A series arc fault detection method based on multi-domain deep feature combination

Abstract

The present invention discloses a series arc fault detection method based on the combination of multi-domain deep features. By extracting three one-dimensional features, namely time domain features, frequency domain features and wavelet packet energy features, from the load current signal data of the electrical appliance as input, a constructed arc fault detection model extracts deep features from each domain, and the extracted deep features are combined to perform fault detection by a fully connected neural network to obtain a fault detection result. In this fault detection method, since the input features are all one-dimensional features, the problem of large computational complexity of two-dimensional features in previous image data can be effectively solved. At the same time, this detection method adopts multi-feature inputs of time domain statistical features, frequency domain statistical features and wavelet packet energy features in the current signal data to comprehensively consider the collected current signal data and improve the accuracy of the detection results.

Description

Technical Field

[0001] The present invention relates to the technical field of series arc fault detection, and in particular to a series arc fault detection method based on the combination of multi-domain deep features. Background Art

[0002] In low-voltage power distribution systems, due to the diverse load types and complex electrical wiring environments, arc faults can occur when insulation damage, loose connections, or voltage reaches a certain threshold, potentially causing electrical fires. Arc faults can be categorized as series arc faults, parallel arc faults, and ground arc faults. Series arcs, however, are low current and highly concealed, making them difficult to detect with traditional circuit protection devices. Therefore, effectively identifying series arc faults is crucial for preventing electrical fires and ensuring electrical safety.

[0003] Currently, numerous researchers are conducting research from various perspectives on detecting series arc faults in low-voltage distribution systems. Some have studied the physical characteristics of arc faults, such as arc light, arc sound, temperature, and electromagnetic radiation, and have utilized sensors and other technologies to detect arc faults. These methods have high requirements for the on-site environment and installation location, resulting in practical limitations. They are generally used to detect arcs within specific switchgear. In low-voltage distribution systems, series arc faults occur randomly, making arc fault data difficult to record. Acquiring large amounts of fault data through experiments is time-consuming and expensive. Therefore, some researchers are looking to develop mathematical arc models and simulate arc faults. Major research results include the Cassie, Mayr, Schavemaker, and Habedank arc models. Replacing complex arc experiments with computer simulations has significantly reduced costs and achieved the desired goal. However, due to limitations in parameters and application conditions, the method remains at the simulation stage. Detection methods based on both the physical characteristics of arc faults and mathematical arc models have certain limitations. Therefore, some researchers are focusing on analyzing arc current signals to detect arc faults. Arc fault detection primarily analyzes the time-frequency characteristics of current signals and identifies them. With the rise of artificial intelligence, high-performance intelligent algorithms have been widely used in fault diagnosis. Combining intelligent algorithms with arc characteristics has become a key research direction in arc fault detection. A previous time-domain visualization recognition method based on a multi-layer convolutional neural network model combined high-dimensional features of arc images with a multi-layer convolutional neural network model, demonstrating excellent recognition performance for series arcs in typical loads. The paper "Research on a Series Arc Fault Identification Method Based on an Improved CNN" improves the convolutional neural network model by combining it with the time-frequency spectrum of current, improving the accuracy and efficiency of arc fault identification. The paper "Series AC Arc Fault Detection Method Based on a High-Frequency Coupling Sensor and Convolution Neural Network" proposes a novel high-frequency coupling sensor to extract the characteristics of low-voltage series arc faults. This sensor is converted into a two-dimensional grayscale feature image and then fed into a neural network model, enabling accurate identification of both the series arc and the load type. Although the powerful visual processing capabilities of intelligent algorithms can effectively identify series arc faults, there are still problems such as slow convergence speed, large amount of calculation and single input features. Summary of the Invention

[0004] In view of this, the present invention provides a series arc fault detection method based on the combination of multi-domain deep features to solve the problems of large amount of calculation and single input features in the previous series arc fault detection.

[0005] The technical solution provided by the present invention is specifically a series arc fault detection method based on the combination of multi-domain deep features, which includes the following steps:

[0006] S1: Real-time collection of current signal data of electrical loads;

[0007] S2: extracting time domain statistical features, frequency domain statistical features and wavelet packet energy features from the current signal data respectively;

[0008] S3: Input the extracted time domain statistical features, frequency domain statistical features and wavelet packet energy features into the trained arc fault detection model to obtain the fault detection results.

[0009] Preferably, in step S3, the arc fault detection model is composed of a one-dimensional stacked neural network, a deep feature joint module and a fully connected neural network;

[0010] The one-dimensional stacked neural network is used to perform further deep feature extraction on the input time domain statistical features, frequency domain statistical features and wavelet packet energy features;

[0011] The input end of the deep feature combination module is connected to the output end of the one-dimensional stacked neural network, and is used to combine the deep features output by the one-dimensional stacked neural network to form a new feature;

[0012] The input end of the fully connected neural network is connected to the output end of the deep feature combination module, and the fault detection result is output based on the new features output by the deep feature combination module.

[0013] Further preferably, the one-dimensional stacked neural network consists of a one-dimensional convolutional layer, a batch normalization layer, a one-dimensional maximum pooling layer and a fully connected layer.

[0014] Further preferably, the training process of the arc fault detection model is as follows:

[0015] S301: collecting current signal data of the electrical load in a normal working state and in a series arc fault state respectively;

[0016] S302: Extracting time domain statistical features, frequency domain statistical features, and wavelet packet energy features from the current signal data;

[0017] S303: normalizing the time domain statistical features, frequency domain statistical features, and wavelet packet energy features, and performing redundancy removal to obtain reconstructed features of each domain;

[0018] S304: Inputting the reconstructed features of each domain into an arc fault detection model for training to obtain a trained arc fault detection model.

[0019] Further preferably, in step S303, the redundancy elimination process is specifically as follows:

[0020] The normalized time domain statistical features, frequency domain statistical features and wavelet packet energy features are sorted by importance using a tree algorithm, and the top 90-95% important features are selected as the reconstruction features of each domain.

[0021] Further preferably, in step S304, the reconstructed features of each domain are input into an arc fault detection model for training to obtain a trained arc fault detection model, specifically:

[0022] The reconstructed features of each domain of the input are subjected to deep feature extraction by a one-dimensional stacked neural network to obtain the deep features of each domain;

[0023] Combining the deep features of each domain to form a new feature;

[0024] The newly formed features are used as input to train the fully connected neural network.

[0025] Further preferably, when training the fully connected neural network, the Radam optimization algorithm is used to optimize the fully connected neural network.

[0026] The present invention provides a series arc fault detection method based on the combination of multi-domain deep features. By extracting three one-dimensional features, namely time domain features, frequency domain features and wavelet packet energy features, from the electrical load current signal data as input, a constructed arc fault detection model extracts deep features from each domain, and combines the extracted deep features. A fully connected neural network is used to perform fault detection to obtain a fault detection result.

[0027] In the series arc fault detection method based on the combination of multi-domain deep features provided by the present invention, since the input features are all one-dimensional features, it can effectively solve the problem of large computational complexity of two-dimensional features in previous image data. At the same time, the detection method adopts multi-feature inputs of time domain statistical features, frequency domain statistical features and wavelet packet energy features in current signal data to comprehensively consider the collected current signal data and improve the accuracy of the detection results.

[0028] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0031] Figure 1 A flow chart of a series arc fault detection method based on the combination of multi-domain deep features provided by an embodiment of the present invention;

[0032] Figure 2 A schematic diagram of the structure of an arc fault detection model in a series arc fault detection method based on the combination of multi-domain deep features provided by an embodiment of the present invention;

[0033] Figure 3 A schematic diagram of the training process of an arc fault detection model in a series arc fault detection method based on multi-domain deep feature combination provided by an embodiment of the present invention;

[0034] Figure 4 This is the circuit diagram of the series arc fault experimental platform;

[0035] Figure 5 The distribution diagram of the maximum value, minimum value, pulse factor and range in the extracted time domain features;

[0036] Figure 6 The ranking diagrams of the importance of time domain features using random forest, gradient boosting tree, extreme gradient boosting tree and extreme random tree respectively;

[0037] Figure 7 It is the importance ranking diagram of time domain features;

[0038] Figure 8 The training and validation accuracy curves of the arc fault model during the detection process;

[0039] Figure 9 The loss value change curve of the arc fault model during training and verification during the detection process;

[0040] Figure 10 Test accuracy curves for each detection model;

[0041] Figure 11 This is a comparison graph of different optimization algorithms. DETAILED DESCRIPTION

[0042] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0043] In order to solve the problems of large amount of computation and single input features in the past when performing series arc fault detection, this embodiment provides a series arc fault detection method based on the combination of multi-domain deep features. Figure 1 , the detection method is as follows:

[0044] S1: Real-time collection of current signal data of electrical loads;

[0045] S2: Extract the time domain statistical features, frequency domain statistical features and wavelet packet energy features from the current signal data respectively;

[0046] S3: Input the extracted time domain statistical features, frequency domain statistical features and wavelet packet energy features into the trained arc fault detection model to obtain the fault detection results.

[0047] In step S2, the extraction of frequency domain statistical features is performed by converting the detected time domain current signal data into frequency domain current signal data through Fourier transform, and then extracting the frequency domain features.

[0048] See also Figure 2 ,In step S3, the arc fault detection model is composed of a one-dimensional stacked neural network 1D-CNN, a deep feature joint module ADD, and a fully connected neural network FCNN;

[0049] Among them, the one-dimensional stacked neural network 1D-CNN is used to further extract deep features from the input time domain statistical features, frequency domain statistical features and wavelet packet energy features;

[0050] The input end of the deep feature combination module ADD is connected to the output end of the one-dimensional stacked neural network 1D-CNN, and is used to combine the deep features output by the one-dimensional stacked neural network 1D-CNN to form a new feature;

[0051] The input end of the fully connected neural network FCNN is connected to the output end of the deep feature joint module ADD, and the fault detection result is output based on the new features output by the deep feature joint module ADD.

[0052] See also Figure 3, the training process of the above arc fault detection model is as follows:

[0053] S301: collecting current signal data of the electrical load in a normal working state and in a series arc fault state respectively;

[0054] S302: Extracting time domain statistical features, frequency domain statistical features, and wavelet packet energy features from the current signal data;

[0055] S303: normalizing the time domain statistical features, frequency domain statistical features, and wavelet packet energy features, and performing redundancy removal to obtain reconstructed features of each domain;

[0056] S304: Inputting the reconstructed features of each domain into an arc fault detection model for training to obtain a trained arc fault detection model.

[0057] In step S303, the redundancy removal process is specifically as follows:

[0058] The normalized time domain statistical features, frequency domain statistical features and wavelet packet energy features are sorted by importance using a tree algorithm, and the top 90-95% important features are selected as the reconstruction features of each domain.

[0059] In order to avoid the one-sidedness of the feature importance ranking results of a single tree algorithm, a variety of different tree algorithms were selected to score the importance of each domain feature, and the comprehensive average of the scoring results of each model was used as the basis for the final feature importance ranking. The features with the top 90-95% importance were selected as the reconstruction features of each domain.

[0060] In step S304, the reconstructed features of each domain are input into the arc fault detection model for training to obtain a trained arc fault detection model, specifically:

[0061] The reconstructed features of each domain of the input are subjected to deep feature extraction by a one-dimensional stacked neural network to obtain the deep features of each domain;

[0062] Combining the deep features of each domain to form a new feature;

[0063] The newly formed features are used as input to train the fully connected neural network.

[0064] Among them, when training the fully connected neural network, it is preferred to use the Radam optimization algorithm to optimize the fully connected neural network.

[0065] In order to further explain the present invention, the present invention is further explained below in combination with our specific experimental cases.

[0066] The following describes the steps in the order of signal acquisition and feature extraction, data preprocessing and data set construction, arc fault detection model, and experimental results and analysis.

[0067] 1. Signal acquisition and feature extraction

[0068] 1.1 Signal Acquisition

[0069] Establish a series arc fault simulation experimental platform, such as Figure 4 As shown, the experimental platform mainly consists of an arc fault generating device 1, an experimental load 2, a 220V / 50Hz AC power supply 3, a sampling resistor 4, a switch 5 and an oscilloscope 6.

[0070] The arc generating device 1 is designed according to the UL1699 standard for AFCIs and primarily consists of a fixed carbon rod electrode, a moving metal electrode, and a stepper motor. According to the standard, the fixed graphite electrode is a 6.4mm diameter carbon rod, and the moving metal electrode is a 10mm diameter copper rod with a sharpened contact. The stepper motor controls the slow movement of the copper rod until an arc fault occurs.

[0071] To better simulate the actual electrical environment, six typical electrical appliances were selected as sample loads: an incandescent lamp, an incandescent lamp in series with an inductor, a hair dryer, an induction cooker, a computer, and a hand drill. Load characteristics can be categorized as linear or nonlinear. The sample load parameters are shown in Table 1.

[0072] Table 1 Experimental load and parameters

[0073]

[0074] 1.2 Time Domain Feature Extraction

[0075] Time domain features are important indicators for measuring signal characteristics. Eleven dimensionless time domain statistical features and six dimensionless time domain statistical features commonly used in signal analysis are selected to extract the time domain features of the current signal, as shown in Table 2.

[0076] Table 2 Time domain statistical characteristics

[0077]

[0078]

[0079] 1.3 Frequency Domain Feature Extraction

[0080] Fourier transform can convert time domain signals into frequency domain signals, thereby processing complex signals and extracting more feature information. For a finite length discrete signal x(n), 0≤n<N, it can be expressed as:

[0081]

[0082] The frequency domain of the current signal is analyzed using Fourier transform, and 10 frequency domain statistical features are extracted, as shown in Table 3.

[0083] Table 3 Frequency domain statistical characteristics

[0084]

[0085] Where F(k), k=1, 2, ..., N represents the frequency spectrum; N is the half length of the frequency sequence; and f(k) is the frequency component.

[0086] 1.4 Wavelet packet energy feature extraction

[0087] Wavelet packet can decompose both low-frequency and high-frequency signals, and this decomposition is non-redundant and non-omission, and can perform better time-frequency localization analysis on signals containing a large amount of medium and high-frequency information. Wavelet packet transform is an extension of wavelet transform. Orthogonal wavelet packet basis w n,j,k (t) = 2 -j / 2 w n (2 -j tk,n∈Z / Z -1 ,j,k∈Z) is orthogonalized by a scale parameter It is generated by a two-scale difference equation, which is:

[0088]

[0089] The wavelet packet decomposition coefficients are obtained by projecting the signal S(t) onto the orthogonal wavelet basis space, namely:

[0090]

[0091] When the wavelet packet basis function is a set of orthogonal bases, the wavelet packet transform has the property of energy conservation. The wavelet packet energy at a single scale is the sum of the squares of the wavelet packet coefficients at that scale.

[0092]

[0093] Considering the irregularity of arc fault current signal, the Daubechies series wavelet with variable support length is selected from the perspective of support length, vanishing moment order, regularity and symmetry. The wavelet function is db5, the decomposition level is 5, and 32 wavelet packet energy features can be extracted, which are E1-E 32 .

[0094] 2. Data preprocessing and dataset construction

[0095] Preprocessing feature data, removing redundant and invalid features, and improving data quality are beneficial to improving model detection accuracy and efficiency and building an effective data set.

[0096] 2.1 Normalization

[0097] The above extracted features are one-dimensional features of arc faults, some of which are dimensionless and are greatly affected by the load itself. In order to reduce the interference of the load itself on the characteristic parameters, all characteristic parameters are normalized and the dimensionless parameters are converted into dimensionless parameters.

[0098]

[0099] 2.2 Feature Selection

[0100] Among the initially extracted time domain, frequency domain and wavelet packet energy features, the quality of each feature cannot be determined, and there may be redundant and invalid features. Figure 5 The distribution of the maximum value, minimum value, pulse factor and range in the time domain features has 12 categories with labels ranging from 0 to 11. Figure 5 It can be seen that there is a large amount of overlap in the distribution of the four features. This type of feature has a small contribution to the model's detection results, and the model cannot effectively detect arc faults based on this feature. It may even increase the amount of computation, affecting the model's detection accuracy and effectiveness.

[0101] To address the above issues, we use a tree algorithm to sort features by importance, eliminating low-importance features while ensuring feature validity, reducing the number of features and improving data quality. To avoid the one-sidedness of the feature importance sorting results of a single tree algorithm, we selected four different tree algorithms to score the importance of each domain feature, and used the comprehensive average of the scoring results of each model as the basis for the final feature importance sorting. Figure 6 As shown in Figure 4, the ranking of the importance of time domain features by random forest, gradient boosting tree, extreme gradient boosting tree and extreme random tree is shown. The importance score of each feature is shown in Table 4.

[0102] Table 4. Time domain feature importance scores

[0103] feature RF GBDT XGBoost ExtRa Trees mean <![CDATA[T1]]> 0 0 0 0 0 <![CDATA[T2]]> 0 0 0.01 0.01 0.005 <![CDATA[T3]]> 0 0 0 0.01 0.0025 <![CDATA[T4]]> 0.04 0 0.03 0.04 0.0275 <![CDATA[T5]]> 0.03 0 0 0.04 0.0175 <![CDATA[T6]]> 0.04 0.08 0.07 0.07 0.065 <![CDATA[T7]]> 0.15 0.01 0.06 0.12 0.085 <![CDATA[T8]]> 0.06 0.05 0.07 0.06 0.06 <![CDATA[T9]]> 0.07 0.01 0 0.06 0.035 <![CDATA[T 10 ]]> 0.04 0.14 0.15 0.04 0.095 <![CDATA[T 11 ]]> 0.06 0.03 0.11 0.09 0.0725 <![CDATA[T 12 ]]> 0.08 0.01 0 0.05 0.035 <![CDATA[T 13 ]]> 0.22 0.45 0.09 0.19 0.2375 <![CDATA[T 14 ]]> 0.02 0.02 0.02 0.03 0.0225 <![CDATA[T 15 ]]> 0.12 0.12 0.27 0.13 0.16 <![CDATA[T 16 ]]> 0.02 0 0 0.02 0.01 <![CDATA[T 17 ]]> 0.03 0.09 0.11 0.05 0.07

[0104] In order to avoid excessive loss of features and affect the accuracy of model detection, the top 95% of the important features are retained. The ranking results of time domain feature importance are as follows: Figure 7 As shown. Select the kurtosis T in the time domain feature 13 , form factor T 15 , third-order central moment T 10 , standard deviation T7, third-order origin moment T 11 , margin factor T 17, mode T6, root mean square T8, mean square value T9, skewness T 12 , mean T4 and kurtosis factor T 14 , as the model input features. The same method was used to select frequency domain features and wavelet packet energy features, and finally 7 frequency domain features and 22 wavelet packet energy features were selected as model input features, as shown in Table 5.

[0105] Table 5 Feature selection

[0106]

[0107] 2.2 Dataset Construction

[0108] After normalization and feature selection, the data was labeled and hot-encoded to construct feature datasets based on the time domain, frequency domain, and wavelet packet energy. The feature data for each domain was shuffled and partitioned using the same method. 23,200 data samples were randomly selected from the 25,200 data samples in each domain and divided into training and validation sets in a 3:1 ratio. The model was trained and tested using the remaining 2,000 samples as the test set for model evaluation. The dataset composition is shown in Table 6.

[0109] Table 6 Dataset composition

[0110]

[0111] 3. Arc Fault Detection Model

[0112] According to the composition characteristics of the data set, a series arc fault detection model based on 1D-SNN is constructed, and the Aadam optimization algorithm is used to optimize the model to enhance the model robustness and improve the model performance.

[0113] 3.1 Arc Fault Detection Model

[0114] The arc fault detection model primarily consists of a one-dimensional stacked neural network (1D-CNN) and a fully connected neural network (FCNN). A deep feature combination module (ADD) is placed between the 1D-CNN and FCNN layers. This model first uses the 1D-CNN to further extract features from each domain. Next, the ADD module combines the extracted deep features to form new features. Finally, the fully connected FCNN is trained and outputs the detection results.

[0115] The one-dimensional stacked neural network 1D-CNN consists of a one-dimensional convolutional layer, a batch normalization layer, a one-dimensional maximum pooling layer, and a fully connected layer; the fully connected neural network FCNN is mainly composed of a fully connected layer, with activation functions added to the convolutional layer and the fully connected layer.

[0116] For the one-dimensional convolution layer of layer l, its output can be expressed as:

[0117]

[0118] Where, Represents the jth feature of the lth layer, j represents the number of convolution kernels, k represents the convolution kernel, * represents the convolution operator, and M represents the number of channels of the input feature. represents the input features of layer l, Represents the corresponding bias vector, and f represents the activation function used.

[0119] The convolution layer uses the ReLU activation function. For an array X consisting of n elements, the ReLU value of the i-th element in the array can be expressed as:

[0120]

[0121] In order to prevent the feature parameters from falling into the saturation zone during the transmission process, which may lead to overfitting of the model, a batch normalization layer (BN) is used to perform batch normalization on the data during the transmission process to ensure the data quality during the transmission process.

[0122] The operation principle is as follows:

[0123] Batch processing sample mean:

[0124]

[0125] Batch processing sample variance:

[0126]

[0127] Standardization processing:

[0128]

[0129] Scaling and offsetting:

[0130]

[0131] Among them, x i ∈X={x1,x2,…,x n}, X is the sample data for batch processing, y i is the output data after batch normalization, ε>0 is a constant infinitely close to 0, γ and β are reconstruction parameters, which are respectively expressed as scale parameter and offset parameter. When the processed data does not play an optimization role, β=μ X .

[0132] The one-dimensional maximum pooling layer is used to reduce data dimensionality and remove redundant data. Its output can be expressed as:

[0133]

[0134] Where, represents the neuron output of the l+1 layer, W is the pooling area, f i l (t) represents the output value of the tth neuron in the i-th feature of the l-th layer.

[0135] The fully connected layer multiplies each input signal and its corresponding weight and passes the sum through ReLU to calculate the response of the hidden layer unit. The output can be expressed as:

[0136]

[0137] Where w represents weight and b represents bias.

[0138] The output layer is a fully connected layer, and the SoftMax function is used for classification. For an array X consisting of n elements, the SoftMax value of the i-th element in the array can be expressed as:

[0139]

[0140] 3.3Radam Optimization Algorithm and Loss Function

[0141] Radam, or Rectified Adam, is an optimization algorithm with adaptive learning rates. Compared to the traditional Adam optimization algorithm, Radam uses a warm-up approach, using a small learning rate at the beginning of training and a larger learning rate later. This effectively addresses the problem of Adam easily converging to local optima and prevents excessive parameter updates in the wrong direction. The Radam optimization algorithm uses a dynamic heuristic method to automatically reduce variance. A rectifier is constructed based on the actual variance encountered, dynamically adjusting the adaptive momentum until the data variance stabilizes and the model converges to the optimal solution. The calculation process is as follows.

[0142] First, initialize the first and second moments of momentum, m0, v0 = 0, 0, and calculate the maximum length of the simple moving approximation:

[0143]

[0144] Then, the gradient at step t is:

[0145] g t =Δθf t (θ t-1 );

[0146] The second and first moments of momentum are:

[0147]

[0148] m t =β1m t-1 +(1-β1)g t ;

[0149] The momentum bias correction value is:

[0150]

[0151] The moving length is:

[0152]

[0153] When ρ t >4:00

[0154]

[0155]

[0156]

[0157] When ρ t ≤4:

[0158]

[0159] where α t is the step size, {β1,β2} is the decay rate, is the deviation correction value of the second-order moment of momentum, r t is the variance correction value. t >4, θ t is the adaptive momentum update parameter. When ρ t ≤4, θ t For non-adaptive momentum update parameters, the SGDM method is used to update the parameters.

[0160] The loss function uses the classification cross entropy function, and its calculation formula is:

[0161]

[0162] Where, represents the model predicted output, and y represents the true output distribution.

[0163] 4. Experimental results and analysis

[0164] The processed data is fed into a 1D-SNN series arc fault detection model for training, which then outputs the detection results. Simultaneously, single-domain feature detection algorithms and other optimization algorithms are explored and compared with the established model.

[0165] 4.1 Analysis of test results

[0166] The divided training set and validation set are input into the model for training and testing. Each batch is defined to contain 256 data samples. The number of iterations is determined by the callback function. The monitoring point is the accuracy on the training set, accurate to three decimal places, with a continuous step size of 10. The loss function uses the classification cross function and the Radam optimization algorithm to automatically adjust the learning rate. The training and validation accuracy of the 1D-SNN arc fault model during the detection process is as follows: Figure 8 As shown, the loss value changes as Figure 9 shown.

[0167] from Figure 8 、 9 As can be seen, in the early stages of training, training and validation accuracy increased rapidly, while loss decreased rapidly. Training accuracy and loss stabilized after approximately 10 iterations. Validation accuracy and loss also stabilized after approximately 15 iterations. Ultimately, both training and validation accuracy reached over 99%, and loss dropped below 0.1, demonstrating no overfitting. To mitigate randomness in test results, cross-validation was used to verify the effectiveness of the constructed model. The training and validation sets were repartitioned into four groups. Each data set was evaluated, and the remaining three groups were used for training. The results are shown in Table 7.

[0168] Table 7 Cross validation

[0169] Serial number Loss value Accuracy (%) 1 0.0567 99.62 2 0.0515 99.56 3 0.0477 99.68 4 0.0542 99.60 average value 0.0525 99.63

[0170] To verify the generalization ability of the constructed model, we evaluated it using a test set. The composition of the test set and the test results are shown in Table 8. Across the six typical workloads, the highest detection accuracy was 100%, and the lowest was 98.76%. Of the 2000 test samples, 1995 were successfully detected, while 5 failed, resulting in a detection accuracy of 99.75%. This demonstrates good detection results and strong generalization ability of the model.

[0171] Table 8 Test set and test results

[0172] Label sample correct mistake Accuracy 0 166 166 0 100％ 1 151 151 0 100％ 2 167 167 0 100％ 3 152 152 0 100％ 4 168 168 0 100％ 5 166 166 0 100％ 6 172 172 0 100％ 7 182 181 1 99.45％ 8 161 159 2 98.76％ 9 185 184 1 99.46％ 10 176 175 1 99.43 11 154 154 0 100％ total 2000 1995 5 99.75％

[0173] 4.2 Comparison with Single Domain Feature Detection Algorithms

[0174] Aiming at the problem of single input feature in series arc fault detection, a detection method based on multi-domain deep feature combination is proposed. In order to verify the effectiveness and superiority of the proposed method for series arc fault detection, the single-domain feature detection algorithm is explored, and 1D-CNN fault detection models based on time domain features, frequency domain features and wavelet packet energy features are constructed respectively. It is defined that each training batch contains 256 data samples, the number of iterations is 50, the loss function uses the classification cross function, and the optimization algorithm is Radam. The test accuracy of each detection model is as follows Figure 10 As shown in Figure 2, the test accuracy rates for models using single-domain feature detection were 95.20%, 96.79%, and 89.06%, respectively. Frequency domain features performed best, followed by time domain features, and wavelet packet energy features performed worst. Compared to single-domain feature detection, the proposed method, which utilizes multi-domain deep features for combined detection, achieved higher accuracy, achieving a test accuracy of 99.63%.

[0175] 4.3 Comparison of Optimization Algorithms

[0176] In the field of deep learning, the choice of optimization algorithm is crucial and seriously affects the performance of the model. In order to obtain the best detection effect, Adam, SGD, RMSprop and Radam optimization algorithms are compared on the same dataset. The results are as follows: Figure 11 As shown in the figure, compared to SGD and RMSprop, Adam and Radam can dynamically adjust the learning rate, resulting in less fluctuation in loss. After 10 iterations, the loss stabilizes, ultimately achieving good convergence. Compared to traditional Adam, Radam's loss is more stable, exhibiting stronger robustness and providing better model stability. Experimental results demonstrate that the Radam optimization algorithm has a superior optimization effect on the detection model, effectively improving model performance and achieving better detection results.

[0177] Other embodiments of the present invention will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the invention being indicated by the following claims.

[0178] It should be understood that the present invention is not limited to the above description and that various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A series arc fault detection method based on multi-domain deep feature combination, characterized in that: The steps include: S1: Real-time collection of current signal data of electrical loads; S2: extracting time domain statistical features, frequency domain statistical features and wavelet packet energy features from the current signal data respectively; S3: Input the extracted time domain statistical features, frequency domain statistical features, and wavelet packet energy features into the trained arc fault detection model to obtain fault detection results; In step S3, the arc fault detection model is composed of a one-dimensional stacked neural network, a deep feature joint module and a fully connected neural network; The one-dimensional stacked neural network is used to perform further deep feature extraction on the input time domain statistical features, frequency domain statistical features and wavelet packet energy features; The input end of the deep feature combination module is connected to the output end of the one-dimensional stacked neural network, and is used to combine the deep features output by the one-dimensional stacked neural network to form a new feature; The input end of the fully connected neural network is connected to the output end of the deep feature combination module, and the fault detection result is output according to the new features output by the deep feature combination module; The one-dimensional stacked neural network consists of a one-dimensional convolutional layer, a batch normalization layer, a one-dimensional maximum pooling layer and a fully connected layer; The training process of the arc fault detection model is as follows: S301: collecting current signal data of the electrical load in a normal working state and in a series arc fault state respectively; S302: Extracting time domain statistical features, frequency domain statistical features, and wavelet packet energy features from the current signal data; S303: normalizing the time domain statistical features, frequency domain statistical features, and wavelet packet energy features, and performing redundancy removal to obtain reconstructed features of each domain; S304: Inputting the reconstructed features of each domain into an arc fault detection model for training to obtain a trained arc fault detection model; In step S303, the redundancy removal process is specifically as follows: The normalized time-domain statistical features, frequency-domain statistical features, and wavelet packet energy features were ranked by importance using a tree algorithm, and the top 90-95% of the important features were selected as the reconstruction features for each domain. Among them, four different tree algorithms were selected to score the importance of each domain feature, and the comprehensive average of the scoring results of each model was used as the basis for the final feature importance ranking. The four tree algorithms were random forest, gradient boosting tree, extreme gradient boosting tree, and extreme random tree. In step S304, the reconstructed features of each domain are input into the arc fault detection model for training to obtain a trained arc fault detection model, specifically: The reconstructed features of each domain of the input are subjected to deep feature extraction by a one-dimensional stacked neural network to obtain the deep features of each domain; Combining the deep features of each domain to form a new feature; The newly formed features are used as input to train the fully connected neural network.

2. The method for detecting series arc faults based on the combination of multi-domain deep features according to claim 1, characterized in that: When training a fully connected neural network, the Radam optimization algorithm is used to optimize the fully connected neural network.

Images