Arc Fault Detection Methods

Abstract

This application provides an arc fault detection method applicable to the field of power monitoring technology. The method includes: obtaining differential current and target voltage based on the electrical signal to be detected in the circuit under test, wherein the circuit under test includes a main circuit and multiple parallel load branches, and the electrical signal to be detected includes the main current and voltage to be detected; obtaining multi-view features based on the differential current and target voltage; processing the multi-view features using a first feature extraction model to obtain a first feature; and, if the first arc fault detection result obtained by arc fault detection on the first feature indicates that a series arc fault exists in the circuit under test, performing arc fault detection on the second feature obtained by processing the multi-view features using a second feature extraction model to obtain a second arc fault detection result indicating that the circuit under test does not have a series arc fault or that a parallel load branch has a series arc fault.

Description

Technical Field

[0001] This application relates to the field of power monitoring technology, and specifically to a method for detecting electric arc faults. Background Technology

[0002] A significant proportion of electrical fires are caused by arc faults, which can result from insulation aging, cable core breakage, or loose electrical connections. Arc faults generate temperatures of thousands of degrees Celsius, easily igniting surrounding flammable materials. However, existing arc fault detection models have low accuracy in detecting arc faults. Summary of the Invention

[0003] In view of the above problems, this application provides an arc fault detection method.

[0004] According to a first aspect of the embodiments of this application, an arc fault detection method is provided, comprising: obtaining a differential current and a target voltage based on a detectable electrical signal of a circuit under test, wherein the circuit under test includes a main circuit and multiple parallel load branches, the detectable electrical signal includes a detectable main circuit current and a detectable main circuit voltage, the differential current characterizing the difference between cycles of the phase-aligned detectable main circuit current to reflect waveform distortion and non-stationary evolution caused by a series arc fault, and the target voltage characterizing the steady-state reference of the phase-aligned detectable main circuit voltage; obtaining multi-view features based on the differential current and the target voltage, wherein the multi-view features are used to jointly characterize the electrical characteristic deviation of the detectable electrical signal caused by the series arc fault from at least two dimensions, namely, time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution offset; and processing the multi-view features using a first feature extraction model to obtain a first feature. The first feature extraction model is trained using the multi-view features of the sample electrical signal. The sample electrical signal is the electrical signal obtained under the normal operation of the sample circuit. The sample circuit includes the sample main circuit and multiple parallel sample load branches. When the first arc fault detection result obtained by performing arc fault detection on the first feature indicates that the circuit under test has a series arc fault, the second feature obtained by processing the multi-view features using the second feature extraction model is used to perform arc fault detection. This yields a second arc fault detection result indicating that the circuit under test does not have a series arc fault or has a series arc fault in the parallel load branch. The second feature extraction model is trained using the abnormal sample multi-view features of the abnormal sample electrical signal. The abnormal sample electrical signal is the sample electrical signal corresponding to the sample arc fault detection result indicating that the sample circuit has a series arc fault.

[0005] Optionally, the multi-view features include at least two of the following multi-view sub-features: trajectory features, optical flow features, or expert features. The trajectory features characterize the temporal topological distortion of the electrical signal to be detected in the voltage-current plane, the optical flow features characterize the dynamic differences between adjacent cycles, and the expert features characterize the statistical distribution shift of the electrical signal to be detected. The first feature is obtained by processing the multi-view features using a first feature extraction model, including: processing the multi-view features using the encoder in the first feature extraction model to obtain view coding features corresponding to each of the multiple multi-view sub-features; and processing the view coding features corresponding to each of the multiple multi-view sub-features using the cross-view feature fusion module in the first feature extraction model to obtain the first feature. The first feature is characterized as an embedded representation of complementary information between different view features mapped to a low-dimensional discriminative representation space.

[0006] Optionally, the cross-view feature fusion module includes a feature fusion unit and a feature enhancement unit; wherein, the cross-view feature fusion module in the first feature extraction model processes the view coding features corresponding to each of the multiple multi-view sub-features to obtain a first feature, including: processing the view coding features corresponding to each of the multiple multi-view sub-features using the feature fusion unit to obtain a first fused feature, the first fused feature representing the common patterns and complementary fault features of the circuit operating state under different views; based on the first fused feature, processing the view coding features corresponding to each of the multiple multi-view sub-features using the feature enhancement unit to obtain an enhanced feature, the enhanced feature representing the context enhancement representation under the constraints of the global circuit feature structure relationship; and obtaining the first feature based on the enhanced feature.

[0007] Optionally, a first fused feature is obtained by processing the view coding features corresponding to each of the multiple multi-view sub-features using a feature fusion unit, including: performing feature fusion on the view coding features corresponding to each of the multiple multi-view sub-features using multiple projection parameters in the feature fusion unit to obtain the first fused feature; wherein, based on the first fused feature, an enhanced feature is obtained by processing the view coding features corresponding to each of the multiple multi-view sub-features using a feature enhancement unit, including: performing feature fusion on multiple view coding features using multiple enhancement parameters in the feature enhancement unit to obtain global structural relationship parameters, wherein the global structural relationship parameters characterize the structural relationship strength of the view coding features between different projection spaces; and performing feature enhancement on the first fused feature using the global structural relationship parameters to obtain the enhanced feature.

[0008] Optionally, based on the electrical signal to be detected from the circuit under test, differential current and target voltage are obtained, including: using the phase reference point obtained by detecting the phase reference point of the main voltage to be detected as the starting point of the cycle, performing cycle interception on the main current to be detected and the main voltage to be detected respectively to obtain phase-aligned current sequence and voltage sequence; obtaining differential current based on the difference between the reference current sequence and the subsequent current sequence, wherein the reference current sequence is obtained by phase alignment of the current obtained under stable load conditions, and the subsequent current sequence is obtained by phase alignment of the current obtained after the reference current sequence; and obtaining the target voltage based on the phase-aligned voltage sequence.

[0009] Optionally, based on the differential current and the target voltage, multi-view features are obtained, including: performing spectral separation on the differential current to obtain a first differential component and a second differential component, wherein the frequency of the second differential component is greater than the frequency of the first differential component; obtaining trajectory features based on the first differential component and the target voltage; performing optical flow estimation on the trajectory features to obtain optical flow features; obtaining expert features based on the first differential component and the second differential component; and obtaining multi-view features based on at least two of the trajectory features, optical flow features, or expert features.

[0010] Optionally, the differential current is subjected to spectral separation to obtain a first differential component and a second differential component, including: performing phase space reconstruction on the differential current to obtain reconstructed trajectory parameters, wherein the reconstructed trajectory parameters characterize the change in fault dynamic complexity of the circuit under the current operating state; extracting time-domain components from the reconstructed trajectory parameters to obtain multiple time-domain components; clustering the multiple time-domain components according to their respective spectral parameters to obtain a first cluster and a second cluster; and using the time-domain components in the first cluster as the first differential component and the time-domain components in the second cluster as the second differential component.

[0011] Optionally, the arc fault detection method further includes: splitting the first feature to obtain a first split feature and a second split feature; performing an affine transformation on the first split feature to obtain a first scale parameter and a first translation parameter, wherein the first scale parameter controls the scaling degree of the feature and the first translation parameter controls the offset degree of the feature; performing a feature transformation on the second split feature based on the first scale parameter and the first translation parameter to obtain an initial transformed feature; performing an affine transformation on the initial transformed feature to obtain a second scale parameter and a second translation parameter; performing a feature transformation on the first split feature based on the second scale parameter and the second translation parameter to obtain an intermediate transformed feature; fusing the initial transformed feature and the intermediate transformed feature to obtain a second fused feature; and obtaining a first arc fault detection result based on the second fused feature.

[0012] Optionally, the multi-view features of the samples are obtained based on the sample differential current and sample voltage, which are obtained based on the sample electrical signals of the sample circuit. The sample electrical signals include the sample main current and sample main voltage. The sample differential current characterizes the difference between cycles of the phase-aligned sample main current to reflect the waveform distortion and non-stationary evolution caused by the series arc fault. The sample voltage characterizes the steady-state reference of the phase-aligned sample main voltage. The multi-view features of the samples are used to jointly characterize the electrical characteristic deviation of the series arc fault on the sample electrical signal from at least two dimensions: time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution shift. The multi-view features of the samples include at least two sample multi-view sub-features. The first feature extraction model is trained as follows: the sample multi-view features of the sample electrical signal are processed using a first deep learning model to obtain the first sample decoding features, the first sample global structure relationship parameters, the first sample features, and the first sample view encoding features corresponding to each of the multiple sample multi-view sub-features; and based on the first objective loss function, according to the first sample global structure relationship... The first feature extraction model is trained by using the parameters, the first sample features, and the first sample view encoding features corresponding to each of the multiple sample multi-view sub-features, as well as the reconstruction loss function value determined based on the sample multi-view features and the first sample decoding features. The first feature extraction model is obtained by training a first deep learning model, where the abnormal sample multi-view features include at least two abnormal sample multi-view sub-features. The second feature extraction model is trained by using the second deep learning model to extract features from the abnormal sample multi-view features of the abnormal sample electrical signal, obtaining the second sample decoding features, the second sample global structural relationship parameters, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features. The second feature extraction model is then trained based on the second target loss function, the contrast loss function value determined based on the second sample global structural relationship parameters, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features, and the reconstruction loss function value determined based on the abnormal sample multi-view features and the second sample decoding features.

[0013] Optionally, based on the first objective loss function, a first deep learning model is trained to obtain a first feature extraction model, which includes: training a first deep learning model based on the first objective loss function, the contrast loss function value determined by the first sample global structural relationship parameters, the first sample features, and the first sample view encoding features corresponding to each of the multiple sample multi-view sub-features, and the reconstruction loss function value determined by the sample multi-view features and the first sample decoding features; wherein, based on the second objective loss function, a second deep learning model is trained to obtain a second feature extraction model, which includes: training a first deep learning model based on the first objective loss function, the contrast loss function value determined by the first sample features and the multiple first sample view encoding features, the first sample global structural relationship parameters, and the reconstruction loss function value determined by the sample multi-view features and the first sample decoding features; and training a second deep learning model based on the second objective loss function, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features, and the reconstruction loss function value determined by the abnormal sample multi-view features and the second sample decoding features; and the second feature extraction model includes: training a second deep learning model based on the sample data obtained by clustering analysis of the second sample features. The arc fault detection results are quality assessed to obtain sample quality assessment results. If the sample quality assessment results are satisfactory, a second deep learning model is trained based on the second objective loss function, the confidence level of the sample arc fault detection results, the distance between the second sample features and positive sample features, the distance between the second sample features and negative sample features (triple loss function values), the contrast loss function values ​​determined by the second sample global structural relationship parameters, the second sample features, and the second sample view encoding features corresponding to the multi-view sub-features of multiple abnormal samples, and the reconstruction loss function values ​​determined by the multi-view features of abnormal samples and the decoding features of the second samples. This results in a second feature extraction model. Positive sample features are those second sample features whose distance from the second sample feature is greater than or equal to a first predetermined distance among multiple other second sample features with the same arc fault detection results as the second sample feature. Negative sample features are those second sample features whose distance from the second sample feature is less than or equal to a second predetermined distance among multiple other second sample features with different arc fault detection results from the second sample feature. The first predetermined distance is greater than the second predetermined distance.

[0014] A third aspect of this application provides an arc fault detection device, comprising: a first extraction module, configured to obtain differential current and target voltage based on the electrical signal to be detected from a circuit to be detected, wherein the circuit to be detected includes a main circuit and multiple parallel load branches, the electrical signal to be detected includes a main circuit current to be detected and a main circuit voltage to be detected, the differential current characterizing the difference between cycles of the main circuit current to be detected after phase alignment, so as to reflect the waveform distortion and non-stationary evolution caused by a series arc fault, and the target voltage characterizing the steady-state reference of the main circuit voltage to be detected after phase alignment; a second extraction module, configured to obtain multi-view features based on the differential current and target voltage, wherein the multi-view features are used to jointly characterize the electrical characteristic deviation of the electrical signal to be detected caused by the series arc fault from at least two dimensions, namely, time-domain topological distortion, dynamic difference between cycles, or statistical distribution offset; and a first processing module, configured to process the multi-view features using a first feature extraction model. The system obtains a first feature, which is trained using the multi-view features of the sample electrical signal. The sample electrical signal is obtained under normal operation of the sample circuit, which includes the sample main circuit and multiple parallel sample load branches. The system also includes a second processing module, used to perform arc fault detection on the second feature obtained by processing the multi-view features using the second feature extraction model when the first arc fault detection result indicates that the circuit under test has a series arc fault. This results in a second arc fault detection result indicating that the circuit under test does not have a series arc fault or has a parallel load branch with a series arc fault. The second feature extraction model is trained using the multi-view features of the abnormal sample electrical signal, which is the sample electrical signal corresponding to the sample arc fault detection result indicating that the sample circuit has a series arc fault.

[0015] A third aspect of this application provides an electronic device, including: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the above-described arc fault detection method.

[0016] A fourth aspect of this application provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0017] A fifth aspect of this application provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps of the above-described method.

[0018] According to the arc fault detection method provided in this application, since the first feature extraction model is trained from the multi-view features of normal samples, it can complete the abnormal detection of series arc faults without relying on arc fault labels, reducing the dependence on arc fault labels and enabling effective detection of series arc fault anomalies under zero fault sample conditions. The first stage filters out non-fine-grained arc fault samples that indicate whether there is an arc fault in the main circuit. The second feature extraction model is trained using the non-fine-grained arc fault samples filtered out during the training of the first feature extraction model, realizing the identification of parallel branches within abnormal arc faults. No manual labeling of fine-grained arc faults is required throughout the process. Therefore, it has the advantages of requiring no arc fault samples, requiring no fine-grained labels, high detection accuracy, and good engineering application feasibility, making it suitable for intelligent electrical safety protection in multi-load parallel low-voltage power distribution scenarios. Furthermore, through the differential current determined based on the inter-cycle difference of the main current to be detected based on phase alignment, the background load component in the main current to be detected is further processed... Line suppression can effectively enhance the weak distortion characteristics caused by series arc faults and reduce the flooding effect of the masking load on the arc fault characteristics in multi-load parallel scenarios. Therefore, the multi-view features extracted based on differential current and target voltage can jointly characterize electrical feature deviations from at least two dimensions, such as time-domain topological distortion, inter-cycle dynamic differences, or statistical distribution offset. This makes the arc fault feature expression complementary and robust, reducing the detection failure of a single view dimension under complex load conditions, thereby improving the detectability of abnormal features of series arc faults under complex load conditions. In the first stage, the first feature extraction model is used to perform anomaly detection on the multi-view features. If the first arc fault detection result indicates that the circuit under test has a series arc fault, the second stage uses the second feature extraction model to perform series arc fault detection on the multi-view features again. This can further identify whether there are false detections in the first stage or identify parallel load branches with series arc faults, thereby improving the identification accuracy of series arc fault branches and the overall detection reliability. Attached Figure Description

[0019] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments of this application with reference to the accompanying drawings.

[0020] Figure 1 An application scenario of the arc fault detection method according to an embodiment of this application is shown.

[0021] Figure 2 A flowchart of an arc fault detection method according to an embodiment of this application is shown.

[0022] Figure 3A A schematic diagram of a first-stage anomaly detection model according to an embodiment of this application is shown.

[0023] Figure 3BA schematic diagram of a first deep learning model according to an embodiment of this application is shown.

[0024] Figure 4A A schematic diagram illustrating the training effect of the first-stage anomaly detection model according to an embodiment of this application is shown.

[0025] Figure 4B A schematic diagram of clustering second sample features according to an embodiment of this application is shown.

[0026] Figure 5 A block diagram of an electronic device suitable for implementing an arc fault detection method according to an embodiment of this application is shown. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] Arc faults can be classified into three types: parallel arcs, grounding arcs, and series arcs. Compared to the former two, series arc faults, due to the physical limitation of the downstream series load impedance, generate smaller fault currents, making it difficult to trigger overcurrent protection devices or circuit breakers, thus exhibiting strong concealment. Furthermore, the electrical characteristics of series arc faults are highly influenced by the modulation effect of the series load in the branch they occur in. With the increasing variety of electrical loads in modern buildings, accurately extracting and identifying the dynamic characteristics of arc faults has become increasingly challenging.

[0032] Series arc fault detection methods revolve around the time-domain, frequency-domain, or time-frequency-domain feature analysis of current signals. However, in actual low-voltage distribution networks, complex topologies with multiple load branches operating in parallel are common. In such circuit structures, the main current (i.e., the current in the main circuit) is a physical superposition of multiple load branch currents (i.e., the current in the load branches). When a series arc fault occurs in a load branch, the resulting weak distortions such as high-frequency pulses, waveform asymmetry, and current flattening are easily overwhelmed by the current generated by other normally operating high-power masking loads. This feature masking effect causes the fault current in the main circuit to tend to be similar to the normal current in terms of waveform morphology and statistical distribution. This makes series arc fault detection methods based on predetermined thresholds or manually extracted features have a high risk of failure under complex conditions, leading to missed or false alarms.

[0033] In recent years, end-to-end arc fault detection methods based on deep learning models have been introduced into the field of arc fault detection. However, these methods face data acquisition bottlenecks when deployed in practical engineering projects. Their model construction relies on large-scale supervised learning mechanisms, requiring joint parameter optimization using a large number of pre-labeled normal samples and arc fault samples. While collecting normal samples is feasible in real-world environments, artificially simulating and continuously collecting real arc fault samples poses safety risks and lacks practicality in engineering implementation. This leads to arc fault detection models needing to cope with extreme conditions with zero fault samples when deployed in unfamiliar scenarios. Furthermore, the generalization ability of supervised arc fault detection models will decline when the data distribution in actual deployment scenarios deviates from the distribution of prior laboratory databases. In addition to basic arc detection, fine-grained identification of the branch where the arc fault occurs is crucial for subsequent fault isolation and accurate safety maintenance of the power distribution system. Given the difficulty of acquiring arc fault samples in actual deployment scenarios, pre-collecting various load fault labels for training fine-grained branch identification models faces extremely high practical barriers. Therefore, how to effectively overcome the interference of multi-load background current under actual deployment conditions where there are no real arc fault samples and no fine-grained category labels, and thus simultaneously achieve high-precision series arc fault screening and fine-grained load category identification, has become a technical challenge that urgently needs to be overcome in the field of low-voltage electrical safety protection.

[0034] In view of this, embodiments of this application provide an arc fault detection method to solve the technical problems of series arc fault characteristics being submerged by the background current of the main road in multi-parallel load branch scenarios, difficulty in obtaining arc fault samples in actual deployment, and difficulty in obtaining fine-grained arc fault labels for each branch to train the arc fault detection model.

[0035] First, addressing the technical challenge of series arc fault characteristics being overwhelmed by the background current of the main circuit in multi-parallel load branch scenarios, a differential current preprocessing strategy based on dynamic baseline subtraction is proposed. This strategy detects load switching events and selects a steady-state cycle as a benchmark, then aligns and differentially processes subsequent monitoring cycles to enhance the subtle distortion characteristics caused by the series arc. Furthermore, to balance the different spectral information requirements of fault detection and fault load identification, an adaptive spectrum separation strategy is proposed. This strategy decomposes the differential current into a low-frequency first differential component and a high-frequency second differential component, thereby highlighting the high-frequency disturbance characteristics caused by the arc while preserving load-specific information.

[0036] Furthermore, a multi-view feature extraction strategy is proposed, which jointly characterizes the electrical feature deviations of the detected electrical signal caused by the series arc fault from at least two dimensions, including temporal topological distortion, inter-cycle dynamic differences, or statistical distribution shifts, thus depicting the characteristics of the series arc fault from different perspectives. Further, an unsupervised deep learning model is constructed, which maps the multi-view features to a unified low-dimensional discriminative representation space by jointly constraining cross-view feature consistency and sample structure relationships, thereby obtaining a highly discriminative feature representation suitable for subsequent arc fault detection and fault branch identification tasks.

[0037] Furthermore, the model training process employs a two-stage processing mechanism. In the first stage, a first deep learning model is trained using normal samples to obtain a first feature extraction model, and a density estimation network based on normalized flow is trained using the output features of the first samples to obtain a first-stage anomaly detection model. No fault samples are introduced during this training stage. In the second stage, a second deep learning model is trained using the unlabeled abnormal arc samples detected in the first stage to obtain a second feature extraction model. Clustering pseudo-label generation and self-supervised iterative optimization are then performed on the second sample features to achieve series arc fault load identification and correct false detection samples from the first stage. Therefore, without requiring fault samples and fine-grained fault branch labels for supervised training, effective detection and branch identification of series arc faults in multi-parallel load branch scenarios are achieved.

[0038] Figure 1 An application scenario of the arc fault detection method according to an embodiment of this application is shown.

[0039] like Figure 1As shown, the business system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0040] Users can interact with server 105 via network 104 using at least one of the first terminal device 101, second terminal device 102, and third terminal device 103 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, second terminal device 102, and third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0041] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0042] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). Server 105 includes multiple server nodes that can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0043] It should be noted that the arc fault detection method provided in this application embodiment can be executed by the first terminal device 101, the second terminal device 102 or the third terminal device 103, or it can be executed by other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0044] Alternatively, the arc fault detection method provided in this application embodiment can also be executed by server 105. The arc fault detection method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and is capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or server 105.

[0045] Furthermore, the training methods of the first feature extraction model and the second feature extraction model described in the embodiments of this application may be the same as or different from the execution entities of the arc fault detection method described in the embodiments of this application, and the embodiments of this application do not limit this.

[0046] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0047] It should be noted that the sequence numbers of the operations in the following methods are for descriptive purposes only and should not be considered as indicating the execution order of the operations. Unless explicitly stated otherwise, the method does not need to be executed in the exact order shown.

[0048] Figure 2 A flowchart of an arc fault detection method according to an embodiment of this application is shown.

[0049] like Figure 2 As shown, the arc fault detection method includes operations S210 to S240.

[0050] In operation S210, the differential current and target voltage are obtained based on the electrical signal to be detected from the circuit under test.

[0051] In operation S220, multi-view features are obtained based on differential current and target voltage. These multi-view features are used to jointly characterize the electrical feature deviation of the electrical signal to be detected caused by the series arc fault from at least two dimensions, including time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution offset.

[0052] In operation S230, the first feature extraction model is used to process the multi-view features to obtain the first feature. The first feature extraction model is trained using the multi-view features of the sample electrical signal. The sample electrical signal is the electrical signal obtained under the normal operation of the sample circuit. The sample circuit includes the sample main circuit and multiple parallel sample load branches.

[0053] In operation S240, when the first arc fault detection result obtained by performing arc fault detection on the first feature indicates that the circuit under test has a series arc fault, the second feature obtained by processing the multi-view features using the second feature extraction model is subjected to arc fault detection to obtain a second arc fault detection result indicating that the circuit under test does not have a series arc fault or has a series arc fault.

[0054] The arc fault detection method is designed to detect series arcs.

[0055] The circuit under test includes a main circuit and multiple parallel load branches, and the electrical signals under test include the main circuit current and the main circuit voltage under test.

[0056] Users can use 220V / 50Hz AC power supply. Multiple parallel load branches are connected to the main circuit in parallel. The parallel load branches are the power load circuits of the electrical equipment. The main current and voltage to be detected are collected by the data acquisition device on the main circuit.

[0057] For example, if User 1's electrical appliances include air conditioners, electric fans, refrigerators, hair dryers, etc., then the parallel load branches corresponding to each air conditioner, electric fan, refrigerator, and hair dryer are connected in parallel to the main circuit, and the electrical signal to be detected is collected from the data acquisition device at the power supply inlet of User 1.

[0058] Load switching events are detected in the main current to be tested. After a load switching event is detected, a reference cycle that has entered steady state is selected as the dynamic baseline. The subsequent cycles after the dynamic baseline are phase-aligned and differentially processed with the corresponding main currents to be tested to obtain differential currents. The differential current characterizes the difference between cycles of the phase-aligned main current to be tested, reflecting the waveform distortion and non-stationary evolution caused by series arc faults.

[0059] A series arc fault is an arc discharge phenomenon caused by abnormal current conduction within the parallel load branch where the electrical equipment is located.

[0060] Load switching event detection identifies abrupt changes in the amplitude, phase, or harmonic characteristics of the main current under test caused by equipment start-up / shutdown, speed adjustment, or circuit topology changes. After detecting a load switching event, the first complete cycle after the load enters steady state is selected as the dynamic baseline, which reflects the normal current waveform characteristics under the current load combination. Differential current characterizes the difference between cycles of the main current under test after phase alignment, which can amplify and highlight the waveform distortion and non-stationary evolution caused by series arc faults.

[0061] Under normal conditions, the subsequent cycles tend to stabilize after the dynamic baseline is established. However, after an arc fault occurs, the subsequent cycles exhibit continuous non-stationary distortion at the dynamic baseline. Therefore, under normal conditions, the differential current mainly exhibits small fluctuations near zero. However, the differential current after a series arc fault disturbance will exhibit obvious waveform distortion and high-frequency pulse characteristics.

[0062] The subsequent cycles following the dynamic baseline are phase-aligned with the corresponding target backbone voltages of the dynamic baseline. The target voltage is then obtained by averaging and summing the phase-aligned subsequent cycles and the target backbone voltages of the dynamic baseline. The target voltage represents the steady-state reference of the phase-aligned target backbone voltage.

[0063] Adaptive spectrum separation is performed on the differential current to obtain low-frequency normal differential components that characterize load morphology information and high-frequency normal differential components that characterize disturbance information; based on the low-frequency normal differential components, high-frequency normal differential components and target voltage, multi-view features are constructed.

[0064] The first feature is obtained by processing multi-view features using a first feature extraction model. The first feature is an embedding representation that maps complementary information between different view sub-features into a low-dimensional discriminative representation space.

[0065] The first feature extraction model can be based on multi-view... Figure 1 Construction of Consistent Representation Learning Networks, Multi-view Figure 1 The consistent representation learning network consists of a convolutional encoder, a fully connected encoder, a feature aggregation module, a feature enhancement module, and a multilayer perceptron.

[0066] Arc fault detection is performed on the first feature using a density estimation model based on normalized flow, yielding a first arc fault detection result. The first arc fault detection result indicates either the presence of a series arc fault in the circuit under test or the absence of a series arc fault in the circuit under test.

[0067] If the initial arc fault detection result indicates a series arc fault in the circuit under test, further investigation is needed to determine whether a false detection occurred or if the series arc fault is located in the parallel load branch. The second feature is obtained by processing the multi-view features using a second feature extraction model.

[0068] The second feature is a low-dimensional electrical feature that characterizes the multi-view features of the abnormal circuit under test.

[0069] Clustering algorithms can be used to detect arc faults based on the second feature, thus obtaining the second arc fault detection result.

[0070] For example, the second arc fault detection result can be 0, 1, 2, or 3. Among them, 0 represents that there is no series arc fault in the circuit under test, that is, the first feature extraction model has a false detection; 1 represents that there is a series arc fault in the parallel load branch of the air conditioner; 2 represents that there is a series arc fault in the parallel load branch of the electric fan; and 3 represents that there is a series arc fault in the parallel load branch of the refrigerator.

[0071] The first and second feature extraction models have the same model structure but different model parameters.

[0072] The first feature extraction model utilizes the multi-view features of the sample electrical signal for unsupervised training of the original multi-view model. Figure 1 The consistency representation learning network obtains the consistent intrinsic structure of the multi-view features of the electrical signals of normal samples in unsupervised learning, which serves as a reference benchmark for anomaly detection during the model inference stage.

[0073] The sample electrical signal is the electrical signal acquired under normal operating conditions of the sample circuit. During the training process, the model outputs the first sample feature, and the first sample feature is used to perform arc fault detection to obtain the first arc fault detection result of the sample.

[0074] The abnormal sample electrical signal characterization indicates that the sample circuit has a series arc fault, as shown by the first arc fault detection result. The abnormal sample multi-view features are the multi-view characteristics of the abnormal sample electrical signal. For example, the first arc fault detection result is 0 or 1, where 0 represents that the sample circuit does not have a series arc fault, and 1 represents that the sample circuit has a series arc fault.

[0075] The second feature extraction model utilizes the abnormal sample electrical signal features of abnormal sample multi-views for unsupervised training of the original multi-view. Figure 1 The consistency representation learning network obtains the consistent intrinsic structure of the multi-view features of unsupervised learning abnormal sample electrical signals, which serves as a pattern library for arc fault identification and fine detection during the model inference stage.

[0076] Optionally, since the first feature extraction model is trained from the multi-view features of normal samples, it can complete the anomaly detection of series arc faults without relying on arc fault labels, reducing the dependence on arc fault labels and enabling effective detection of series arc fault anomalies under zero fault sample conditions. The first stage filters out non-fine-grained arc fault samples that indicate whether there is an arc fault in the main circuit. The second feature extraction model is trained using the non-fine-grained arc fault samples filtered out in the training of the first feature extraction model, realizing the identification of parallel branches within the abnormal arc fault. No manual labeling of fine-grained arc faults is required throughout the process. Therefore, it has the advantages of not requiring arc fault samples, not requiring fine-grained labels, high detection accuracy, and good engineering application feasibility, making it suitable for intelligent electrical safety protection in multi-load parallel low-voltage power distribution scenarios. Furthermore, by using the differential current determined based on the inter-cycle difference of the main current to be detected based on phase alignment, the differential current suppresses the background load component in the main current to be detected, thus achieving… This method effectively enhances the subtle distortion characteristics caused by series arc faults, mitigating the overwhelming effect of masking loads on arc fault features in multi-load parallel scenarios. Based on differential current and target voltage extraction, multi-view features can jointly characterize electrical feature deviations from at least two dimensions: time-domain topological distortion, inter-cycle dynamic differences, or statistical distribution offsets. This makes the arc fault feature representation complementary and robust, reducing detection failures of single-view dimensions under complex load conditions, thereby improving the detectability of abnormal features of series arc faults under complex load conditions. In the first stage, the first feature extraction model is used to detect anomalies in the multi-view features. If the first arc fault detection result indicates the presence of a series arc fault in the circuit under test, the second stage uses a second feature extraction model to further detect series arc faults in the multi-view features. This further allows for the identification of false detections in the first stage or the identification of parallel load branches with series arc faults, thereby improving the identification accuracy of series arc fault branches and the overall detection reliability.

[0077] Optionally, based on the electrical signal to be detected from the circuit under test, differential current and target voltage are obtained, including: using the phase reference point obtained by detecting the phase reference point of the main voltage to be detected as the starting point of the cycle, performing cycle interception on the main current to be detected and the main voltage to be detected respectively to obtain phase-aligned current sequence and voltage sequence; obtaining differential current based on the difference between the reference current sequence and the subsequent current sequence, wherein the reference current sequence is obtained by phase alignment of the current obtained under stable load conditions, and the subsequent current sequence is obtained by phase alignment of the current obtained after the reference current sequence; and obtaining the target voltage based on the phase-aligned voltage sequence.

[0078] The main voltage and main current to be detected are the raw voltage and raw current collected from the data source.

[0079] The phase reference point is a characteristic moment point extracted from the voltage waveform to identify the start of the cycle. It is usually taken as the positive zero-crossing point or peak point of the fundamental voltage.

[0080] Because the actual frequency of the power grid has a slight drift, the number or length of sampling points for different cycles may be different. Phase alignment is achieved by adjusting each cycle to the same phase grid through interpolation or resampling. For example, each cycle has a fixed 128 sampling points.

[0081] The phase reference point of the main voltage to be detected is detected in real time, and the zero-crossing point of the fundamental wave is identified as the starting point of the cycle. Based on the starting point of the cycle, the main voltage to be detected at the same time is truncated into cycles to obtain data segments of equal period length. Phase alignment is achieved by interpolation and resampling to obtain phase-aligned current and voltage sequences, so that the cycles collected at different times are strictly corresponding in phase coordinates.

[0082] During periods of stable system load, the phase-aligned current sequence is used as the reference current sequence, and the current sequence following the reference current sequence is used as the subsequent current sequence. The two are subtracted point by point to complete the differential processing and output the differential current. The average of the synchronously aligned reference voltage sequence and the subsequent voltage sequence is used as the target voltage output for subsequent multi-view feature construction.

[0083] Differential current is the result of subtracting the reference current sequence from the subsequent current sequence at each sampling point, and is used to strip away the normal load background and highlight the characteristics of abnormal electrical signals.

[0084] The target voltage is a smooth and neutral reference voltage waveform. Since the main voltage to be detected may have slight fluctuations, the mean of the aligned reference voltage sequence and subsequent voltage sequences is used as the voltage reference axis in subsequent multi-view feature extraction.

[0085] By employing a current preprocessing strategy that differentially processes the reference current sequence and subsequent current sequences, the background load component in the main current to be detected is suppressed. This effectively enhances the weak distortion characteristics caused by series arcs, reduces the flooding effect of the masking load on fault characteristics in multi-load parallel scenarios, and thus improves the detectability of abnormal series arc characteristics under complex multi-equipment loads.

[0086] Optionally, based on the differential current and the target voltage, multi-view features are obtained, including: performing spectral separation on the differential current to obtain a first differential component and a second differential component, wherein the frequency of the second differential component is greater than the frequency of the first differential component; obtaining trajectory features based on the first differential component and the target voltage; performing optical flow estimation on the trajectory features to obtain optical flow features; obtaining expert features based on the first differential component and the second differential component; and obtaining multi-view features based on at least two of the trajectory features, optical flow features, or expert features.

[0087] The differential current is spectrally separated using the symplectic geometrical mode decomposition algorithm to obtain a first differential component and a second differential component. The second differential component is a high-frequency differential component, and the first differential component is a low-frequency differential component.

[0088] The target voltage and the first differential component within one cycle are normalized. The normalized target voltage is used as the abscissa and the normalized first differential component is used as the ordinate. Adjacent sampling points are connected sequentially to form a voltage-current trajectory. When the distance between adjacent sampling points is greater than the grid resolution, linear interpolation is performed between adjacent sampling points, and the grid cells through which the trajectory passes are assigned a value of 1, thus obtaining a binary trajectory diagram of the voltage-current trajectory.

[0089] To enhance the ability of the trajectory diagram to represent the characteristics of physical electrical signals, color coding is applied to the trajectory diagram. The color coding uses a color space (Hue Saturation Value, HSV).

[0090] The hue H is used to characterize the direction of motion of the trajectory. The hue value of the grid cell is obtained by averaging the corresponding directions of the sampling points falling within the same grid cell.

[0091] The saturation S is used to characterize the proportion of the first differential component relative to the effective value of the differential current.

[0092] The brightness V is used to characterize the smoothness of the first difference component, which is determined by the rate of change of adjacent sampling points. The smoothness of each sampling point falling within the same grid cell is averaged to obtain the hue value of that grid cell. The hue, saturation, and brightness are then concatenated along the channel dimension to obtain the trajectory feature.

[0093] Trajectory features characterize the morphological and structural information of voltage-current loads.

[0094] A polynomial extended algorithm is used to calculate the optical flow displacement vector between the trajectory features of adjacent cycles. Based on the optical flow displacement vector field, the direction of the displacement vector of each pixel in the trajectory feature is mapped to the hue component in the color space, the amplitude of the displacement vector is mapped to the saturation component, and the lightness component is configured to a set maximum threshold. The components of each channel are fused to generate the optical flow feature. The optical flow feature characterizes the dynamic changes between the trajectory features of adjacent cycles.

[0095] The expert features include amplitude features extracted based on the first difference component and statistical features extracted based on the second difference component.

[0096] In one embodiment, amplitude features As shown in formula (1):

[0097] (1).

[0098] in, Characterizes the square of the j-th first difference component. It represents the number of the first difference components.

[0099] The peak value, standard deviation, kurtosis, skewness, envelope entropy, waveform factor, peak factor, impulse factor, and margin factor of the second difference component are extracted as statistical features using statistical functions.

[0100] Trajectory features, optical flow features, or expert features are stitched together to obtain multi-view features.

[0101] Optionally, the differential current is subjected to spectral separation to obtain a first differential component and a second differential component, including: performing phase space reconstruction on the differential current to obtain reconstructed trajectory parameters, wherein the reconstructed trajectory parameters characterize the change in fault dynamic complexity of the circuit under the current operating state; extracting time-domain components from the reconstructed trajectory parameters to obtain multiple time-domain components; clustering the multiple time-domain components according to their respective spectral parameters to obtain a first cluster and a second cluster; and using the time-domain components in the first cluster as the first differential component and the time-domain components in the second cluster as the second differential component.

[0102] The differential current is reconstructed in phase space to build a trajectory matrix. In one embodiment, the trajectory matrix... As shown in formula (2):

[0103] (2).

[0104] Among them, differential current r represents the embedding dimension. Indicates the delay time. .

[0105] Constructing the covariance matrix based on the trajectory matrix Constructing the Hamiltonian matrix based on the covariance matrix. Where the superscript T denotes matrix transpose, and 0 denotes the result of the matrix transpose and covariance matrix transpose. The zero matrix of dimension matching.

[0106] Calculate the square of the Hamiltonian matrix to obtain the square matrix. Perform a symplectic similarity transformation on the square matrix to construct a symplectic orthogonal matrix. .

[0107] In one embodiment, the symplectic similarity transformation process is shown in formula (3):

[0108] (3).

[0109] Where L represents the residual block matrix, Let represent an upper triangular matrix. The eigenvalues ​​of an upper triangular matrix can be represented as ( According to the properties of the Hamiltonian matrix, the covariance matrix... The eigenvalues ​​can be expressed as: .

[0110] The i-th column vector of a symplectic orthogonal matrix is ​​denoted as , Covariance matrix Eigenvectors corresponding to eigenvalues Construct reconstructed trajectory parameters based on the i-th column vector and the trajectory matrix. .

[0111] The reconstructed trajectory parameters characterize the changes in fault dynamic complexity of the circuit under the current operating state.

[0112] Diagonal averaging of the reconstructed trajectory parameters yields multiple time-domain components.

[0113] To characterize the frequency distribution center and dispersion of each time-domain component, the spectral centroid and frequency variance of each time-domain component are calculated respectively. The spectral centroid and frequency variance of each time-domain component are used to form a feature vector. Based on the feature vector, multiple time-domain components are clustered to obtain the first cluster and the second cluster.

[0114] The time-domain components in the first cluster with the lower spectral centroid are used as the first difference components, and the time-domain components in the second cluster are used as the second difference components.

[0115] The differential current is decomposed into a low-frequency first differential component and a high-frequency second differential component. The first differential component retains the structural information related to the low-frequency load configuration, while the second differential component highlights the pulse and statistical characteristics related to arc fault disturbances. This approach takes into account the different requirements of arc fault detection and branch identification for spectral information, thereby improving the pertinence and effectiveness of feature representation.

[0116] Optionally, the multi-view features include at least two of the following multi-view sub-features: trajectory features, optical flow features, or expert features. The trajectory features characterize the temporal topological distortion of the electrical signal to be detected in the voltage-current plane, the optical flow features characterize the dynamic differences between adjacent cycles, and the expert features characterize the statistical distribution shift of the electrical signal to be detected. The first feature is obtained by processing the multi-view features using a first feature extraction model, including: processing the multi-view features using the encoder in the first feature extraction model to obtain view coding features corresponding to each of the multiple multi-view sub-features; and processing the view coding features corresponding to each of the multiple multi-view sub-features using the cross-view feature fusion module in the first feature extraction model to obtain the first feature. The first feature is characterized as an embedding representation of complementary information between different view sub-features mapped to a low-dimensional discriminative representation space.

[0117] The first feature extraction model consists of two convolutional encoders and one fully connected encoder. The encoder is used to map multiple multi-view sub-features to a latent space of dimension 256.

[0118] A convolutional encoder is used to extract features from trajectory features to obtain view coding features corresponding to the trajectory features; a convolutional encoder is used to extract features from optical flow features to obtain view coding features corresponding to the optical flow features; a fully connected encoder is used to extract features from expert features to obtain view coding features corresponding to the expert features.

[0119] The cross-view feature fusion module can be constructed based on the Global and Cross-view Feature Aggregation (GCFAgg) network.

[0120] The cross-view feature fusion module maps the view encoding features corresponding to each of the multiple multi-view sub-features to a unified and consistent representation space to obtain the first feature.

[0121] Optionally, the cross-view feature fusion module includes a feature fusion unit and a feature enhancement unit; wherein, the cross-view feature fusion module in the first feature extraction model processes the view coding features corresponding to each of the multiple multi-view sub-features to obtain a first feature, including: processing the view coding features corresponding to each of the multiple multi-view sub-features using the feature fusion unit to obtain a first fused feature, the first fused feature representing the common patterns and complementary fault features of the circuit operating state under different views; based on the first fused feature, processing the view coding features corresponding to each of the multiple multi-view sub-features using the feature enhancement unit to obtain an enhanced feature, the enhanced feature representing the context enhancement representation under the constraints of the global circuit feature structure relationship; and obtaining the first feature based on the enhanced feature.

[0122] Since the view encoding features corresponding to each of the multiple multi-view sub-features may still contain view-private information that interferes with the detection of series arc faults, the structural relationship between different views is jointly modeled by the feature fusion unit.

[0123] The cross-view feature fusion module includes a feature fusion unit and a feature enhancement unit.

[0124] In one embodiment, the first fusion feature R is as shown in formula (4):

[0125] (4).

[0126] in, This represents the view encoding feature corresponding to each of the p-th multi-view sub-features. This represents the projection parameter corresponding to the p-th multi-view sub-feature. The projection parameter is a learnable parameter in the feature fusion unit.

[0127] The first fusion feature characterizes the common patterns and complementary fault characteristics of the circuit's operating state under different views.

[0128] The view encoding features corresponding to each of the multiple multi-view sub-features are concatenated to obtain the concatenated features.

[0129] Based on the first fusion feature, the splicing feature is enhanced using a feature enhancement unit to obtain the enhanced feature.

[0130] To remove redundant information introduced by concatenating view encoding features from multiple views, a multilayer perceptron is used to map the enhanced features to obtain the first feature. The multilayer perceptron employs a three-layer fully connected structure.

[0131] In one embodiment, the first feature As shown in formula (5):

[0132] (5).

[0133] in, , , as well as , , Z represents the learnable parameters of the multilayer perceptron, and Z is the concatenated feature determined by the view encoding features corresponding to each of the multiple multi-view sub-features. To enhance features.

[0134] By constructing multi-view features such as trajectory features, optical flow features, and expert features, and using the first feature extraction model to jointly model the multi-view features, complementary information in different views is mapped to a unified low-dimensional discriminative representation space. This can enhance the consistency of fault feature expression of the electrical signal under different views, suppress the interference of view-private redundant information on the detection results, and thus improve the discriminativeness and robustness of series arc fault feature representation.

[0135] Optionally, a first fused feature is obtained by processing the view coding features corresponding to each of the multiple multi-view sub-features using a feature fusion unit, including: performing feature fusion on the view coding features corresponding to each of the multiple multi-view sub-features using multiple projection parameters in the feature fusion unit to obtain the first fused feature; wherein, based on the first fused feature, an enhanced feature is obtained by processing the view coding features corresponding to each of the multiple multi-view sub-features using a feature enhancement unit, including: performing feature fusion on multiple view coding features using multiple enhancement parameters in the feature enhancement unit to obtain global structural relationship parameters, wherein the global structural relationship parameters characterize the structural relationship strength of the view coding features between different projection spaces; and performing feature enhancement on the first fused feature using the global structural relationship parameters to obtain the enhanced feature.

[0136] The view encoding features corresponding to each of the multiple multi-view sub-features are concatenated to obtain the concatenated features.

[0137] In one embodiment, the global structural relationship parameters are as shown in formula (6):

[0138] (6).

[0139] Where softmax represents the normalization function. , , For two enhancement parameters, Z represents the learning parameters, and Z represents the concatenation features.

[0140] The enhancement parameter represents the auxiliary projection matrix in the feature enhancement unit. The global structural relationship parameter is used to characterize the strength of the structural relationship between the stitched features determined based on multiple view encoded features in different projection spaces.

[0141] In one embodiment, the enhancement feature is as shown in formula (7):

[0142] (7).

[0143] Optionally, the method further includes: splitting the first feature to obtain a first split feature and a second split feature; performing an affine transformation on the first split feature to obtain a first scale parameter and a first translation parameter, wherein the first scale parameter controls the scaling degree of the feature and the first translation parameter controls the offset degree of the feature; performing a feature transformation on the second split feature based on the first scale parameter and the first translation parameter to obtain an initial transformed feature; performing an affine transformation on the initial transformed feature to obtain a second scale parameter and a second translation parameter; performing a feature transformation on the first split feature based on the second scale parameter and the second translation parameter to obtain an intermediate transformed feature; fusing the initial transformed feature and the intermediate transformed feature to obtain a second fused feature; and obtaining a first arc fault detection result based on the second fused feature.

[0144] The first-stage anomaly detection model is used to detect faults in the first feature. This model is constructed based on a normalized flow density estimation network. The normalized flow density estimation network is used to establish a bijective mapping relationship from the first sample feature to the standardized latent space under the condition of training only using sample electrical signals, so that the output probability density follows a standard normal distribution.

[0145] Figure 3A A schematic diagram of a first-stage anomaly detection model according to an embodiment of this application is shown.

[0146] like Figure 3A As shown, the first-stage anomaly detection model consists of multiple affine coupling blocks. Within each affine coupling block, the first feature is first permuted and divided into first split features. Second splitting features The first splitting feature is processed by a fully connected network to obtain affine transformation parameters, and the output is divided into first scale parameters. and the first translation parameter The affine coupling block includes two affine coupling sub-transformation modules. In the first affine coupling sub-transformation module, the second split feature is transformed according to the first scale parameter and the first translation parameter to obtain the initial transformed feature. In the second affine coupler transformation module, an affine transformation is performed on the initial transformation features to obtain the second scale parameters. Second translation parameter Based on the second scale parameter and the second translation parameter, feature transformation is performed on the first split feature to obtain the intermediate transformed feature. .

[0147] Density estimation networks based on normalized flow are preferred. =12 affine coupling blocks.

[0148] In one embodiment, the affine coupling transformation process is as shown in equation (8):

[0149] (8).

[0150] in, This indicates element-wise product.

[0151] The initial transformed features and intermediate transformed features are concatenated to obtain the second fused features. .

[0152] Anomaly assessment is performed on the second fusion feature to obtain an anomaly score. The higher the anomaly score, the greater the degree of deviation from the normal distribution.

[0153] An anomaly detection threshold is set. If the anomaly score is abnormally greater than the threshold, the first arc fault detection result is 1; otherwise, the first arc fault detection result is 0.

[0154] The first-stage anomaly detection model is obtained by training a normalized flow density estimation network using the features of the first sample. The training objective is to optimize the model parameters to maximize the likelihood probability of the features of the first sample.

[0155] In one embodiment, the negative log-likelihood loss of the normalized flux density estimation model As shown in formula (9):

[0156] (9).

[0157] in, Characterizes the second fusion feature of the sample. The first sample feature characterizing the sample electrical signal, This represents the partial derivative function.

[0158] By minimizing the negative log-likelihood loss, the first sample feature of the sample electrical signal under normal conditions is improved. After normalized flow mapping, it should be as close as possible to the second fusion feature of the sample. The origin region is used to achieve high-precision modeling of normal sample distribution.

[0159] A normalized current density estimation network trained only on normal samples is used to obtain a first-stage anomaly detection model, which is then used to perform series arc anomaly screening on the samples to be detected.

[0160] In the first stage, an anomaly detection model trained solely on sample electrical signals under normal operating conditions is used to detect arc faults in the electrical signals under test. This avoids dependence on real arc fault samples and enables effective screening of series arc anomaly samples under zero fault sample conditions. This overcomes the application bottleneck of related supervised learning methods in real deployment scenarios where it is difficult to obtain fault samples.

[0161] Optionally, the first feature extraction model is obtained by training a first deep learning model using training samples. The training samples include 12 types of residential electrical loads to construct multi-load parallel operating conditions, specifically: kettle (1650W), electric cooker (370W), washing machine (250W), vacuum cleaner (640W), electric drill (230W), hair dryer (1360W), induction cooker (1800W), microwave oven (380W), fluorescent lamp (120W), computer (180W), LED light (80W), and monitor (100W). The parallel sample load branches are connected in series with the arc generator, and the remaining loads are connected in parallel to the main sample circuit as masking loads. The experiment covers both single-load and multi-load operating conditions, with the number of parallel masking loads ranging from 1 to 8 in the multi-load condition. A complete cycle of the fundamental voltage or current is used as one sample. The arc generator is used to simulate arc faults in a real environment.

[0162] The training set consists only of sample electrical signals under normal operating conditions, containing a total of 12,285 samples; the test set includes 8,190 samples, and 820 arc fault samples for each of the 12 types of electrical equipment loads. All arc fault labels and load fine-grained labels are only used for model performance evaluation after training and do not participate in model training supervision.

[0163] Sample multi-view features are extracted from the sample differential current and sample voltage of the sample electrical signal. These features are used to jointly characterize the electrical feature deviation of the sample electrical signal caused by the series arc fault from at least two dimensions, namely, time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution offset.

[0164] The sample multi-view feature includes at least two sample multi-view sub-features. For example, sample trajectory features, sample optical flow features, and sample expert features. Sample trajectory features characterize the temporal topological distortion of the sample electrical signal in the voltage-current plane. Sample optical flow features characterize the dynamic differences between adjacent sample cycles. Sample expert features characterize the statistical distribution shift of the sample electrical signal.

[0165] Figure 3B A schematic diagram of a first deep learning model according to an embodiment of this application is shown.

[0166] like Figure 3BAs shown, the first deep learning model includes a convolutional encoder, a convolutional decoder, a fully connected encoder, a fully connected decoder, a cross-view feature fusion module (GCFAgg), and a multilayer perceptron.

[0167] Using a convolutional encoder to analyze sample trajectory features Feature extraction is performed to obtain the first sample view encoding features corresponding to the sample trajectory features. Using a convolutional encoder to analyze the optical flow features of the samples Feature extraction is performed to obtain the first sample view encoding features corresponding to the sample optical flow features. Using a fully connected encoder to analyze sample expert features Feature extraction is performed to obtain the first sample view encoded features corresponding to the sample expert features. .

[0168] The first sample view encoded features corresponding to the sample trajectory features and sample optical flow features are decoded using a convolutional decoder to obtain the first sample decoded features of the sample trajectory features. First sample decoding features of optical flow characteristics of samples The first sample view encoded features of the sample expert features are decoded using a fully connected decoder to obtain the first sample decoded features of the sample expert features. .

[0169] The cross-view feature fusion module processes the encoded features of multiple first sample views to obtain the global structural relationship parameters and features of the first sample. The generation process of the first-sample global structural relationship parameters is the same as that during model inference, and will not be elaborated here. The first-sample global structural relationship parameters are used to characterize the structural relationship strength between the stitched features determined based on the encoded features of multiple first-sample views in different projection spaces.

[0170] The first sample view is encoded using a multilayer perceptron. ) Perform feature mapping to obtain multiple sample mapping features ( ).

[0171] In one embodiment, a structure-guided multi-view contrastive learning (SgCL) network is used to calculate the contrastive loss function value. The first objective loss function is shown in formula (10):

[0172] (10).

[0173] in, Indicates the number of samples. Parameters characterizing the global structural relationships of the first sample. The first sample feature characterizing the b-th sample, The sample mapping feature that represents the first sample view encoding feature of the b-th sample. The sample mapping feature represents the view encoding feature corresponding to the p-th first sample of the q-th sample, where C is the cosine similarity. This represents the temperature parameter.

[0174] The training objective of the first objective loss function is to ensure that samples with high structural relationships maintain consistent representations, while samples with low structural relationships play a major separating role during training. By minimizing the contrastive loss function value, the consistency of multiple views in representing the fault features of the same sample can be enhanced, and the interference of view-specific redundant information on the detection of series arc faults can be suppressed.

[0175] In one embodiment, the loss function value is reconstructed. As shown in formula (11):

[0176] (11).

[0177] in, Represents the square of the L2 norm. The first sample decoding feature that represents the p-th first sample view encoding feature of the q-th sample. The multi-view sub-features characterizing the p-th sample of the q-th sample.

[0178] The reconstruction loss is used to constrain the ability of the low-dimensional latent representation to learn from the multi-view features of the original input samples.

[0179] The total loss is determined by comparing the loss function value and reconstructing the loss function value. The first deep learning model is then trained based on the total loss to obtain the first feature extraction model.

[0180] The training of the first feature extraction model consists of two steps: First, the encoder and decoder are pre-trained separately for 600 epochs using the reconstruction loss function value; then, the entire first deep learning model is trained for 80 epochs under the total loss constraint; finally, the first deep learning model is frozen, and a density estimation network based on normalized flow is trained using the first sample feature representation to obtain the first-stage anomaly detection model, which is trained for 200 epochs. The optimizer used is Adam, with a learning rate of 0.0005 and a batch size of 256. The preferred sampling rate is 30kHz, the preferred image resolution parameter is 32, and the phase space reconstruction delay time is [not specified]. The preferred setting is 1, temperature parameter The preferred value is 0.4, and the preferred value for anomaly detection threshold is 0.25.

[0181] The first-stage anomaly detection model is used to detect faults in the first sample features, and the sample arc fault detection results of the first sample features are obtained.

[0182] Samples with an arc fault detection result of 1 in the first sample feature are identified as abnormal samples and fed into the second stage as training samples for the second feature extraction model, using them as unlabeled arc samples. Samples with an arc fault detection result of 0 in the first sample feature are identified as normal samples. Since the first feature extraction model in the first stage is trained only on normal samples, its output of abnormal samples may contain false positives. These false positives will be further corrected in the second stage by combining the final clustering results.

[0183] A second deep learning model is used to extract features from the multi-view features of the abnormal sample electrical signals, resulting in second sample features. The second deep learning model has the same model structure and parameters as the first deep learning model.

[0184] Cluster analysis is performed on the second set of sample features to group samples with similar feature distributions into the same category. Based on the cluster category to which each sample belongs, a corresponding arc fault detection result is generated. The arc fault detection result is used to characterize the initial category affiliation of unlabeled arc samples in the feature space; the arc fault detection result is a pseudo-label. For example, the arc fault detection result might be 1.

[0185] The quality assessment results of the sample arc fault detection results are calculated based on the sample clustering quality evaluation index, and the confidence level of each sample is determined based on the distance relationship between the cluster center and the next nearest cluster center to which each sample belongs.

[0186] The sample quality assessment results are used to characterize the intra-class compactness and inter-class separation of the current clustering results, and the confidence score is used to characterize the reliability of the current pseudo-labels of each sample and serves as the basis for self-supervised iterative optimization.

[0187] The quality evaluation metrics for sample clustering include the silhouette coefficient (SC) and the Davies-Bouldin index. The silhouette coefficient is used to characterize the intra-cluster cohesion and inter-cluster segregation of the samples, while the Davies-Bouldin index is used to characterize the ratio between intra-cluster dispersion and inter-cluster segregation in the clustering results.

[0188] The sample arc fault detection results are dynamically verified based on the sample quality assessment results. When the sample quality assessment results indicate that the current clustering structure is better than the previous iteration result, the sample arc fault detection results are re-clustered and updated, and the next round of self-supervised optimization is continued. Otherwise, the update is stopped.

[0189] Based on the sample arc fault detection results, the cluster center and the next nearest cluster center of the sample are determined. The confidence level is then calculated based on the distance from the sample to its respective cluster center and the distance to the next nearest cluster center. In one embodiment, the confidence level... As shown in formula (12):

[0190] (12).

[0191] in, Characterizes the distance from a sample to its second nearest cluster center. The confidence level represents the distance from a sample to its cluster center. The higher the confidence level, the stronger the correlation between the sample and its cluster center, and the more reliable its pseudo-label.

[0192] The triplet loss function value is determined based on the confidence level of the sample arc fault detection results, the distance between the second sample feature and the positive sample feature, and the distance between the second sample feature and the negative sample feature.

[0193] Positive sample features can be selected from multiple other second sample features that are identical to the sample arc fault detection result of the second sample feature, and the distance between the second sample feature and the second sample feature is greater than or equal to a first predetermined distance.

[0194] Negative sample features can be selected from multiple other second sample features that are different from the sample arc fault detection result of the second sample feature. The distance between the negative sample feature and the second sample feature is less than or equal to the second predetermined distance.

[0195] The first predetermined distance is greater than the second predetermined distance.

[0196] In one embodiment, multiple second sample features A, B, C, D, and E are clustered. Second sample features A, B, and C form one cluster, with a sample arc fault detection result of 1. D and E form another cluster, with a sample arc fault detection result of 0. For second sample feature A, the second sample feature B that is farthest from second sample feature A is selected from other second sample features B and C belonging to the same cluster as a positive sample feature. The second sample feature D that is closest to second sample feature A is selected from other second sample features D and E belonging to different clusters as a negative sample feature.

[0197] In one embodiment, the second objective loss function is shown in equation (13):

[0198] (13).

[0199] in, express , This represents the interval hyperparameter. The positive sample feature corresponding to the second sample feature of the q-th sample. The negative sample feature corresponding to the second sample feature of the q-th sample. The second sample feature of the q-th sample has a triplet loss function value of . .

[0200] The interval hyperparameter in the second objective loss function is preferably set to 1.

[0201] By introducing confidence levels into the second objective loss function, the weight of high-confidence samples in the optimization process can be increased, while the interference of pseudo-label noise that may be introduced by low-confidence samples can be reduced. The triplet loss function value can enhance the separability between different cluster categories and improve the compactness within the same cluster category.

[0202] Calculate the contrastive loss function value and reconstruction loss function value of the multi-view features of the abnormal samples. Based on the total loss determined by the triplet loss function value, contrastive loss function value and reconstruction loss function value, train the second deep learning model to obtain the second feature extraction model.

[0203] By iteratively optimizing the total loss using a preset loss value, the second sample features output from the second-stage training gradually acquire stronger intra-class compactness and inter-class separability. Based on these second sample features, final clustering and discrimination are performed to obtain the arc fault detection results corresponding to each abnormal sample. Furthermore, the identification results of the parallel branch where the series arc fault occurs are determined based on these arc fault detection results.

[0204] The arc fault detection results from the final clustering in the second stage are jointly analyzed with the first arc fault detection results from the first stage. For samples whose first-sample-feature arc fault detection result indicates an arc fault in the circuit under test, but whose actual arc fault detection result indicates no arc fault in the circuit under test, these are corrected to normal samples. For samples whose first-sample-feature arc fault detection result indicates an arc fault in the circuit under test, and whose actual arc fault detection result indicates an arc fault in a parallel load branch, their abnormal attributes are retained, and their corresponding final category is output as the series arc fault branch identification result. Through this method, the correction of false detection samples from the first stage and the identification of fault branches for true arc fault samples are achieved.

[0205] In the second stage, the second feature extraction model is trained and iteratively optimized for 100 epochs, with the K-Means clustering algorithm being the preferred choice for cluster analysis. In this second stage, for the unlabeled anomalous samples screened out in the first stage, clustering, clustering quality evaluation, confidence estimation, and a self-supervised iterative optimization mechanism are introduced. This enables the identification of fault branches within anomalous samples without the need for manual labeling of fine-grained faults. Furthermore, the false detection samples from the first stage are corrected using the final clustered sample arc fault detection results, thereby simultaneously improving the accuracy of fault branch identification and the overall detection reliability.

[0206] The laboratory offline training server was configured with a 12th-generation Intel Core i7-12700 processor, 32GB of RAM, and a single GeForce RTX 4060 graphics card; the online deployment platform was a Raspberry Pi 4B, configured with a 1.5GHz 64-bit quad-core Arm Cortex A72 processor and 8GB of RAM. The final output of the second sample feature's arc fault detection results was evaluated using a 13-class confusion matrix, where 12 classes were series arc fault load categories and 1 was the normal category. For the fault detection task, the 12 fault classes were merged into a single anomalous class, with accuracy, false negative rate, and false positive rate used as metrics. For the fault branch identification task, the Macro-F1 score was used as the metric.

[0207] Figure 4A A schematic diagram illustrating the training effect of the first-stage anomaly detection model according to an embodiment of this application is shown.

[0208] like Figure 4A As shown, the anomaly score is a quantitative indicator output by the first-stage anomaly detection model, used to measure the degree to which a sample deviates from the normal pattern, typically ranging from 0 to 1. The first-stage anomaly detection model outputs the following results: for normal samples without arc faults, the anomaly scores are mainly concentrated in the low-score region and mostly below the judgment threshold. However, for arc fault samples with arc faults, the anomaly scores are generally distributed in the higher-score region. Although there is a slight overlap near the threshold, the overall separation is significant. This indicates that the first-stage anomaly detection model can effectively form a screening boundary for arc fault samples under training conditions using only normal samples, and provides a relatively reliable set of candidate anomaly samples for the second stage.

[0209] Figure 4B A schematic diagram of clustering second sample features according to an embodiment of this application is shown.

[0210] like Figure 4BAs shown, the horizontal and vertical axes represent the two principal component dimensions of the second feature after dimensionality reduction, used to visualize the high-dimensional feature space. After clustering the second sample features, clusters represented by different shaped labels are output. The category of the cluster represents the sample arc fault detection result, including thirteen categories: normal, kettle, electric cooker, washing machine, vacuum cleaner, electric drill, microwave oven, induction cooker, hair dryer, fluorescent lamp, computer, LED light, and monitor. After the second stage of cluster optimization and self-supervised iterative optimization, the second sample features finally exhibit a better clustering structure in the low-dimensional space, with most different load categories forming cluster structures with clear boundaries and compact intra-class structures. This indicates that the constructed two-stage self-supervised training method can not only achieve anomaly detection of series arc faults, but also further achieve fine-grained identification of fault branches and correct false detection samples introduced in the first stage.

[0211] In this embodiment, a fault detection accuracy of 99.0%, a false negative rate of 1.5%, a false positive rate of 0.4%, and a Macro-F1 score of 98.0% were achieved. These results were obtained during the training phase using only normal samples, without fault samples or fine-grained labels. Compared to related supervised methods, this application still demonstrates competitive detection and identification performance under zero-fault sample conditions.

[0212] Furthermore, this application also verifies the feasibility of engineering deployment of the model. Experimental results show that the model trained in this application can be deployed online on the Raspberry Pi 4B platform, with a single sample processing time of 84.71 ms.

[0213] By employing dynamic baseline subtraction differential preprocessing, adaptive spectrum separation, training a first deep learning model based on normal samples in the first stage to obtain a first feature extraction model and a first-stage anomaly detection model, and training a second deep learning model based on unlabeled abnormal arc samples in the second stage to obtain a second feature extraction model, and clustering self-supervised optimization, this system can effectively screen series arc faults, identify fault branches, and correct false detection samples in multi-load parallel scenarios. It has the advantages of requiring no fault samples, requiring no fine-grained labels, being highly adaptable to complex background currents, having high detection accuracy, and being easy to deploy at the edge.

[0214] Figure 5 A block diagram of an electronic device suitable for implementing an arc fault detection method according to an embodiment of this application is shown.

[0215] Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0216] like Figure 5As shown, an electronic device 500 according to an embodiment of this application includes a processor 501, which can perform various appropriate actions and processes according to a program stored in ROM 502 or a program loaded from storage portion 508 into RAM 503, wherein ROM 502 is a read-only memory and RAM 503 is a random access memory. The processor 501 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 501 may also include onboard memory for caching purposes. The processor 501 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0217] RAM 503 stores various programs and data required for the operation of electronic device 500. Processor 501, ROM 502, and RAM 503 are interconnected via bus 504. Processor 501 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 502 and / or RAM 503. It should be noted that programs may also be stored in one or more memories other than ROM 502 and RAM 503. Processor 501 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0218] Optionally, the electronic device 500 may also include an input / output (I / O) interface 505, which is also connected to the bus 504. The electronic device 500 may also include one or more of the following components connected to the input / output (I / O) interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the input / output (I / O) interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 510 as needed so that computer programs read from it can be installed into the storage section 508 as needed.

[0219] Optionally, the method flow according to the embodiments of this application can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by processor 501, it performs the functions defined in the system of the embodiments of this application. Optionally, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0220] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the arc fault detection method according to the embodiments of this application.

[0221] Optionally, the computer-readable storage medium may be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0222] For example, optionally, the computer-readable storage medium may include the ROM 502 and / or RAM 503 described above and / or one or more memories other than ROM 502 and RAM 503.

[0223] Embodiments of this application also include a computer program product, which includes a computer program containing program code for performing the methods provided in the embodiments of this application. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the arc fault detection method provided in the embodiments of this application.

[0224] When the computer program is executed by the processor 501, it performs the functions defined in the system / apparatus of this application embodiment. Optionally, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0225] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 509, and / or installed from a removable medium 511. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0226] Optionally, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0227] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments of this application can be combined and / or combined in various ways. All such combinations and / or combinations fall within the scope of this application.

[0228] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A method for detecting electric arc faults, characterized in that, The method for detecting series electric arcs includes: Based on the electrical signal to be detected from the circuit under test, differential current and target voltage are obtained. The circuit under test includes a main circuit and multiple parallel load branches. The electrical signal to be detected includes the main current to be detected and the main voltage to be detected. The differential current characterizes the difference between cycles of the main current to be detected after phase alignment, so as to reflect the waveform distortion and non-stationary evolution caused by series arc fault. The target voltage characterizes the steady-state reference of the main voltage to be detected after phase alignment. Based on the differential current and the target voltage, a multi-view feature is obtained. The multi-view feature is used to jointly characterize the electrical characteristic deviation of the series arc fault on the electrical signal to be detected from at least two dimensions, namely, time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution offset. The multi-view features are processed using a first feature extraction model to obtain a first feature. This first feature extraction model is trained using the multi-view features of the sample electrical signal. The sample electrical signal is an electrical signal obtained under normal operating conditions of the sample circuit. The sample circuit includes a main sample circuit and multiple parallel sample load branches. If the first arc fault detection result obtained by performing arc fault detection on the first feature indicates that the circuit under test has a series arc fault, then the second feature obtained by processing the multi-view feature using the second feature extraction model is used to perform arc fault detection, resulting in a second arc fault detection result indicating that the circuit under test does not have a series arc fault or has a series arc fault in the parallel load branch. The second feature extraction model is trained using the abnormal sample multi-view feature of the abnormal sample electrical signal, and the abnormal sample electrical signal is the sample electrical signal corresponding to the sample arc fault detection result indicating that the sample circuit has a series arc fault.

2. The method according to claim 1, characterized in that, The multi-view feature includes at least two of the following multi-view sub-features: trajectory feature, optical flow feature, or expert feature. The trajectory feature characterizes the temporal topological distortion of the electrical signal to be detected in the voltage-current plane. The optical flow feature characterizes the dynamic difference between adjacent cycles. The expert feature characterizes the statistical distribution shift of the electrical signal to be detected. The step of processing the multi-view features using a first feature extraction model to obtain the first feature includes: The encoder in the first feature extraction model processes the multi-view features to obtain view encoding features corresponding to each of the multi-view sub-features; and The first feature is obtained by using the cross-view feature fusion module in the first feature extraction model to process the view encoding features corresponding to each of the multiple multi-view sub-features. The first feature is represented as an embedding representation of the complementary information between different view sub-features mapped to a low-dimensional discriminative representation space.

3. The method according to claim 2, characterized in that, The cross-view feature fusion module includes a feature fusion unit and a feature enhancement unit; The step of using the cross-view feature fusion module in the first feature extraction model to process the view encoding features corresponding to each of the multiple multi-view sub-features to obtain the first feature includes: The feature fusion unit processes the view encoding features corresponding to each of the multiple multi-view sub-features to obtain a first fusion feature. The first fusion feature represents the common patterns and complementary fault features of the circuit operation state under different views. Based on the first fusion feature, the feature enhancement unit processes the view encoding features corresponding to each of the multiple multi-view sub-features to obtain enhanced features, which represent a context-enhanced representation under the constraints of global circuit feature structure relationships; and The first feature is obtained based on the enhanced features.

4. The method according to claim 3, characterized in that, The process of using the feature fusion unit to process the view encoding features corresponding to each of the multiple multi-view sub-features to obtain the first fused feature includes: The first fused feature is obtained by fusing the view coding features corresponding to each of the multiple multi-view sub-features using multiple projection parameters in the feature fusion unit; The step of processing the view encoding features corresponding to each of the multiple multi-view sub-features based on the first fusion feature to obtain the enhanced features includes: The feature enhancement unit utilizes multiple enhancement parameters to perform feature fusion on multiple view-encoded features, obtaining global structural relationship parameters. These global structural relationship parameters characterize the strength of the structural relationship between view-encoded features in different projection spaces. The enhanced features are obtained by using the global structural relationship parameters to enhance the first fused features.

5. The method according to any one of claims 1 to 4, characterized in that, The step of obtaining the differential current and target voltage based on the electrical signal to be detected from the circuit under test includes: Using the phase reference point obtained by detecting the phase reference point of the main voltage to be detected as the starting point of the cycle, the main current and the main voltage to be detected are respectively subjected to cycle interception to obtain phase-aligned current sequence and voltage sequence. The differential current is obtained based on the difference between the reference current sequence and the subsequent current sequence. The reference current sequence is obtained by phase-aligning the currents obtained under stable load conditions, and the subsequent current sequence is obtained by phase-aligning the currents obtained after the reference current sequence. The target voltage is obtained based on the phase-aligned voltage sequence.

6. The method according to any one of claims 2 to 4, characterized in that, The process of obtaining multi-view features based on the differential current and the target voltage includes: The differential current is subjected to spectral separation to obtain a first differential component and a second differential component, wherein the frequency of the second differential component is greater than the frequency of the first differential component. The trajectory features are obtained based on the first differential component and the target voltage; Optical flow estimation is performed on the trajectory features to obtain the optical flow features; The expert features are obtained based on the first difference component and the second difference component; and The multi-view features are obtained based on at least two of the trajectory features, the optical flow features, or the expert features.

7. The method according to claim 6, characterized in that, The step of performing spectral separation on the differential current to obtain a first differential component and a second differential component includes: The differential current is reconstructed in phase space to obtain reconstructed trajectory parameters, which characterize the change in fault dynamic complexity of the circuit under the current operating state. The reconstructed trajectory parameters are subjected to temporal component extraction to obtain multiple temporal components; Based on the spectral parameters of each of the multiple time-domain components, the multiple time-domain components are clustered to obtain a first cluster and a second cluster; and The time-domain components in the first cluster are used as the first difference components, and the time-domain components in the second cluster are used as the second difference components.

8. The method according to any one of claims 1 to 4, characterized in that, The method further includes: The first feature is split to obtain a first split feature and a second split feature; An affine transformation is performed on the first split feature to obtain a first scale parameter and a first translation parameter. The first scale parameter is used to control the scaling degree of the feature, and the first translation parameter is used to control the offset degree of the feature. Based on the first scale parameter and the first translation parameter, the second split feature is transformed to obtain the initial transformed feature; An affine transformation is performed on the initial transformation features to obtain the second scale parameter and the second translation parameter; Based on the second scale parameter and the second translation parameter, the first split feature is transformed to obtain the intermediate transformed feature; The initial transformation features and the intermediate transformation features are fused to obtain a second fused feature; and The first arc fault detection result is obtained based on the second fusion feature.

9. The method according to any one of claims 1 to 4, characterized in that, The sample multi-view features are obtained based on the sample differential current and sample voltage, which are obtained based on the sample electrical signal of the sample circuit. The sample electrical signal includes the sample main current and sample main voltage. The sample differential current characterizes the difference between cycles of the phase-aligned sample main current to reflect the waveform distortion and non-stationary evolution caused by the series arc fault. The sample voltage characterizes the steady-state reference of the phase-aligned sample main voltage. The sample multi-view features are used to jointly characterize the electrical characteristic deviation of the sample electrical signal caused by the series arc fault from at least two dimensions: time-domain topological distortion, inter-cycle dynamic difference, or statistical distribution offset. The sample multi-view features include at least two sample multi-view sub-features. The first feature extraction model is trained in the following manner: The sample electrical signal is processed using a first deep learning model to obtain first sample decoding features, first sample global structural relationship parameters, first sample features, and first sample view encoding features corresponding to each of the multiple sample multi-view sub-features; and Based on the first target loss function, the first deep learning model is trained according to the contrast loss function value determined by the first sample global structure relationship parameter, the first sample feature and the first sample view encoding feature corresponding to each of the multiple sample multi-view sub-features, and the reconstruction loss function value determined by the sample multi-view feature and the first sample decoding feature, to obtain the first feature extraction model. The abnormal sample multi-view features include at least two abnormal sample multi-view sub-features, and the second feature extraction model is trained in the following manner: The abnormal sample electrical signal is subjected to feature extraction using a second deep learning model to obtain the abnormal sample multi-view features, the second sample decoding features, the second sample global structural relationship parameters, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features; and Based on the second objective loss function, the second deep learning model is trained according to the contrast loss function value determined by the second sample global structure relationship parameters, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features, as well as the reconstruction loss function value determined by the abnormal sample multi-view features and the second sample decoding features, to obtain the second feature extraction model.

10. The method according to claim 9, characterized in that, The first feature extraction model is obtained by training the first deep learning model based on the first target loss function, the contrast loss function value determined according to the first sample global structural relationship parameters, the first sample features, and the first sample view encoding features corresponding to each of the multiple sample multi-view sub-features, and the reconstruction loss function value determined according to the sample multi-view features and the first sample decoding features, including: Based on the first target loss function, the first deep learning model is trained according to the similarity between the first sample features and multiple first sample view encoding features, the contrast loss function value determined by the first sample global structure relationship parameters, and the reconstruction loss function value determined by the sample multi-view features and the first sample decoding features, to obtain the first feature extraction model. The second feature extraction model is trained based on the second objective loss function, using the contrast loss function value determined by the second sample global structural relationship parameters, the second sample features, and the second sample view encoding features corresponding to each of the multiple abnormal sample multi-view sub-features, and the reconstruction loss function value determined by the abnormal sample multi-view features and the second sample decoding features, to obtain the second feature extraction model. This includes: Based on the sample arc fault detection results obtained from cluster analysis of the second sample features, a quality assessment is performed to obtain the sample quality assessment result; and If the sample quality assessment result is satisfactory, the second deep learning model is trained based on the second objective loss function, the confidence level of the sample arc fault detection result, the triplet loss function value determined by the distance between the second sample feature and the positive sample feature, and the distance between the second sample feature and the negative sample feature, the contrast loss function value determined by the second sample global structure relationship parameter, the second sample feature, and the second sample view encoding feature corresponding to each of the multiple abnormal sample multi-view sub-features, and the reconstruction loss function value determined by the abnormal sample multi-view feature and the second sample decoding feature, to obtain the second feature extraction model. The positive sample feature is one of the other second sample features with the same sample arc fault detection result as the second sample feature, whose distance from the second sample feature is greater than or equal to a first predetermined distance. The negative sample feature is one of the other second sample features with a different sample arc fault detection result than the second sample feature, whose distance from the second sample feature is less than or equal to a second predetermined distance, where the first predetermined distance is greater than the second predetermined distance.

Images