An alternating current arc fault detection method based on an adaptive multi-scale time-frequency fusion network

Abstract

The application discloses an alternating current arc fault detection method based on an adaptive multi-scale time-frequency fusion network and relates to the technical field of alternating current arc fault detection. Time sequence data of current signals of different types of loads under normal working and arc fault conditions are collected, classified, arranged and labeled, and a data set is constructed; an adaptive multi-scale time-frequency fusion network is constructed, which is composed of a multi-scale time domain convolution module, a learnable frequency domain transformation module and a joint feature fusion module, input current sample sequences and outputs low-dimensional discriminant feature vectors; a two-stage model training strategy is adopted for training, input current sample sequences to be detected, and detection of alternating series arc faults and identification of load types are realized. The alternating series arc fault can be accurately and stably detected and the load type can be identified, and the problems of insufficient multi-scale feature capture, non-adaptive frequency domain feature extraction and high feature redundancy are overcome, thereby providing strong support for power safety guarantee.

Description

Technical Field

[0001] This invention relates to the field of AC arc fault detection technology, specifically an AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network. Background Technology

[0002] When an AC series arc fault occurs in a circuit, it can easily ignite surrounding flammable materials, posing a serious safety hazard. However, because the current amplitude of a series arc fault is relatively small, it is difficult to be effectively detected by traditional fuses and circuit breakers. Therefore, an accurate method for detecting series arc faults is crucial.

[0003] Currently, current signal-based detection methods have achieved some success by extracting current time-frequency domain features and combining them with machine learning models for classification. However, these methods rely on manually designed features, resulting in a complex feature extraction process and poor generalization performance. In recent years, deep learning methods have been widely used in arc fault detection; however, existing methods still have the following shortcomings: most methods use single-scale convolution kernels, making it difficult to simultaneously capture multi-scale mixed features such as high-frequency transient pulses and power frequency periodic distortions in arc currents; at the same time, frequency domain analysis relies on short-time Fourier transforms with fixed parameters, making it impossible to adaptively extract task-relevant optimal frequency domain features; furthermore, the extracted features have high dimensionality and redundancy, leading to high model computational overhead and making it difficult to meet the real-time requirements of embedded devices. Summary of the Invention

[0004] To address the shortcomings of the prior art, this invention provides an AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network. This method can accurately and stably detect AC series arc faults and identify the load type, overcoming the problems of insufficient multi-scale feature capture, non-adaptive frequency domain feature extraction, and high feature redundancy, thus providing strong support for ensuring power safety.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network, comprising the following steps:

[0006] S1. Dataset Construction: Build an AC arc fault simulation experimental platform, collect time series data of current signals of different types of loads under normal operation and arc fault conditions, classify and label the collected samples, and construct the original current sample dataset.

[0007] S2. Detection Model Construction: An adaptive multi-scale time-frequency fusion network AMTFF-Net is constructed as the detection model. The adaptive multi-scale time-frequency fusion network adopts a dual-path parallel architecture, including a time-domain path and a frequency-domain path. It consists of three parts: a multi-scale time-domain convolution module, a learnable frequency-domain transformation module, and a joint feature fusion module. The input of the detection model is the current sampling sequence, and the output is a low-dimensional discriminative feature vector.

[0008] S3. Two-stage model training: The constructed adaptive multi-scale time-frequency fusion network is trained using a two-stage model training strategy. After training, the current sampling sequence to be detected is input into the trained adaptive multi-scale time-frequency fusion network to extract the corresponding low-dimensional discriminative feature vector, which is then input into the trained random forest classifier to realize the detection of AC series arc faults and the identification of load type.

[0009] Furthermore, in step S1, the sampling frequency when acquiring the current signal is 100kHz, each sample is truncated to cover 2000 sampling points covering one power frequency cycle, the time span is 20ms, and the samples are labeled according to the load type and working status corresponding to them.

[0010] Furthermore, the operation process of the multi-scale temporal convolution module includes:

[0011] S2.1.1 Input the 1×2000 current sampling sequence into three parallel one-dimensional convolutional layers with kernel sizes of 15, 63 and 255 respectively. Each convolutional layer has 16 output channels, resulting in three sets of convolutional feature maps of different scales with a size of 16×2000.

[0012] S2.1.2 Perform global average pooling on each group of 16×2000 convolutional feature maps to compress each group of convolutional feature maps into a 1×16 feature vector;

[0013] S2.1.3. Concatenate the 1×16 feature vectors output from each branch along the channel dimension to obtain a 1×48 concatenated feature vector.

[0014] S2.1.4 Input the 1×48 concatenated feature vector into the fully connected attention network to calculate the 1×3 attention weight vector;

[0015] S2.1.5 Broadcast the attention weight vector and achieve adaptive fusion of multi-scale features through weighted summation to obtain a multi-scale temporal feature map.

[0016] Furthermore, the operation process of the learnable frequency domain transform module includes:

[0017] S2.2.1 Input the 1×2000 current sampling sequence into the first convolutional layer. The convolution kernel size is 64, the stride is 8, and the number of output channels is 32. Simulate the time-frequency analysis process of short-time Fourier transform and output a simulated spectrum with a size of 32×251. The convolution kernel size corresponds to the window length of time-frequency analysis, the sliding stride corresponds to the time window movement interval, and the number of output channels corresponds to the number of frequency segments.

[0018] S2.2.2 After batch normalization and ReLU activation function processing, the 32×251 analog spectrogram is input into the second convolutional layer with a kernel size of 3, a stride of 1, an output channel number of 32, and an output size of 32×251.

[0019] S2.2.3 After batch normalization and ReLU activation, the feature map output by the second convolutional layer is subjected to max pooling downsampling, and the final output is a frequency domain feature map with a size of 32×125.

[0020] Furthermore, the operation process of the joint feature fusion module includes:

[0021] S2.3.1 Align the multi-scale temporal feature map output by the multi-scale temporal convolution module with the frequency domain feature map output by the learnable frequency domain transformation module in terms of dimensions, and concatenate them in the channel dimension to obtain the concatenated feature map.

[0022] S2.3.2 Input the spliced ​​feature map into a 1×1 convolutional layer, and perform linear combination of all channels after splicing through 1×1 convolution to achieve cross-domain feature fusion;

[0023] S2.3.3. Perform batch normalization, ReLU activation function, adaptive global average pooling and Flatten operation on the fused feature map in sequence to convert it into a one-dimensional feature vector;

[0024] S2.3.4. Introduce a Dropout layer and set a dropout rate to prevent the model from overfitting;

[0025] S2.3.5. Input the processed feature vector into the fully connected layer and compress it into a 1×10 low-dimensional discriminative feature vector.

[0026] Furthermore, in step S3, the specific process of the two-stage model training strategy includes:

[0027] S3.1 Divide the dataset constructed in step S1 into a training set, a validation set, and a test set according to the stratified sampling method;

[0028] S3.2, First stage: Train the AMTFF-Net feature extractor, connect a fully connected classifier to the output of the joint feature fusion module, and train the entire neural network in an end-to-end manner with the output dimension of the classifier corresponding to the number of load type labels.

[0029] S3.3, Second Stage: Train the random forest classifier, remove the fully connected classifier, fix all parameters of the AMTFF-Net feature extractor, input all samples in the training and validation sets into the AMTFF-Net feature extractor with fixed parameters, extract the corresponding low-dimensional discriminative feature vectors, and input the extracted low-dimensional discriminative feature vectors into the random forest model for training.

[0030] Furthermore, the hyperparameters of the two-stage model training strategy during the training process are set as follows: AdamW optimizer is used, initial learning rate is set to 0.001, weight decay is 1e-4, batch size is set to 16, total number of training rounds is 50, learning rate decay and early stopping strategy are used to prevent overfitting, and cross-entropy loss function is used.

[0031] Furthermore, after the training is completed, the two-stage model training strategy uses a test set to test the combined model of the trained AMTFF-Net feature extractor and random forest classifier, and evaluates the test results using accuracy, precision, recall, F1 score, and false positive rate.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention can simultaneously capture high-frequency transient pulses and power frequency periodic distortion features in current signals through multi-scale time-domain convolution, overcoming the limitation of the limited receptive field of a single-scale convolution kernel. At the same time, through learnable frequency domain transformation, the optimal frequency domain transformation basis can be learned end-to-end, and frequency domain discriminative features can be adaptively extracted. In addition, through adaptive feature fusion and compression, high-dimensional features are compressed into low-dimensional discriminative vectors, which helps to significantly reduce the computational cost of the model. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method of the present invention;

[0034] Figure 2 This is an overall architecture diagram of the adaptive multi-scale time-frequency fusion network constructed in this invention;

[0035] Figure 3 yes Figure 2 Structure diagram of a multi-scale temporal convolution module;

[0036] Figure 4 yes Figure 2 The structure diagram of the learnable frequency domain transform module;

[0037] Figure 5 yes Figure 2 Structure diagram of the joint feature fusion module;

[0038] Figure 6 This is a diagram of the two-stage model training strategy in this invention;

[0039] Figure 7 This is a diagram of the experimental equipment for the AC arc fault simulation experimental platform in the embodiment;

[0040] Figure 8 These are the current time-domain waveforms under different loads in the embodiments;

[0041] Figure 9 These are current spectrum images under different loads in the embodiments;

[0042] Figure 10 This is a confusion matrix diagram of the classification results obtained from the test in the embodiment. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0044] like Figures 1-6 As shown, an AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network is presented, and its overall process combines... Figure 1 As shown, the main steps include three core steps: dataset construction, detection model construction, and two-stage model training. Current data under multiple load conditions is collected by building an AC arc fault simulation experimental platform. A dataset containing both normal operation and arc fault states is constructed. An adaptive multi-scale time-frequency fusion network (AMTFF-Net), composed of a multi-scale time-domain convolution module, a learnable frequency-domain transform module, and a joint feature fusion module, is built as a feature extractor. A two-stage model training strategy is used to complete model training, ultimately achieving high-precision detection of AC series arc faults and load type identification. Specifically:

[0045] Step 1: Dataset Construction;

[0046] An AC arc fault simulation experimental platform was built, which includes a load selection cabinet, an arc-suppressing device, and a port wiring cabinet. For different types of loads, under normal operation and arc fault conditions, time series data of current signals in the load circuit were collected at a set sampling frequency. The number of sampling points covering one power frequency cycle for each sample (hereinafter, we will use 2000 sampling points per sample with a time span of 20ms as an example) is used as the single sample length. All collected samples are classified, organized, and labeled according to their corresponding load type and operating state to construct the original current sample dataset.

[0047] Step 2: Detection model construction;

[0048] An adaptive multi-scale time-frequency fusion network (AMTFF-Net) is constructed as the detection model, and its overall architecture combines... Figure 2 As shown, a dual-path parallel architecture is adopted, including a time-domain path and a frequency-domain path, consisting of three parts: a multi-scale time-domain convolution module, a learnable frequency-domain transform module, and a joint feature fusion module. The detection model takes a 1×2000 single-cycle current sampling sequence as input and outputs a 10-dimensional fused feature as a low-dimensional discriminative feature vector, which is used by the subsequent classifier to identify fault state and load type.

[0049] S2.1, Construction of multi-scale temporal convolution module:

[0050] Structure combination of multi-scale temporal convolution modules Figure 3 As shown, this module is used to extract waveform features of current signals at different time scales, while simultaneously capturing high-frequency transient pulses and power frequency periodic distortion features. Multiple sets of one-dimensional convolutional layers with different kernel sizes are deployed in parallel to cover current features of different durations through multi-scale receptive fields. The specific implementation process is as follows:

[0051] 1. The input is a current sampling sequence of 1×2000, which is fed into three parallel one-dimensional convolutional layers with kernel sizes of 15, 63 and 255 respectively. The number of output channels of each convolutional layer is 16, resulting in three sets of convolutional feature maps of different scales with a size of 16×2000.

[0052] 2. Perform global average pooling on each group of 16×2000 convolutional feature maps to compress each group of convolutional feature maps into a 1×16 feature vector, thereby achieving global representation of features at different scales.

[0053] 3. Concatenate the 1×16 feature vectors output from each branch along the channel dimension to obtain a 1×48 concatenated feature vector;

[0054] 4. Input the 1×48 concatenated feature vector into the fully connected attention network to calculate the 1×3 attention weight vector. The three elements of the weight vector correspond to the contribution of the three branches at different scales.

[0055] 5. Broadcast the attention weight vector and achieve adaptive fusion of multi-scale features through weighted summation to obtain the final multi-scale temporal feature map.

[0056] S2.2 Construction of a learnable frequency domain transform module:

[0057] The structure of the learnable frequency domain transform module is combined Figure 4 As shown, a parameter-learnable one-dimensional convolutional layer is used to replace the traditional fixed-parameter short-time Fourier transform. During training, the convolutional kernel parameters are optimized through backpropagation with the goal of minimizing the classification loss. This enables end-to-end learning of the frequency domain transform basis for arc fault classification tasks and adaptive extraction of the optimal frequency domain discriminative features. The specific implementation process is as follows:

[0058] 1. The input is a current sampling sequence of 1×2000, which is input to the first convolutional layer with a kernel size of 64, a stride of 8, and 32 output channels. It simulates the time-frequency analysis process of short-time Fourier transform and outputs a simulated spectrum with a size of 32×251. The kernel size corresponds to the window length of the time-frequency analysis, the stride corresponds to the time window movement interval, and the number of output channels corresponds to the number of frequency bands.

[0059] 2. After batch normalization (BatchNorm1d) and ReLU activation function processing, the 32×251 simulated spectrum is input into the second convolutional layer with a kernel size of 3, a stride of 1, and 32 output channels. Higher-order frequency domain features are further extracted from the simulated spectrum, and the output size is still 32×251.

[0060] 3. The feature map output from the second convolutional layer is processed by batch normalization (BatchNorm1d) and ReLU activation function, and then max pooling is performed to achieve downsampling to reduce feature redundancy. The final output is a frequency domain feature map with a size of 32×125.

[0061] S2.3 Construction of the joint feature fusion module:

[0062] Structural integration of joint feature fusion module Figure 5 As shown, an adaptive fusion method is used to combine the multi-scale time-domain feature map output from the time-domain path with the frequency-domain feature map output from the frequency-domain path, and compress it into a low-dimensional discriminative feature vector. The specific implementation process is as follows:

[0063] 1. Align the multi-scale temporal feature map output by the multi-scale temporal convolution module with the frequency domain feature map output by the learnable frequency domain transformation module in terms of dimensions, and concatenate them in the channel dimension to obtain the concatenated feature map.

[0064] 2. Input the spliced ​​feature map into a 1×1 convolutional layer. Linearly combine all channels after splicing through 1×1 convolution to adaptively adjust the contribution of each feature channel from the time domain and frequency domain, thereby achieving the fusion of cross-domain features.

[0065] 3. Perform batch normalization (BatchNorm1d), ReLU activation function, adaptive global average pooling (AdaptiveAvgPool1d), and Flatten operations on the fused feature map in sequence to convert the two-dimensional feature map into a one-dimensional feature vector;

[0066] 4. Introduce a Dropout layer with a dropout rate of 0.3 to prevent model overfitting;

[0067] 5. Input the processed feature vector into the fully connected layer and compress it into a 1×10 low-dimensional discriminative feature vector to determine whether an arc fault has occurred, thus completing the feature extraction process.

[0068] Step 3: Two-stage model training strategy;

[0069] A two-stage model training strategy is employed to train and test the constructed adaptive multi-scale time-frequency fusion network, combining... Figure 6 As shown, by conducting phased training to balance the effectiveness of feature extraction and the accuracy of classification, the accuracy and robustness of AC series arc fault detection are improved. The specific process is as follows:

[0070] 1. Dataset partitioning: Divide the dataset constructed in step one into a training set, a validation set, and a test set using stratified sampling.

[0071] 2. First stage: Training the AMTFF-Net feature extractor

[0072] A fully connected classifier is connected to the output of the joint feature fusion module. The output dimension of the classifier corresponds to the number of load type labels. The entire neural network is trained in an end-to-end manner.

[0073] 3. Second stage: Training the random forest classifier

[0074] After the first stage of training is completed, the fully connected classifier is removed, all parameters of the AMTFF-Net feature extractor are fixed, and all samples in the training and validation sets are input into the AMTFF-Net feature extractor with fixed parameters to extract the corresponding low-dimensional discriminative feature vectors. The extracted low-dimensional discriminative feature vectors are then input into the random forest model for random forest classifier training to complete the final classification.

[0075] 4. Model Testing and Evaluation: The trained AMTFF-Net feature extractor and random forest classifier combination model is tested using a test set. The effectiveness and accuracy of the test results are evaluated using multiple metrics such as accuracy, precision, recall, F1 score, and false positive rate.

[0076] Example

[0077] The current data in this embodiment comes from an AC arc fault simulation experimental platform, combined with... Figure 7 As shown, this experimental platform is mainly used to study the impact of a series AC arc fault caused by an electric arc generated in a circuit when carbonized cable is energized on the load current. The power supply of the experimental platform is 220V, 50Hz. A current transformer is used to measure the current in the load circuit. The collected current is transmitted to the host for storage through a data acquisition unit NI / PCI6229. The sampling frequency is 100kHz. There are 7 types of loads in the experiment, as shown in Table 1:

[0078] Table 1 Load Types and Their Operating Conditions

[0079]

[0080] The aforementioned experimental platform was used to collect current signals from each load under two operating states: normal operation and arc fault. Figure 8 The figure shows the current time-domain waveforms under different load conditions, including (a) a resistor, (b) a vacuum cleaner, (c) a switching power supply, (d) an electronic light regulator, (e) a halogen lamp, (f) a fluorescent lamp, and (g) a handheld drill. It can be seen from the figure that under normal operating conditions, the current waveforms of different loads have distinct characteristics: the resistor exhibits an approximately sinusoidal wave, the switching power supply an approximately pulse wave, and the vacuum cleaner an approximately triangular wave. When an arc fault occurs, the current waveform shows abnormal characteristics such as slight distortion, reduced amplitude, current flattening, and changes in the conduction angle. Furthermore, combined with… Figure 9As shown, the current spectrum images under different load conditions are displayed, including (a) resistor, (b) vacuum cleaner, (c) switching power supply, (d) electronic light regulator, (e) halogen lamp, (f) fluorescent lamp, and (g) handheld electric drill. It can be seen from the figure that under the arc fault condition, the current amplitude at the 50Hz main frequency decreases, while the amplitude at the 3rd and 5th harmonics increases significantly. In addition, different load types also show different frequency domain characteristics.

[0081] All collected samples were organized and labeled according to their corresponding load type and operating state. Each sample consisted of 2000 sampling points, with a time span of 20ms, covering one complete cycle of power frequency AC. A dataset containing 53,639 sets of raw current samples with 14 labels was ultimately constructed. The specific sample label settings and quantities are shown in Table 2.

[0082] Table 2. Dataset label definitions and dataset size

[0083]

[0084] The process for detecting AC series arc faults using arc fault data acquired through this experimental platform mainly includes the following aspects:

[0085] First, the dataset is divided into training, validation, and test sets according to hierarchical sampling, which are used for the training and testing of the adaptive multi-scale time-frequency fusion network.

[0086] The raw current samples collected are used as model input with a single cycle (2000 sampling points). No additional time-frequency transformation preprocessing is required. The samples are directly input into the constructed adaptive multi-scale time-frequency fusion network (AMTFF-Net).

[0087] A two-stage model training strategy is adopted: In the first stage, the AMTFF-Net feature extractor is trained, and a fully connected classifier is connected after the joint feature fusion module to train the entire neural network in an end-to-end manner; In the second stage, the random forest classifier is trained, the parameters of the feature extractor are fixed, and the output 10-dimensional discriminative feature vector is input into the random forest model to complete the final classification.

[0088] The hyperparameters during training were set as follows: AdamW optimizer was used, initial learning rate was 0.001, weight decay was 1e-4, batch size was 16, and the total number of training epochs was 50. Learning rate decay and early stopping strategies were employed to prevent overfitting. The cross-entropy loss function was used. Hierarchical five-fold cross-validation was used for performance evaluation. The experimental results are shown in Table 3.

[0089] Table 3 Detailed results of five-fold cross-validation

[0090]

[0091] As shown in Table 3, the neural network classification accuracy of AMTFF-Net at each fold is no less than 99.49%, with minimal fluctuation. After adopting a two-stage model training strategy, the average accuracy of the random forest classifier trained based on 10-dimensional features is 99.77%, with a standard deviation of only ±0.10%.

[0092] To further verify the effectiveness of the extracted features, the 10-dimensional features extracted by the method of this invention were compared experimentally with 10 traditional handcrafted features (including 6 time-domain features, 2 frequency-domain features, and 2 time-frequency domain hybrid features). Under the condition that the classifier and experimental settings were exactly the same, the comparison results are shown in Table 4:

[0093] Table 4 Performance Comparison of Different Feature Extraction Methods

[0094]

[0095] As can be seen from Table 4, the features automatically extracted by the method of the present invention are significantly better than traditional manual features in all evaluation indicators. The classification accuracy is improved by about 4% compared with the best comparison feature group, the F1 score is about 4.5% higher, and the false alarm rate is only 1.2‰, which is far lower than the false alarm level of manual features.

[0096] To quantitatively evaluate the effectiveness of each core module in the adaptive multi-scale time-frequency fusion network, four sets of ablation experiments were designed, and the results are shown in Table 5:

[0097] Table 5 Ablation Experiment Results

[0098]

[0099] Ablation experiments show that the absence of any core module will lead to a decrease in the overall performance of the adaptive multi-scale time-frequency fusion network. Among them, the absence of frequency domain information has the most significant impact on classification accuracy, while the replacement of the fusion strategy will cause a significant increase in the false alarm rate. This fully demonstrates the necessity and complementarity of each module in the design.

[0100] To evaluate the generalization ability and robustness of the adaptive multi-scale time-frequency fusion network in unknown scenarios, a newly collected test set (containing approximately 4000 samples, collected at a different time than the original dataset) was used to validate the trained network. Applying a feature extractor with fixed parameters and a trained random forest classifier to this new test set yielded an overall classification accuracy of 99.19% with no false positives. Detailed classification confusion matrices were also provided. Figure 10 As shown.

[0101] According to the confusion matrix, the adaptive multi-scale time-frequency fusion network can accurately distinguish between normal operation and arc fault operation states for the vast majority of load categories. A small number of misclassifications are mainly concentrated between halogen lamps and fluorescent lamps, because both types of loads rely on electronic transformers, and their normal operation and arc fault current waveform characteristics are highly similar. It is worth emphasizing that all misclassifications occur within the same state (between normal samples or between fault samples), with no cross-state misclassifications; that is, fault samples were not classified as normal, and normal samples were not classified as faulty. This indicates that the key safety attributes of the adaptive multi-scale time-frequency fusion network of this invention are strictly guaranteed, demonstrating high reliability in practical applications.

[0102] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0103] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting AC arc faults based on an adaptive multi-scale time-frequency fusion network, characterized in that: Includes the following steps: S1. Dataset Construction: Build an AC arc fault simulation experimental platform, collect time series data of current signals of different types of loads under normal operation and arc fault conditions, classify and label the collected samples, and construct the original current sample dataset. S2. Detection Model Construction: An adaptive multi-scale time-frequency fusion network AMTFF-Net is constructed as the detection model. The adaptive multi-scale time-frequency fusion network adopts a dual-path parallel architecture, including a time-domain path and a frequency-domain path. It consists of three parts: a multi-scale time-domain convolution module, a learnable frequency-domain transformation module, and a joint feature fusion module. The input of the detection model is the current sampling sequence, and the output is a low-dimensional discriminative feature vector. S3. Two-stage model training: The constructed adaptive multi-scale time-frequency fusion network is trained using a two-stage model training strategy. After training, the current sampling sequence to be detected is input into the trained adaptive multi-scale time-frequency fusion network to extract the corresponding low-dimensional discriminative feature vector, which is then input into the trained random forest classifier to realize the detection of AC series arc faults and the identification of load type.

2. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 1, characterized in that: In step S1, the sampling frequency for acquiring the current signal is 100kHz, and each sample is extracted from 2000 sampling points covering one power frequency cycle, with a time span of 20ms. The samples are labeled according to their corresponding load type and working status.

3. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 2, characterized in that: The operation of the multi-scale temporal convolution module includes: S2.1.1 Input the 1×2000 current sampling sequence into three parallel one-dimensional convolutional layers with kernel sizes of 15, 63 and 255 respectively. Each convolutional layer has 16 output channels, resulting in three sets of convolutional feature maps of different scales with a size of 16×2000. S2.1.2 Perform global average pooling on each group of 16×2000 convolutional feature maps to compress each group of convolutional feature maps into a 1×16 feature vector; S2.1.

3. Concatenate the 1×16 feature vectors output from each branch along the channel dimension to obtain a 1×48 concatenated feature vector. S2.1.4 Input the 1×48 concatenated feature vector into the fully connected attention network to calculate the 1×3 attention weight vector; S2.1.5 Broadcast the attention weight vector and achieve adaptive fusion of multi-scale features through weighted summation to obtain a multi-scale temporal feature map.

4. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 3, characterized in that: The operation of the learnable frequency domain transform module includes: S2.2.1 Input the 1×2000 current sampling sequence into the first convolutional layer. The convolution kernel size is 64, the stride is 8, and the number of output channels is 32. Simulate the time-frequency analysis process of short-time Fourier transform and output a simulated spectrum with a size of 32×251. The convolution kernel size corresponds to the window length of time-frequency analysis, the sliding stride corresponds to the time window movement interval, and the number of output channels corresponds to the number of frequency segments. S2.2.2 After batch normalization and ReLU activation function processing, the 32×251 analog spectrogram is input into the second convolutional layer with a kernel size of 3, a stride of 1, an output channel number of 32, and an output size of 32×251. S2.2.3 After batch normalization and ReLU activation, the feature map output by the second convolutional layer is subjected to max pooling downsampling, and the final output is a frequency domain feature map with a size of 32×125.

5. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 4, characterized in that: The operation process of the joint feature fusion module includes: S2.3.1 Align the multi-scale temporal feature map output by the multi-scale temporal convolution module with the frequency domain feature map output by the learnable frequency domain transformation module in terms of dimensions, and concatenate them in the channel dimension to obtain the concatenated feature map. S2.3.2 Input the spliced ​​feature map into a 1×1 convolutional layer, and perform linear combination of all channels after splicing through 1×1 convolution to achieve cross-domain feature fusion; S2.3.

3. Perform batch normalization, ReLU activation function, adaptive global average pooling and Flatten operation on the fused feature map in sequence to convert it into a one-dimensional feature vector; S2.3.

4. Introduce a Dropout layer and set a dropout rate to prevent the model from overfitting; S2.3.

5. Input the processed feature vector into the fully connected layer and compress it into a 1×10 low-dimensional discriminative feature vector.

6. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 1, characterized in that: In step S3, the specific process of the two-stage model training strategy includes: S3.1 Divide the dataset constructed in step S1 into a training set, a validation set, and a test set according to the stratified sampling method; S3.2, First stage: Train the AMTFF-Net feature extractor, connect a fully connected classifier to the output of the joint feature fusion module, and train the entire neural network in an end-to-end manner with the output dimension of the classifier corresponding to the number of load type labels. S3.3, Second Stage: Train the random forest classifier, remove the fully connected classifier, fix all parameters of the AMTFF-Net feature extractor, input all samples in the training and validation sets into the AMTFF-Net feature extractor with fixed parameters, extract the corresponding low-dimensional discriminative feature vectors, and input the extracted low-dimensional discriminative feature vectors into the random forest model for training.

7. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 6, characterized in that: The hyperparameters of the two-stage model training strategy during training are set as follows: AdamW optimizer is used, initial learning rate is set to 0.001, weight decay is 1e-4, batch size is set to 16, total number of training rounds is 50, learning rate decay and early stopping strategy are used to prevent overfitting, and cross-entropy loss function is used.

8. The AC arc fault detection method based on an adaptive multi-scale time-frequency fusion network according to claim 6, characterized in that: After training, the two-stage model training strategy uses a test set to test the combined model of the AMTFF-Net feature extractor and random forest classifier. The test results are evaluated using accuracy, precision, recall, F1 score, and false positive rate.

Images