An unsupervised arc fault detection method and system based on a physically guided deep autoencoder and multi-source feature fusion

Abstract

This invention discloses an unsupervised fault arc detection method and system based on a physically guided deep autoencoder and multi-source feature fusion. The method includes: acquiring the real-time current signal of the protected line; preprocessing it by framing and normalizing using a sliding window, and calculating the frequency domain feature vector and time domain statistical feature vector of the current analysis window; inputting the preprocessed real-time current signal into the encoder of a one-dimensional convolutional autoencoder with frozen parameters to extract hidden layer feature vectors, simultaneously obtaining the reconstruction output and calculating the reconstruction error scalar; concatenating the hidden layer feature vector, reconstruction error scalar, frequency domain feature vector, and time domain statistical feature vector into a feature layer to form a multi-source fused feature vector; and inputting the multi-source fused feature vector into a pre-trained single-class support vector machine to obtain the output value of the decision function to determine whether a fault arc exists.

Description

Technical Field

[0001] This invention belongs to the field of power system fault diagnosis and artificial intelligence technology. Specifically, it relates to an unsupervised detection method and system for low-voltage AC series fault arcs that integrates a physical information-guided autoencoder, multi-source electrical feature extraction and single-class support vector machine. It is applicable to real-time detection and early warning of fault arcs in various scenarios such as household electricity and industrial power distribution. Background Technology

[0002] Electric arc faults are one of the main causes of electrical fires. Among them, series arc faults are particularly dangerous because their fault current is small and they are greatly affected by the type of load, making them difficult to detect by traditional overcurrent protection devices (such as circuit breakers and fuses).

[0003] Traditional arc fault detection methods are mainly divided into three categories: those based on physical quantity detection (arc light, arc sound, electromagnetic radiation), those based on threshold discrimination of current and voltage waveform characteristics, and those based on machine learning classification methods. The first type of method requires the sensor to be placed near the fault point, which limits its practical application and makes it unsuitable for environments requiring non-invasive detection. The second type of method relies on manually set features and thresholds, which have insufficient generalization ability when the load types are complex and diverse. The characteristics of normal and fault states of different loads overlap, resulting in high false alarm and false negative rates.

[0004] In recent years, deep learning techniques have been introduced into the field of arc fault detection, with convolutional neural networks, long short-term memory networks, and their variants being extensively studied. However, most existing deep learning methods employ supervised classification, requiring a large number of labeled arc fault samples for training. In real-world scenarios, acquiring arc fault samples is costly, scarce, and their diverse fault modes make labeling difficult, limiting the engineering application of supervised methods. Meanwhile, some studies have used autoencoders for unsupervised anomaly detection, but these typically rely solely on reconstruction errors and fixed thresholds for judgment, easily leading to misjudgments under complex conditions such as nonlinear loads. Furthermore, existing technologies include multi-source feature fusion schemes for arc fault detection, but these schemes mostly employ supervised classifiers, still relying on a large number of labeled fault samples. Additionally, research on using physical information neural networks for arc detection mainly involves modifying the network structure to embed physical knowledge, resulting in high complexity and difficulty in compatibility with unsupervised frameworks.

[0005] In summary, existing technologies cannot simultaneously solve the following problems: no need for fault sample labeling, automatic extraction of effective features, full utilization of complementary information between deep features and reconstruction errors, fusion of multi-source electrical features, and introduction of physical priors to improve generalization ability. Summary of the Invention

[0006] This invention addresses the challenges of fault sample acquisition, poor generalization of artificial features, limited feature sources, and lack of physical interpretability in existing fault arc detection technologies. It provides an unsupervised fault arc detection method and system based on a physically guided deep autoencoder and multi-source feature fusion. The method constructs a one-dimensional convolutional autoencoder and trains it unsupervised using only normal operating condition samples. Physical constraints such as fundamental amplitude phase consistency, total harmonic distortion rate consistency, and power frequency periodicity are introduced into the loss function to guide the model in learning a normal pattern manifold that conforms to the physical laws of circuits. After freezing the trained autoencoder parameters, hidden layer feature vectors and reconstruction error scalars are extracted and concatenated with frequency domain feature vectors and time domain statistical feature vectors to form a multi-source fused feature vector. This fused feature vector is then used to train a single-class support vector machine to construct a compact decision boundary for normal samples. This achieves the goals of fault arc detection without fault sample labeling, improved model generalization ability and physical interpretability, and enhanced detection accuracy under complex nonlinear load conditions.

[0007] According to one aspect of the present invention, an unsupervised fault arc detection method based on a physically guided deep autoencoder and multi-source feature fusion is provided, comprising: acquiring a real-time current signal of a protected line; preprocessing it by framing and normalizing it through a sliding window, and calculating the frequency domain feature vector and time domain statistical feature vector of the current analysis window; inputting the preprocessed real-time current signal into the encoder of a one-dimensional convolutional autoencoder with frozen parameters to extract hidden layer feature vectors; and inputting the preprocessed real-time current signal into the one-dimensional convolutional autoencoder with frozen parameters to obtain a reconstruction output and calculate a reconstruction error scalar; concatenating the hidden layer feature vector, the reconstruction error scalar, the frequency domain feature vector, and the time domain statistical feature vector into a feature layer to form a multi-source fusion feature vector; inputting the multi-source fusion feature vector into a pre-trained single-class support vector machine to obtain the output value of a decision function, and determining whether a fault arc exists based on the output value of the decision function.

[0008] As a further technical solution, the training process of the one-dimensional convolutional autoencoder uses only normal working condition samples, and its training loss function includes a physical information guiding term; the physical information guiding term includes at least one of the following: fundamental amplitude and phase consistency constraint, total harmonic distortion rate consistency constraint, and power frequency periodicity constraint.

[0009] As a further technical solution, the training loss function is expressed as: ,in For reconstruction error, It is the sum of the squared difference between the fundamental amplitude and the squared difference between the input signal and the reconstructed signal. It is the square of the difference between the total harmonic distortion rate of the input signal and the reconstructed signal. To reconstruct the sum of squares of the waveform differences between adjacent periods of the signal, , , This is a hyperparameter.

[0010] As a further technical solution, the encoder of the one-dimensional convolutional autoencoder embeds an efficient channel attention mechanism after each convolutional layer. The efficient channel attention mechanism captures cross-channel interaction information between channels through a one-dimensional adaptive convolutional kernel.

[0011] As a further technical solution, the frequency domain feature vector is the amplitude sequence of the first N harmonic components after performing a fast Fourier transform on the real-time current signal, where N is 10 to 30.

[0012] As a further technical solution, the time-domain statistical feature vector includes at least one or more of the following: peak-to-peak value, variance, skewness, and kurtosis.

[0013] As a further technical solution, the single-class support vector machine adopts a radial basis kernel function.

[0014] As a further technical solution, a delayed determination step is also included: when the output values ​​of the decision function of M consecutive analysis windows all indicate the presence of a fault arc, a fault alarm signal is output; otherwise, the previous stable state is maintained; the value of M ranges from 3 to 5.

[0015] According to one aspect of this invention, an unsupervised fault arc detection system based on a physically guided deep autoencoder and multi-source feature fusion is provided, comprising: a signal acquisition and preprocessing module for acquiring real-time current signals of the protected line, performing preprocessing through sliding window framing and normalization, and calculating the frequency domain feature vector and time domain statistical feature vector of the current analysis window; a frozen encoder module for inputting the preprocessed real-time current signal into the encoder of a one-dimensional convolutional autoencoder with frozen parameters to extract hidden layer feature vectors; a reconstruction error calculation module for inputting the preprocessed real-time current signal into the one-dimensional convolutional autoencoder with frozen parameters to obtain reconstruction output and calculate reconstruction error scalar; a multi-source feature fusion module for concatenating the hidden layer feature vectors, reconstruction error scalar, frequency domain feature vectors, and time domain statistical feature vectors into a feature layer to form a multi-source fused feature vector; and a single-class support vector machine discrimination module for inputting the multi-source fused feature vector into a pre-trained single-class support vector machine to obtain the output value of the decision function and determine whether a fault arc exists.

[0016] According to one aspect of the present invention, a fault arc detection device is provided, comprising a current sensor, a signal conditioning circuit, an analog-to-digital converter, a memory, and a processor. The current sensor is connected to the analog-to-digital converter via the signal conditioning circuit. The analog-to-digital converter and the memory are respectively connected to the processor. When the processor executes a program stored in the memory, it implements the steps of the method.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. No need for fault sample annotation, lowering the threshold for engineering applications: The training phase only requires current signals under normal operating conditions, completely eliminating the dependence on a large number of labeled fault arc samples, significantly reducing data preparation costs and deployment difficulty.

[0019] 2. Enhanced discrimination capability through multi-source feature fusion: The feature layer fuses four distinct sources of information: autoencoder hidden layer features, reconstruction error, frequency domain harmonic features, and time domain statistical features. This enables a single-class SVM to make comprehensive judgments from four dimensions: deep manifold, anomaly degree, frequency domain distribution, and statistical characteristics. The detection accuracy is significantly improved under complex working conditions such as nonlinear loads.

[0020] 3. Enhancing Model Generalization and Interpretability through Physical Information Guidance: By introducing fundamental amplitude phase consistency constraints, total harmonic distortion rate consistency constraints, and period consistency constraints into the autoencoder loss function, the model is guided to learn normal pattern manifolds that conform to the physical laws of circuits. This method is simple to implement, has low computational cost, and can effectively avoid overfitting the model to non-essential noise in the training data. It also exhibits stronger robustness when load type changes or unknown loads occur.

[0021] 4. Deep fusion of features and errors: The hidden layer features and reconstruction errors are concatenated and then input into a single-class SVM, enabling the classifier to simultaneously learn the intrinsic coupling relationship between the position of the signal in the manifold space and the degree to which the signal deviates from the manifold. This results in higher accuracy under nonlinear load conditions compared to using only a single feature.

[0022] In summary, this invention organically combines four core modules—physically guided deep autoencoder, multi-source feature fusion, concatenation of hidden layer features and reconstruction errors, and single-class SVM decision-making—to form an unsupervised fault arc detection scheme that requires no fault sample labeling, has high detection accuracy, strong generalization ability, and physical interpretability, achieving synergistic technical effects that cannot be achieved by a single technical feature. Attached Figure Description

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

[0024] Figure 1 A flowchart illustrating an unsupervised fault arc detection method based on physical-guided deep autoencoder and multi-source feature fusion, provided for an embodiment of the present invention; Figure 2 This is a schematic diagram of a one-dimensional convolutional autoencoder structure with attention mechanism provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of multi-source feature fusion provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the physical guidance loss function provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0026] This invention aims to solve the following problems existing in current fault arc detection technology: The problem of scarce fault samples: Existing supervised methods rely on a large number of labeled fault arc samples, which are difficult to obtain in real-world scenarios and have high labeling costs; Dependence on manual features and insufficient generalization ability: Traditional methods rely on manually designed time-frequency domain features, and their detection performance drops sharply when the load type is complex or switching occurs; Single feature source: Most existing methods only utilize time-domain waveforms and do not effectively integrate multi-dimensional electrical features such as frequency domain and statistical domain, which limits the richness of discrimination information; The model lacks physical interpretability: Purely data-driven deep models may learn noise patterns that are unrelated to fault arcs, leading to misjudgments and reduced generalization ability.

[0027] Figure 1 This is an overall flowchart of the method of the present invention, which shows the complete process of data acquisition, multi-source preprocessing, physical guided autoencoder training, multi-source feature extraction and fusion, single-class SVM training, online detection and latency determination.

[0028] like Figure 1 As shown, the present invention provides an unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion, comprising the following steps: Step S1: Data acquisition and multi-source signal preprocessing.

[0029] Acquire the raw current signal of the protected line. The sampling frequency is , The value range is 10kHz to 50kHz. The long sequence current signal is divided into multiple analysis units by a sliding window. The length of each analysis unit is L sampling points, where L corresponds to an integer multiple of one power frequency cycle.

[0030] For each analysis unit Perform the following preprocessing:

[0031] The amplitude is normalized to the [0,1] interval;

[0032] Calculate the Fast Fourier Transform of this window to obtain the frequency domain amplitude spectrum. ,in Take the front One harmonic amplitude;

[0033] Calculate the time-domain statistical eigenvector This includes: peak-to-peak value, variance, skewness, and kurtosis.

[0034] During the training phase, only current signal samples under normal operating conditions and their corresponding frequency domain spectra and statistical characteristics are used; fault sample annotation is not required. The training sample set is denoted as... All of them are normal current signals.

[0035] Step S2: Construct a physically guided one-dimensional convolutional autoencoder.

[0036] like Figure 2 As shown, a one-dimensional convolutional autoencoder is constructed, including an encoder module and a decoder module, and an attention mechanism is embedded after each convolutional layer of the encoder.

[0037] The encoder module includes, in sequence:

[0038] The first convolutional layer has a kernel size of 7, a stride of 2, and 32 output channels. It is followed by a batch normalization layer and a ReLU activation function, and then connected to the attention module.

[0039] The second convolutional layer has a kernel size of 5, a stride of 2, and 64 output channels. It is followed by a batch normalization layer and a ReLU activation function, and then connected to the attention module.

[0040] The third convolutional layer has a kernel size of 3, a stride of 2, and 128 output channels. It is followed by a batch normalization layer and a ReLU activation function, and then connected to the attention module.

[0041] Flattening layers and fully connected layers: flatten the convolutional output and map it to the hidden layer feature vector. ,in For hidden layer dimensions.

[0042] The decoder module has a symmetrical structure to the encoder and includes, in sequence:

[0043] Fully connected layers and reshaping layers: Incorporating hidden layer features Map back to the size of the convolutional feature map;

[0044] First transposed convolutional layer: kernel size 3, stride 2, number of output channels 64, followed by batch normalization and ReLU;

[0045] The second transposed convolutional layer has a kernel size of 5, a stride of 2, and 32 output channels, followed by batch normalization and ReLU.

[0046] The third transposed convolutional layer has a kernel size of 7, a stride of 2, and 1 output channel, followed by a Tanh activation function, which outputs the reconstructed signal. .

[0047] Physical guidance mechanisms: such as Figure 4 As shown, in the loss function of the autoencoder, in addition to the conventional reconstruction error, a physical constraint term is introduced to guide the model to learn a reconstruction mapping that conforms to the physical laws of the circuit.

[0048] The total loss function is:

[0049]

[0050] in: , is the mean square error reconstruction loss;

[0051] This represents the loss in the consistency of the fundamental amplitude and phase. and These represent the fundamental amplitude and initial phase of the input signal and the reconstructed signal, respectively.

[0052] This represents the consistency loss of total harmonic distortion. The highest harmonic order;

[0053] This represents the periodic consistency loss. This represents the number of sampling points per power frequency cycle.

[0054] Hyperparameters Normalization settings are configured based on the magnitude of each loss term.

[0055] During training, only normal operating condition samples are used, and the Adam optimizer is employed.

[0056] Step S3: Multi-source feature extraction.

[0057] Freeze the parameters of the trained autoencoder. For any input sample Calculate the following features in sequence:

[0058] Hidden layer feature vectors: ;

[0059] Reconstruction error scalar: ;

[0060] Frequency domain eigenvectors: , is the vector of the amplitude of the first K harmonics after normalization;

[0061] Time-domain statistical eigenvectors: , where is the peak-to-peak value, variance, skewness, and kurtosis.

[0062] Step S4: Multi-source feature fusion.

[0063] like Figure 3 As shown, the four features obtained in step S3 are concatenated at the feature layer to form a multi-source fusion feature vector:

[0064]

[0065] This fusion feature includes: nonlinear manifold information extracted from the time-domain waveform by the deep autoencoder. The degree of abnormality of the signal deviating from the normal manifold Frequency domain harmonic distribution information and time-domain statistical information This enables multi-perspective and multi-level characterization of current signals.

[0066] Step S5: Training a single-class support vector machine.

[0067] The fused feature vector obtained in step S4 As input to a single-class support vector machine. The single-class SVM is trained using only the fused feature vectors corresponding to samples from normal operating conditions.

[0068] The optimization problem for a single-class SVM is:

[0069]

[0070]

[0071] in, Let be the normal vector of the hyperplane. The distance parameter is the distance from the origin. As slack variables, To control the parameters of the upper bound of the training error and the lower bound of the support vectors, This is a kernel function mapping.

[0072] After training, the single-class SVM constructs a compact boundary in the feature space, enclosing the fused features of normal samples inside.

[0073] Step S6: Online detection and delay determination of fault arc.

[0074] In the actual detection stage, the current signal sample to be detected is preprocessed in step S1 to calculate its time-domain waveform, frequency-domain spectrum and statistical characteristics; then step S3 (calculating hidden layer features and reconstruction error through the frozen autoencoder) and step S4 (fusion features) are executed; finally, the fused features are input into the single-class SVM model trained in step S5.

[0075] Single-class SVM output decision function value :

[0076] If the decision value This is considered a normal signal.

[0077] If the decision value This was determined to be a fault arc signal.

[0078] To further reduce false alarms caused by transient noise interference, a delay determination mechanism is introduced: continuous When all analysis units are identified as faulty arcs, a fault alarm signal is finally output; otherwise, the previous stable state is maintained. The value range is 3 to 5.

[0079] As a preferred embodiment, this embodiment describes the arc detection of series faults in a low-voltage AC power distribution system as follows.

[0080] 1. Experimental Platform Setup

[0081] A low-voltage AC arc fault experimental platform was constructed. The experimental circuit was powered by a 220V, 50Hz AC power supply, and the following typical loads were connected to the load terminals:

[0082] Purely resistive loads: Incandescent lamps (200W, 500W), electric furnaces (1000W);

[0083] Non-linear loads: Desktop computer switching power supply (300W), LED driver (50W), inverter air conditioner (1.5 HP), microwave oven (800W);

[0084] Mixed load: A combination of the above loads.

[0085] Bus current signals are acquired using current transformers (CTs, 1000:1 ratio, 20kHz bandwidth). After signal conditioning circuitry (filtering, amplification, biasing), the signals are digitized by a data acquisition card at a 20kHz sampling rate. Arc faults are simulated using a dedicated arc generator, employing both carbonization path and arc gap methods. Each fault arc sample contains at least 10 consecutive arc cycles.

[0086] 2. Dataset Construction

[0087] Collect the following dataset:

[0088] Training set: Various workloads are running under normal operating conditions, collecting 100,000 analysis unit samples (256 points per unit), used only for training autoencoders and single-class SVMs, and does not contain any fault samples.

[0089] Validation set: 20,000 samples under normal operating conditions, used for early shutdown and overparameter adjustment.

[0090] Test set: Contains 20,000 samples of normal operating conditions and 20,000 samples of fault arcs, used for final performance evaluation.

[0091] Sliding window parameters: window length 256 points (one power frequency cycle), step size 128 points.

[0092] 3. Structural details of the self-encoder

[0093] Hidden layer dimensions The attention module employs an efficient channel attention mechanism, and its kernel size is determined by an adaptive function.

[0094] 4. Specific settings for the physical guidance loss function

[0095] Fundamental amplitude and phase For input signals and reconstructed signal Discrete Fourier Transform (DFT) was performed separately, and the precise 50Hz component was extracted using the interpolation FFT method. Fundamental frequency loss weighting was applied. .

[0096] Total Harmonic Distortion (THD): Calculated as the ratio of the sum of squares of the amplitudes of the 2nd to 19th harmonics (100Hz to 950Hz) to the fundamental frequency amplitude. THD loss weight. .

[0097] Periodic consistency loss: (One cycle) .

[0098] Training process: Adam optimizer, initial learning rate 0.001, batch size=128, maximum epochs=200, early stopping patience=20.

[0099] 5. Multi-source feature extraction and fusion

[0100] For each input sample, extract the following features:

[0101] Hidden features 32-dimensional;

[0102] Reconstruction error : 1-dimensional;

[0103] Frequency domain characteristics Take the amplitude of the first 20 harmonics of the FFT (from the fundamental to the 19th harmonic), normalize it to the [0,1] interval, 20 dimensions;

[0104] Time-domain statistical characteristics Peak-to-peak value, variance, skewness, kurtosis, 4 dimensions.

[0105] The concatenated vectors yield a 57-dimensional fused feature vector.

[0106] 6. Single-class SVM training

[0107] Z-score normalization was applied to the 57-dimensional fused features. The single-class SVM used the RBF kernel, and the hyperparameters were... nuclear width Selection was made through cross-validation.

[0108] 7. Test Results

[0109] The method of this invention was comprehensively evaluated on a test set that included purely resistive loads, nonlinear loads (switching power supplies, inverter air conditioners, microwave ovens), and mixed loads. The test set contained a total of 40,000 analysis units, including 20,000 samples of normal operating conditions and 20,000 samples of fault arcing.

[0110] Experimental results show that the method of this invention achieves excellent detection performance under all load types. For purely resistive loads (incandescent lamps, electric furnaces), the detection accuracy reaches 99.6%, recall 99.4%, precision 99.7%, and F1 score 99.5%. Under nonlinear load scenarios, the method maintains high robustness: under switching power supply loads, the accuracy is 99.2%, recall 98.9%, precision 99.3%, and F1 score 99.1%; under inverter air conditioner loads, the accuracy is 98.9%, recall 98.6%, precision 99.0%, and F1 score 98.8%; under microwave oven loads, the accuracy is 98.5%, recall 98.2%, precision 98.7%, and F1 score 98.4%. For mixed load scenarios with multiple load combinations, the detection accuracy is 98.6%, recall 98.3%, precision 98.8%, and F1 score 98.5%.

[0111] The overall weighted average performance is as follows: accuracy 99.0%, recall 98.8%, precision 99.2%, and F1 score 99.0%. These results verify that the method of this invention can achieve high accuracy in fault arc detection under various load conditions, and maintains excellent discrimination ability even under complex conditions such as nonlinear loads and mixed loads.

[0112] 8. Real-time testing

[0113] The trained model was deployed on an ARM Cortex-M7 microcontroller (STM32H743, 400MHz), and the CMSISNN acceleration library was used for neural network inference. Breakdown of single detection cycle time:

[0114] Self-encoder encoder forward: 3.2ms;

[0115] FFT and harmonic extraction: 1.8ms;

[0116] Statistical characteristic calculation: 0.3ms;

[0117] Single-class SVM decision: 1.5ms;

[0118] Total time: 6.8ms.

[0119] Since the power frequency cycle is 20ms, the detection time of this method is much shorter than the cycle length, meeting the real-time requirements. If a delayed judgment is used (for three consecutive fault alarms), the total decision time is approximately 20.4ms, which is still acceptable.

[0120] Based on the same inventive concept as the above-described method embodiments, this invention also provides an unsupervised fault arc detection system based on a physically guided deep autoencoder and multi-source feature fusion, for executing the above-described method. Those skilled in the art will understand that the system's implementation corresponds to the above-described method embodiments, and the specific module division and functions are as follows:

[0121] Multi-source signal acquisition and preprocessing module: used to acquire the real-time current signal of the protected line, perform preprocessing through sliding window framing and normalization, and calculate the frequency domain feature vector and time domain statistical feature vector of the current analysis window;

[0122] The physically guided autoencoder module includes an encoder and a decoder with an attention mechanism, used to extract the hidden layer feature vector of the preprocessed current signal and calculate the reconstruction error scalar; its training loss includes a physical information guidance term, which includes at least one of the following: fundamental amplitude and phase consistency constraint, total harmonic distortion rate consistency constraint, and power frequency periodicity constraint.

[0123] Multi-source feature fusion module: used to concatenate the hidden layer feature vector, the reconstruction error scalar, the frequency domain feature vector and the time domain statistical feature vector in a feature layer to form a multi-source fused feature vector;

[0124] Single-class support vector machine discrimination module: used to input the multi-source fused feature vector into a pre-trained single-class support vector machine, obtain the output value of the decision function, and determine whether there is a fault arc;

[0125] Delay Decision and Output Module: Used to output a fault alarm signal when the output values ​​of the decision function in multiple consecutive analysis windows all indicate the presence of a fault arc.

[0126] Based on the same inventive concept as the above-described method embodiments, the present invention also provides a fault arc detection device, including a current sensor, a signal conditioning circuit, an analog-to-digital converter, a memory, and a processor. The current sensor is connected to the analog-to-digital converter via the signal conditioning circuit, and the analog-to-digital converter and the memory are respectively connected to the processor. When the processor executes a program stored in the memory, it implements the steps of the above-described method embodiments.

[0127] Specifically, the current sensor is mounted on the protected line to collect real-time current signals and convert them into voltage signals; the signal conditioning circuit is used to filter, amplify, and level up the voltage signals; the analog-to-digital converter is used to convert analog signals into digital signals; the processor performs the following operations: performing sliding window framing and normalization preprocessing on the digital current signals, calculating frequency domain feature vectors and time domain statistical feature vectors; inputting the preprocessed current signals into the encoder of a one-dimensional convolutional autoencoder with frozen parameters to extract hidden layer feature vectors; inputting the preprocessed current signals into a one-dimensional convolutional autoencoder with frozen parameters to obtain reconstructed outputs and calculate the reconstruction error scalar; concatenating the hidden layer feature vectors, reconstruction error scalars, frequency domain feature vectors, and time domain statistical feature vectors into a fused feature vector; inputting the fused feature vector into a pre-trained single-class support vector machine to obtain decision function values ​​and determine whether a fault arc exists; and outputting an alarm signal when all M consecutive analysis windows are judged as faults.

[0128] In summary, this invention discloses an unsupervised fault arc detection method and system based on physically guided deep autoencoders and multi-source feature fusion. The method includes: acquiring current signals and calculating their frequency domain spectrum and time domain statistical features; constructing a one-dimensional convolutional autoencoder with an attention mechanism, introducing physical guidance terms such as fundamental amplitude phase consistency, total harmonic distortion rate consistency, and period consistency into the loss function, and training only with normal samples; extracting hidden layer features, reconstruction errors, frequency domain features, and statistical features, and concatenating them to form a multi-source fusion feature vector; training a single-class SVM with the fusion features; calculating the fusion features of the test sample during detection and inputting them into the SVM to determine the fault. This invention eliminates the need for fault sample labeling, improves model generalization through physical guidance, enhances discriminative ability through multi-source feature fusion and feature error concatenation, is suitable for complex nonlinear load environments, has high robustness and real-time performance, and is applicable to fault arc early warning in various AC power supply systems.

[0129] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. An unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion, characterized in that, include: The real-time current signal of the protected line is acquired, preprocessed by sliding window framing and normalization, and the frequency domain feature vector and time domain statistical feature vector of the current analysis window are calculated. The preprocessed real-time current signal is input into the encoder of a one-dimensional convolutional autoencoder with frozen parameters to extract the hidden layer feature vector; and the preprocessed real-time current signal is input into the one-dimensional convolutional autoencoder with frozen parameters to obtain the reconstruction output and calculate the reconstruction error scalar. The hidden layer feature vector, reconstruction error scalar, frequency domain feature vector, and time domain statistical feature vector are concatenated in a feature layer to form a multi-source fusion feature vector. The multi-source fusion feature vector is input into a pre-trained single-class support vector machine to obtain the output value of the decision function, and the presence of a fault arc is determined based on the output value of the decision function.

2. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, The training process of the one-dimensional convolutional autoencoder uses only normal operating condition samples, and its training loss function includes a physical information guiding term; the physical information guiding term includes at least one of the following: fundamental amplitude and phase consistency constraint, total harmonic distortion rate consistency constraint, and power frequency periodicity constraint.

3. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 2, characterized in that, The training loss function is expressed as follows: ， in For reconstruction error, It is the sum of the squared difference between the fundamental amplitude and the squared difference between the input signal and the reconstructed signal. It is the square of the difference between the total harmonic distortion rate of the input signal and the reconstructed signal. To reconstruct the sum of squares of the waveform differences between adjacent periods of the signal, , , This is a hyperparameter.

4. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, The encoder of the one-dimensional convolutional autoencoder embeds an efficient channel attention mechanism after each convolutional layer. The efficient channel attention mechanism captures cross-channel interaction information through a one-dimensional adaptive convolutional kernel.

5. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, The frequency domain feature vector is the amplitude sequence of the first N harmonic components after performing a fast Fourier transform on the real-time current signal, where N ranges from 10 to 30.

6. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, The time-domain statistical feature vector includes at least one or more of the following: peak-to-peak value, variance, skewness, and kurtosis.

7. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, The single-class support vector machine uses a radial basis kernel function.

8. The unsupervised fault arc detection method based on physically guided deep autoencoder and multi-source feature fusion according to claim 1, characterized in that, It also includes a delayed judgment step: when the decision function output values ​​of M consecutive analysis windows all indicate the presence of a fault arc, a fault alarm signal is output; otherwise, the previous stable state is maintained; the value of M ranges from 3 to 5.

9. An unsupervised fault arc detection system based on a physically guided deep autoencoder and multi-source feature fusion, characterized in that, include: The signal acquisition and preprocessing module is used to acquire the real-time current signal of the protected line, perform preprocessing by sliding window framing and normalization, and calculate the frequency domain feature vector and time domain statistical feature vector of the current analysis window. The freeze encoder module is used to input the preprocessed real-time current signal into the encoder of the one-dimensional convolutional autoencoder with freeze parameters to extract the hidden layer feature vector. The reconstruction error calculation module is used to input the preprocessed real-time current signal into the one-dimensional convolutional autoencoder of the frozen parameters to obtain the reconstruction output and calculate the reconstruction error scalar. The multi-source feature fusion module is used to concatenate the hidden layer feature vector, reconstruction error scalar, frequency domain feature vector and time domain statistical feature vector in a feature layer to form a multi-source fused feature vector. The single-class support vector machine discrimination module is used to input the multi-source fused feature vector into a pre-trained single-class support vector machine, obtain the output value of the decision function, and determine whether there is a fault arc.

10. A fault arc detection device, comprising a current sensor, a signal conditioning circuit, an analog-to-digital converter, a memory, and a processor, wherein the current sensor is connected to the analog-to-digital converter via the signal conditioning circuit, and the analog-to-digital converter and the memory are respectively connected to the processor, characterized in that, When the processor executes a program stored in memory, it implements the steps of the method as described in any one of claims 1 to 8.

Images