Dimensions.

[0125] In this embodiment, the proposed model is implemented on the PyTorch platform, using classification cross-entropy as the loss function. The AdamW optimizer is employed, with a batch size of 64 and an adaptive learning rate strategy, trained for 70 epochs. The initial learning rate is 0.0001, and the minimum learning rate is 0.000001.

[0126] In this embodiment of the invention, the segment selection algorithm based on model parameters in step S200 includes: based on the internal state parameters of the basic diagnostic model; evaluating the correlation between each data segment in the fault indication sample in the first feature data set and the preset fault category; and selecting the data segment whose correlation satisfies the preset condition as a resource in the fault feature resource library.

[0127] In an optional implementation, the segment selection algorithm based on model parameters in step S200 can adopt a method based on activation value statistical analysis to determine important segments by analyzing the distribution characteristics of activation values ​​in each layer of the model.

[0128] In another alternative implementation, the segment selection algorithm based on model parameters in step S200 can adopt a method based on attention weight analysis, which identifies key segments by visualizing the weight distribution of the attention mechanism.

[0129] In this embodiment of the invention, the segment selection algorithm based on model parameters in step S200 is an interpretable time-domain waveform key segment selection algorithm based on multi-dimensional class activation graph (dCAM).

[0130] Specifically, in step 107, the proposed interpretable time-domain waveform key segment selection algorithm based on dimension-wise classactivation map (dCAM) is used. Combined with the parameters of the arc fault diagnosis model, waveform segments strongly associated with the arc in the prior difference database are automatically extracted to form a prior arc-associated waveform segment library.

[0131] The input to the dCAM-based interpretable time-domain waveform key segment selection algorithm is cycle-level arc samples from a priori differential database and the features F extracted from the arc samples by the time-domain feature extraction branch of the priori scenario arc fault diagnosis model. k The weights W of the fully connected layer in the prior scenario arc fault diagnosis model k .

[0132] The output of the dCAM algorithm is the importance assessment of the arc samples in the prior difference database at each sampling point. The higher the importance of the sampling point, the more related it is to the arc.

[0133] The prerequisite for selecting key segments of interpretable time-domain waveforms based on dCAM is that the prior scenario arc fault diagnosis model correctly diagnoses the arc sample and the confidence level is greater than the threshold.

[0134] The dCAM algorithm evaluates the importance of arc sample sampling points in steps A1 to A5:

[0135] Step A1: For a given cycle-level arc sample in the prior difference database, input I = [I 1 ,I 2 ,I 3 ], its size is 3*n, through I 1 ,I 2 ,I 3 A new 3*3*n matrix C(I) is constructed by random permutations and combinations, where n is the number of sampling points in one cycle of the waveform. C(I) requires that each row and column contains no duplicate elements. Therefore, C(I) has 6 possible permutations and combinations:

[0136] C1(I)=[I 1 ,I 2 ,I 3 ;I 3 ,I 1 ,I 2 ;I 2 ,I 3 ,I 1 ];

[0137] C2(I)=[I 1 ,I 3 ,I 2 ;I 2 ,I 1 ,I 3 ;I 3 ,I 2 ,I 1 ];

[0138] C3(I)=[I 2 ,I 3 ,I 1 ;I 3 ,I 1 ,I 2 ;I 1 ,I 2 ,I 3 ];

[0139] C4(I) = [I 2 ,I 1 ,I3 ;I 3 ,I 2 ,I 1 ;I 1 ,I 3 ,I 2 ];

[0140] C5(I) = [I 3 ,I 1 ,I 2 ;I 1 ,I 2 ,I 3 ;I 2 ,I 3 ,I 1 ];

[0141] C6(I)=[I 3 ,I 2 ,I 1 ;I 1 ,I 3 ,I 2 ;I 2 ,I 1 ,I 3 ];

[0142] In this embodiment, n = 250.

[0143] Step A2: For C(I) under the 6 permutations and combinations, calculate its class activation map (CAM) to obtain CAM(C1(I)), CAM(C2(I)), CAM(C3(I)), CAM(C4(I)), CAM(C5(I)), and CAM(C6(I)). The calculation formula is as follows:

[0144]

[0145] In the formula, K s The number of channels representing the features extracted from the arc sample by the temporal feature extraction branch of the prior scenario arc fault diagnosis model, and the number of neurons in the fully connected layer of the prior scenario arc fault diagnosis model, where k represents the k-th channel or the k-th neuron. K represents the weights that the fully connected layer uses to map features to the output category c, i.e., arc or normal. In this embodiment, K s =64.

[0146] Step 3: Calculate the mean of CAM(C1(I)), CAM(C2(I)), CAM(C3(I)), CAM(C4(I)), CAM(C5(I)), and CAM(C6(I)). Right now

[0147]

[0148] Step 4: According to The three-dimensional input I = [I] is obtained. 1 ,I 2 ,I 3 The corresponding dCAM has a size of 3*n, which is 3*250 in this embodiment:

[0149]

[0150] In the formula, μ and σ represent respectively The mean and variance of.

[0151] Step 5: Calculate the mean of dCAM in the first dimension to obtain the importance of each point in the original difference waveform. Its size is 1*n, which is 1*250 in this embodiment.

[0152] The method for selecting key points in the time-domain waveform is that the importance of the sampling point is greater than 0.5 times the average importance of all sampling points in that cycle.

[0153] The selected arc-related sampling points are filled and deleted to ensure that the arc-related sampling points are continuous and without isolated points, so as to extract the key segments of the complete arc-related waveform and then construct a priori arc-related waveform segment library.

[0154] In an optional implementation, the fault feature resource library in step S200 can be a fault mode library built based on expert knowledge, containing typical feature templates for various known fault types.

[0155] In another alternative implementation, the fault feature resource library in step S200 can be a set of feature clusters constructed based on cluster analysis, and potential fault modes in the data can be extracted through unsupervised learning methods.

[0156] In this embodiment of the invention, the fault feature resource library in step S200 is a priori arc-related waveform segment library, which is composed of waveform segments strongly associated with the arc extracted by the above-mentioned dCAM algorithm.

[0157] In this embodiment of the invention, in step S300, normal operating electrical data of the second scenario is obtained, and a feature extraction strategy is applied to generate a second feature data set.

[0158] In an alternative implementation, the second scenario in step S300 can be a test environment with a similar load type to the first scenario but a different operating mode, used to verify the basic generalization ability of the model.

[0159] In another optional implementation, the second scenario in step S300 can be a special electrical environment with specific industry characteristics, such as hospitals, data centers and other places with high electrical safety requirements.

[0160] In this embodiment of the invention, the second scenario in step S300 is an actual user power consumption environment where the load composition and operating status differ from the first scenario. Specifically, step 201 collects the main circuit current and terminal voltage data of multi-branch circuits in unfamiliar application scenarios, covering the appliance start-up and shutdown process, the appliance start-up transient process, and the appliance steady-state operation process, constructs a normal database for unfamiliar scenarios, and performs offline analysis on the collected data.

[0161] In an alternative implementation, the second feature data set in step S300 can be a feature set extracted based on traditional signal processing methods, such as frequency domain features based on Fourier transform.

[0162] In another alternative implementation, the second feature data set in step S300 can be a high-dimensional feature representation automatically learned based on a deep feature extraction network.

[0163] In this embodiment of the invention, the second feature data set in step S300 is a normal differential database for unfamiliar scenarios. Specifically, in step 202, the current differential strategy guided by the key anchor points in steps 102-104 is used to differentially analyze the normal database for unfamiliar scenarios to obtain the normal differential database for unfamiliar scenarios.

[0164] In this embodiment of the invention, step S400, which combines the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, includes: superimposing or embedding data fragments from the fault feature resource library into the background data of the second feature data set according to a preset fusion rule.

[0165] In an optional implementation, the synthetic fault feature data set in step S400 can be generated by a fault simulation method based on a physical model, and synthetic data can be generated by establishing a mathematical model based on the fault mechanism.

[0166] In another alternative implementation, the synthetic fault feature data set in step S400 can be created by a data generation method based on generative adversarial networks (GANs), training a generator to produce realistic fault feature data.

[0167] In this embodiment of the invention, the fault feature data set synthesized in step S400 is a pseudo-arc differential database for unfamiliar scenarios. Specifically, in step 203, the prior arc-related waveform segment library from step 107 is randomly injected into the normal differential database for unfamiliar scenarios to construct the pseudo-arc differential database for unfamiliar scenarios.

[0168] In an optional implementation, the combination method in step S400 can be a linear superposition method, which simply adds the fault characteristic signal to the normal background signal.

[0169] In another alternative implementation, the combination method in step S400 can be a frequency domain fusion method, which combines different signal components in the frequency domain and then converts them back to the time domain.

[0170] In this embodiment of the invention, step S400, which involves adjusting the basic diagnostic model using the second feature data set and the synthetic fault feature data set, includes: keeping the parameters of some convolutional modules in the basic diagnostic model unchanged; and using the second feature data set and the synthetic fault feature data set to optimize and update the parameters of the self-attention module, the subsequent processing module, and the remaining trainable modules in the basic diagnostic model.

[0171] In an alternative implementation, the method for adjusting the basic diagnostic model in step S400 can be to retrain all parameters, using new scene data to train the entire model end-to-end.

[0172] In another alternative implementation, the method for adjusting the basic diagnostic model in step S400 can be a transfer learning approach, which updates only the parameters of the last few layers of the model while keeping the underlying feature extraction capability unchanged.

[0173] In this embodiment of the invention, the adjustment of the basic diagnostic model in step S400 adopts a fine-tuning strategy, specifically freezing the parameters of the first convolutional block and performing gradient updates and fine-tuning on the other model parameters.

[0174] Furthermore, such as Figure 3 As shown, the process of analyzing unfamiliar application scenarios using the non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method includes steps 201 to 205:

[0175] Step 201: Collect the main circuit current and terminal voltage data of multi-branch circuits in unfamiliar application scenarios, covering the start-up and shutdown process of electrical appliances, the transient process of electrical appliances starting up, and the steady-state operation process of electrical appliances. Construct a normal database for unfamiliar scenarios and perform offline analysis on the collected data.

[0176] Step 202: Use the current differential strategy guided by the key anchor points in steps 102 to 104 to perform differential analysis on the normal database of unfamiliar scenarios to obtain the normal differential database of unfamiliar scenarios.

[0177] Step 203: Randomly inject the prior arc-related waveform segment library from Step 107 into the unfamiliar scene normal difference database to construct the unfamiliar scene pseudo arc difference database;

[0178] The strategy for randomly injecting the prior arc-related waveform segment library into the normal difference database of unfamiliar scenarios is as follows:

[0179] I synthetic =(1-w)·S SAF +w·I normal +β;

[0180] In the formula, I normal S represents a normal differential waveform in an unfamiliar scene. SAF Represents the prior arc-related waveform segment, I synthetic β represents the pseudo-arc differential waveform in an unfamiliar scenario, w represents the time-varying weighting function, the purpose of which is to achieve smooth generation of the pseudo-arc differential waveform, and β represents random noise.

[0181] Step 204: Fine-tune the prior arc fault diagnosis model mentioned in step 106 using the unfamiliar scene normal difference database and the unfamiliar scene pseudo arc difference database to obtain the unfamiliar scene localized arc fault diagnosis model.

[0182] The strategy for fine-tuning the prior arc fault diagnosis model in step 106 using the unfamiliar scene normal difference database and the unfamiliar scene pseudo arc difference database is to freeze the parameters of the first convolutional block of the time domain and frequency domain feature extraction branch of the prior scene arc fault diagnosis model, and perform gradient updates and fine-tuning on other model parameters to obtain an unfamiliar scene localized arc fault diagnosis model.

[0183] Step 205: Use the arc fault diagnosis model for unfamiliar scenarios localized in Step 204 to perform cycle-level arc monitoring in unfamiliar scenarios.

[0184] Specifically, in step 204, the strategy for fine-tuning the prior arc fault diagnosis model in step 106 using the unfamiliar scene normal difference database and the unfamiliar scene pseudo arc difference database is to freeze the parameters of the first convolutional block of the time domain and frequency domain feature extraction branch of the prior scene arc fault diagnosis model, and perform gradient updates and fine-tuning on other model parameters to obtain an unfamiliar scene localized arc fault diagnosis model.

[0185] In an optional implementation, the optimized diagnostic model in step S400 can be a multi-model fusion system based on ensemble learning, which improves detection accuracy by combining the prediction results of multiple sub-models.

[0186] In another alternative implementation, the optimized diagnostic model in step S400 can be an adaptive model based on online learning, which can continuously update model parameters according to real-time running data.

[0187] In this embodiment of the invention, the optimized diagnostic model in step S500 is a localized arc fault diagnosis model for unfamiliar scenarios. Specifically, step 205 utilizes the localized arc fault diagnosis model for unfamiliar scenarios from step 204 to perform cycle-level arc monitoring in the unfamiliar scenario.

[0188] Example 3, the third embodiment of the present invention, differs from the previous two embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0189] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0190] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0191] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0192] Example 4 is the fourth embodiment of the present invention. This embodiment provides a non-invasive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation system, including:

[0193] The initial model training module is used to acquire electrical measurement data of the first scenario, apply a feature extraction strategy to the electrical measurement data of the first scenario, and generate a first feature data set.

[0194] The feature resource generation module is used to apply a segment selection algorithm based on model parameters to the fault indication samples in the first feature data set according to the basic diagnostic model, and generate a fault feature resource library.

[0195] The target scenario data processing module is used to acquire the normal operating electrical data of the second scenario and apply a feature extraction strategy to generate a second feature data set.

[0196] The model adaptation module is used to combine the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, and to adjust the basic diagnostic model using the second feature data set and the synthetic fault feature data set to obtain an optimized diagnostic model suitable for the second scenario.

[0197] The fault detection module is used to perform arc fault detection in the second scenario using an optimized diagnostic model.

[0198] Example 5, refer to Figures 1-10 This is the fifth embodiment of the present invention, which provides a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method.

[0199] To verify the effectiveness of the method used in this invention, relevant experiments were conducted in this embodiment. The electrical appliances in the arc test platform of the first scenario include resistive loads such as electric kettles, series-wound motor loads such as electric drills, induction motor loads such as fans, brushed motor loads such as high-speed / low-speed hair dryers, switching power supply loads such as laptops, eddy current loads such as induction cookers, half-wave rectifier loads such as low-speed / low-speed hair dryers, and magnetron loads such as microwave ovens. A multi-branch circuit refers to multiple loads operating in parallel simultaneously, including arc loads and normal loads. The main circuit current and terminal voltage data of the multi-branch circuit refer to the total current and supply voltage, including all load currents, that can be collected at the meter.

[0200] To better illustrate the current differential strategy guided by the key anchor point of this invention, Figure 4 The differential diagram illustrates the main circuit current waveform under multi-branch conditions. Under the original background load conditions, the kettle and hair dryer are turned on at 0.3 seconds and 0.8 seconds, respectively, and an arc fault occurs in the kettle branch at 1.5 seconds. When the appliances are turned on, a sudden change in the main circuit current is caused, and the critical load change point detector detects and marks it. Subsequently, after the current enters steady state, the critical load steady-state point detector detects the critical load steady-state point and saves the current of the next cycle after the critical load steady-state point as the differential reference current. Each cycle after the differential reference current is used as the differential current, and the differential current is obtained by subtracting the reference current from the differential current. Figure 4 The original current waveform within the dashed line in the figure shows the original current waveforms of multiple branches within the dashed box. It was found that the original current of the main circuit did not change significantly before and after the arc occurred. Figure 4 The differential current waveform diagram within the dashed line shows the original current waveforms of the multiple branches within the dashed box. It is found that the differential current is small before the arc occurs, but becomes significantly larger after the arc occurs. The differential current of the main circuit changes significantly before and after the arc occurs, indicating that the current differential strategy can amplify the weak arc fault characteristics that are masked by the branch load.

[0201] Figure 5 The original current waveforms under normal and arcing states under different loads in a single-branch configuration are presented. It is found that under a single-branch configuration, significant changes occur before and after arcing at different loads, exhibiting characteristics such as zero-wave, pulse, and distortion. However, as... Figure 6 The diagram shows the original current waveforms under different loads and under arc conditions in a multi-branch system. When an arc-generating appliance is connected in parallel with a normal appliance, the arc characteristics in the main circuit are submerged, and the current waveform changes slightly before and after the arc, making it difficult to detect the arc. Figure 7 The diagram shows the differential current waveforms under different loads in normal and arc states under multiple branches. It can be seen that when the differential strategy is adopted, the waveform of normal electrical appliances is weakened, the characteristics of arc faults are decoupled from the main circuit current, and the differential current of each electrical appliance changes greatly before and after the arc occurs.

[0202] Figure 8 The results of the importance assessment of spectral amplitude features based on the random forest algorithm show that many low-frequency harmonics have high importance scores, while only a few high-frequency harmonics have importance scores of around 0.5.

[0203] Figure 9 This is a structural diagram of a two-branch CNN-Transformer model driven by time-frequency information fusion.

[0204] Figure 10 This is a schematic diagram of the localization method for unfamiliar scene models based on dCAM.

[0205] It should be noted that this invention proposes a non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method, integrating a current differential strategy guided by key anchor points, an arc fault detection model driven by time-frequency information fusion, and a model localization method for unfamiliar scenarios to improve the detection performance of the prior model in actual unfamiliar deployment scenarios. On the one hand, the current differential strategy amplifies the weak arc fault features masked by branch loads, and using differential current as input can avoid the deep learning model learning irrelevant background load features. On the other hand, by constructing a dual-branch CNN-Transformer model to fuse the time-domain and frequency-domain information of the differential current, the model can simultaneously capture dynamic features such as abrupt changes in the differential current signal and the distribution and intensity of different frequency components. These two aspects improve the model's ability to handle unfamiliar samples from the perspective of feature extraction. Meanwhile, addressing the issue of not being able to obtain real arc samples in unfamiliar scenarios, the interpretable time-domain waveform key segment selection algorithm can automatically extract segments strongly correlated with arc features from the prior database. This allows for the construction of a pseudo-arc difference database for unfamiliar scenarios, enabling the localization of the prior model in these scenarios. This reduces the model's false detection rate while preventing the model from forgetting arc features, further improving the model's detection performance in unfamiliar scenarios. Furthermore, the current difference strategy and unfamiliar scenario model localization method proposed in this invention can be integrated with previous research to enhance the ability of existing deep learning arc detection methods to handle unfamiliar scenarios.

[0206] In summary, this invention clarifies the key issues of arc feature masking and poor model scene adaptability in non-intrusive AC series arc fault detection. Based on the feature extraction strategy of key state points and the deep learning model of multi-dimensional feature fusion, it achieves accurate extraction and identification of weak arc fault features. Through the segment selection algorithm based on model parameters and the construction of a synthetic fault feature dataset, it realizes adaptive optimization of the model in unfamiliar scenarios, provides intelligent safety early warning information for important power systems, reduces equipment failures and accidents, improves system operating efficiency, and reduces costs while increasing benefits.

[0207] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method, characterized by: include, Acquire electrical measurement data for a first scenario, apply a feature extraction strategy to the electrical measurement data for the first scenario, and generate a first feature data set; wherein, the feature extraction strategy uses key state points in the electrical measurement data of the first scenario to establish a processing benchmark, and trains a basic diagnostic model based on the multidimensional features in the first feature data set; Based on the basic diagnostic model, a segment selection algorithm based on model parameters is applied to the fault indication samples in the first feature data set to generate a fault feature resource library. Acquire normal operating electrical data for the second scenario, and apply the feature extraction strategy to generate a second feature data set; The resources in the fault feature resource library are combined with the data in the second feature data set to generate a synthetic fault feature data set. The basic diagnostic model is then adjusted using the second feature data set and the synthetic fault feature data set to obtain an optimized diagnostic model suitable for the second scenario. Arc fault detection is performed in the second scenario using the optimized diagnostic model.

2. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 1, characterized in that: The feature extraction strategy involves analyzing the changes in predetermined characteristic parameters in the electrical measurement data, determining the key state points, and establishing the processing benchmark based on the key state points.

3. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 2, characterized in that: The basic diagnostic model includes multiple feature learning pathways; In this process, multiple feature learning pathways extract and process feature information of different dimensions from the first feature data set, and the basic diagnostic model integrates the processing results of the multiple feature learning pathways to perform fault diagnosis.

4. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 3, characterized in that: The segment selection algorithm based on model parameters includes: Based on the internal state parameters of the aforementioned basic diagnostic model; Evaluate the correlation between each data segment in the fault indication sample of the first feature data set and the preset fault category; Data segments whose correlation meets the preset conditions are selected as resources in the fault feature resource library.

5. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 4, characterized in that: The step of combining the resources in the fault feature resource library with the data in the second feature data set to generate the synthetic fault feature data set includes: superimposing or embedding data fragments from the fault feature resource library into the background data of the second feature data set according to preset fusion rules.

6. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 5, characterized in that: The basic diagnostic model includes a convolutional module for extracting local features from feature information of different dimensions, a self-attention module for capturing long-range dependencies between local features or between features of different dimensions, and a post-processing module for making classification decisions based on the information processed by the convolutional module and the self-attention module. The step of adjusting the basic diagnostic model using the second feature data set and the synthetic fault feature data set includes: keeping the parameters of some of the convolutional modules in the basic diagnostic model unchanged; and using the second feature data set and the synthetic fault feature data set to optimize and update the parameters of the self-attention module, the subsequent processing module, and the remaining trainable modules in the basic diagnostic model.

7. The non-invasive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in claim 6, characterized in that: The first scenario is a controlled experimental environment containing multiple known types of loads, and electrical measurement data including normal operating conditions and series arc fault conditions are generated and collected in a controllable manner; the second scenario is an actual user power environment where the load composition and operating status differ from the first scenario.

8. A non-invasive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation system, employing the non-invasive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method as described in any one of claims 1 to 7, characterized in that, include: The initial model training module is used to acquire electrical measurement data of the first scenario, apply a feature extraction strategy to the electrical measurement data of the first scenario, and generate a first feature data set. The feature resource generation module is used to apply a segment selection algorithm based on model parameters to the fault indication samples in the first feature data set according to the basic diagnostic model to generate a fault feature resource library. The target scenario data processing module is used to acquire the normal operating electrical data of the second scenario and apply the feature extraction strategy to generate a second feature data set. The model adaptation module is used to combine the resources in the fault feature resource library with the data in the second feature data set to generate a synthetic fault feature data set, and to adjust the basic diagnostic model using the second feature data set and the synthetic fault feature data set to obtain an optimized diagnostic model suitable for the second scenario. The fault detection module is used to perform arc fault detection in the second scenario using the optimized diagnostic model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scene adaptation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the non-intrusive AC series arc fault detection generalization enhancement and unfamiliar scenario adaptation method as described in any one of claims 1 to 7.

Images