CN117909818A - Arc fault identification method and system based on multi-cycle features and multi-scale convolution - Google Patents

Abstract

The present invention discloses an arc fault identification method and system based on multi-periodic features and multi-scale convolution. The present invention uses FFT transformation to obtain frequency and corresponding amplitude, realizes the selection of periodic features and the two-dimensional representation of one-dimensional time series data. The commonly used time series data is expanded into data that can be interpreted in two dimensions and a convolution block is designed to expand the analysis based on the attention mechanism and multi-scale, so as to obtain an improved fault arc identification model; comparative experiments are carried out in a fault arc data set covering samples of various load types and different sampling resolutions. The present invention is reasonably designed, and the features in the frequency domain are used to select the period without artificial feature extraction in advance, and then the local perception, weight sharing and multi-layer abstraction of the convolutional neural network are used to complete the feature extraction. Under the premise of meeting the fault identification accuracy, the model parameters are reduced as much as possible to optimize the network architecture, and it can be widely used in the field of household appliance safety.

Description

Arc fault identification method and system based on multi-cycle features and multi-scale convolution

Technical Field

The present invention belongs to the technical field of arc fault identification, and relates to an arc fault identification method and system based on multi-cycle features and multi-scale convolution.

Background technique

With the continuous development of society, the degree of electrification is getting higher and higher, so electrical fires have occurred frequently in recent years. Many serious fire accidents are simply caused by fault arcs in the line that are lower than the rated current or expected short-circuit current. These dangerous arcs may occur in unreasonable or aging power supply lines, electrical plugs, and power cords of household appliances, internal wiring harnesses, or component insulation. And when a fault arc occurs, it may not be quickly detected by the leakage, overcurrent, and short-circuit protection devices on the line, making it difficult to cut off the power supply in time, which can easily cause a fire. Therefore, as one of the direct causes of electrical fires, real-time and accurate identification of fault arcs is very important for various safe electricity use scenarios.

There are many types of household loads, and the electrical characteristics of different types vary greatly, which brings great challenges to arc fault identification. Researchers have never stopped studying arcs. Physical changes such as arc light, heat, and electromagnetism are generally used for arc identification in confined spaces. The recognition accuracy of a single detection sensor is not as good as the joint action of multiple sensors. However, this type of acquisition has high requirements for sensors, installation, and environment, so its practical application is very limited. Based on this, arc fault detection methods proposed by the time-frequency domain characteristics of arc voltage and current have gradually developed, but manual extraction is affected by prior knowledge, and there is still room for improvement in adaptability and generalization ability, but this optimization is difficult and cumbersome. At the same time, with the continuous increase of nonlinear loads and the existence of multiple loads running at the same time, the manually extracted features often miss or misdetect on difficult samples like this, and it is difficult to ensure recognition accuracy.

In recent years, there are two main types of research on arc fault identification: 1. Arc feature extraction method: mainly analyze the time-frequency domain characteristics of the arc, and use mathematical morphology to extract effective features. Combine machine learning classifiers or artificial neural networks to obtain the final output. 2. Research on deep learning models and algorithms: Most of them extract features by designing different convolutional networks, and some methods preprocess the time series data into images and input them into subsequent classification models for processing. These methods analyze the impact of fault arcs on current and voltage, the impact of different loads on fault arc characteristics, and apply heuristic methods or machine learning to make decisions. Among them, deep learning models have been repeatedly proven to have significant advantages in various performance indicators in arc fault identification.

Despite the advantages of deep learning over manual feature extraction, there are still some challenges in the application of deep learning in arc fault identification. First, there is no good balance between model size and performance. In order to better fit the fault arc scenario, a deeper network and many convolution kernels of different scales are usually used. However, such models usually have a large number of parameters and are not easy to meet the lightweight requirements of practical applications. Secondly, with the continuous development of convolutional neural networks, most of them are still applicable to two-dimensional data. Although some studies on fault arcs preprocess one-dimensional time series data into images to facilitate model migration, such processing usually leads to an exponential increase in the size of the input data, which affects subsequent processing and increases the amount of calculation.

Summary of the invention

The purpose of the present invention is to address the deficiencies of the prior art and provide an arc fault identification method and system based on multi-cycle features and multi-scale convolution, which performs a two-dimensional interpretation and expansion of time series data, and ensures that the increase in the size of multi-cycle data compared to the original data is linear, rather than exponential, while facilitating subsequent data input. This method integrates the adaptive features of multi-scale convolution and multi-cycle, and cooperates with the convolution attention module to extract effective arc features. In order to streamline the model size, the convolution kernel size is limited to ensure recognition accuracy, and the maximum convolution kernel size is limited to 3, and the depth of the neural network is controlled within 3 layers. Finally, the fully connected layer is used to realize the recognition of arc faults.

In a first aspect, the present invention provides an arc fault identification method based on multi-cycle features and multi-scale convolution, the method comprising the following steps:

Step (1), collecting arc current and voltage data, and slidingly dividing them into fixed time windows;

Step (2), two-dimensional interpretation expansion:

First, feature embedding is performed on the current and voltage data processed in step (1); second, the first k different frequency corresponding periods with the highest channel amplitude mean are selected by fast Fourier transform; finally, the current and voltage data processed in step (1) are converted into two-dimensional data under k different frequency corresponding periods, k ≥ 1, to obtain two-dimensional data with periodic characteristics;

Step (3), multi-scale convolution based on attention mechanism:

The two-dimensional data with periodic features obtained in step (2) is cyclically input into a multi-scale convolution module to obtain a plurality of output data with multi-scale features having attention performance; wherein the multi-scale convolution module includes two multi-scale convolution operations and a convolution attention module arranged in parallel, and a GELU activation function is applied after each multi-scale convolution operation; each output corresponds to a selected period;

Step (4), feature fusion based on adaptive weights: based on the amplitude mean of the frequency corresponding period, multiple multi-scale features with attention performance are fused with multi-period features;

Step (5), linear classification:

The fusion result of the multi-cycle features in step (4) is layer normalized and GELU activated in turn, and finally flattened and input into the fully connected layer to obtain the classification result.

Preferably, step (2) is:

2-1 Perform position embedding on the data processed in step (1);

2-2 Mark and embed the data processed in step (1);

2-3 Add the position embedding result obtained in step 2-1 and the tag embedding result obtained in step 2-2;

2-4 Perform fast Fourier transform on the addition result of step 2-3, and then compare the amplitude mean of the signal on all channels at different frequencies, select the frequencies corresponding to the k highest amplitude mean values and form a frequency list, and store the corresponding channel amplitude mean values in a weight matrix W;

2-5 Perform a two-dimensional interpretation of the original signal after feature embedding processing according to the frequency list returned in step 2-4, and rearrange the original signal with T _k = 1/f _k as the period length; if N is not divisible by T _k , it is necessary to pad the signal with zeros at the end so that the signal length becomes N' _k , ensuring that N' _k is divisible by T _k , and obtain k two-dimensional data with periodic features. Each column represents the time point with the same phase during the N'/T _k cycles divided in steps 2-4, and each row/> Represents data within a period of T _k .

Preferably, the position embedding in step 2-1 adopts a sine-cosine position encoding method to generate a continuous vector for each time point in the current and voltage data, and this vector can capture the relative relationship between the positions;

Among them, pos represents the position of the data, d represents the dimension of the position encoding, and i represents the index of the dimension.

Preferably, the marker embedding in step 2-2 is convolved using a circular filling method to achieve the representation of discrete collection values using a set of continuous vectors.

Preferably, steps 2-4 are as follows:

First, the kth element of the complex sequence _Xi on the original signal channel i is calculated by fast Fourier transform to obtain the complex amplitude X[k] _i :

Where N represents the length of the signal, k represents the index of the converted frequency, _xi [n] represents the nth element of the original signal sequence on the i-th signal channel after feature embedding processing, and j represents the imaginary unit;

Secondly, the complex amplitude X[k] _{i is} used to calculate the amplitude corresponding to the frequency f _k And calculate the amplitude mean AMP _k on all channels:

Where Re(·) represents the real part, Im(·) represents the imaginary part;

Finally, the channel amplitude means AMP _k corresponding to different frequencies f _k are compared, the frequencies corresponding to the highest k channel amplitude means are selected and returned in the form of a list, and the corresponding channel amplitude means are stored in the weight matrix W = {AMP ₁ , AMP ₂ , ..., AMP _k }.

Preferably, step (3) is:

3-1 The two-dimensional data with periodic features obtained in step (2) is passed through a multi-scale convolution module to obtain multiple scale feature vectors {Feature ₁ , Feature ₂ ,..., Feature _n };

The multi-scale convolution module includes a 1×3 convolution block, a 3×1 convolution block, and a 1×1 convolution block arranged in parallel; wherein the expansion rates of the 1×3 convolution block and the 3×1 convolution block are different;

3-2 The multiple scale feature vectors {Feature ₁ , Feature ₂ , ..., Feature _n } obtained in step 3-1 are stacked to generate a new dimension, and then the average is calculated on this new dimension to obtain the output result of the multi-scale convolution, and the GELU activation function is used on the output result to obtain the final result:

X _H*W*C = GELU(Avg(Stack(Feature ₁ ,Feature ₂ ,...,Feature _n )) Formula (6)

Stack(·) indicates stacking operation, Avg(·) indicates averaging; H, W, C indicate the length, width, and height dimensions of the X _H*W*C feature map;

3-3 Calculate channel attention and spatial attention for the final result X _H*W*C of step 3-2 respectively;

The calculation method of the channel attention is to first perform average pooling AvgPooling and maximum pooling MaxPooling along the channel on the final result X _H*W*C after step 3-2, and obtain two channel weights A _c and _Mc , and then these two channel weights are passed through a shared MLP to obtain their respective new attention weights A' _c and Mc' _c , and then the final channel attention ChannelAttention(X _H*W*C ) is obtained by addition and sigmoid function mapping;

ChannelAttention(X _H*W*C )＝Sigmoid(A' _c +M' _c ) Formula (7)

The calculation method of the spatial attention is to perform average pooling and maximum pooling on the spatial feature map along the channel according to the final result X _H*W*C of step 3-2, and then stack the two feature outputs, and then pass the stacked data through a point convolution and sigmoid to obtain the final spatial attention;

3-4 Multiply the final result X _H*W*C obtained in step 3-2 by the channel attention ChannelAttention(X _H*W*C ) calculated in step 3-3 to obtain feature X1; then multiply feature X1 by the spatial attention SpatialAttention(X _H*W*C ) calculated in step 3-3 to obtain the final attention-based feature output X2;

X1＝X _H*W*C ×ChannelAttention(X _H*W*C ) Formula (8)

X2＝X1×SpatialAttention(X _H*W*C ) Formula (9)

3-5 Take the attention-based feature output X2 of step 3-4 as the input of the multi-scale convolution module, repeat steps 3-1 to 3-2, and get the final output

Preferably, step (4) is specifically:

4-1 Transform the T _k obtained in step (3) into the two-dimensional data output Restore to one-dimensional data X ^k , and then cut the length of the one-dimensional data X ^k to the window length of step (1), that is, the data corresponding to the time point exceeding the length of the original signal is discarded;

4-2 stack the one-dimensional data corresponding to each T _k along one dimension, and then multiply it by the weight matrix W obtained in step 2-4, and finally form a feature output based on multi-cycle fusion;

4-3 Based on the idea of residual design block, the output obtained in step 4-2 is added to the k two-dimensional data with periodic features obtained in step 2-5 to obtain the fusion result of multi-periodic features.

In a second aspect, the present invention provides an arc fault identification system for implementing the above method, comprising:

Data acquisition module, used to collect arc current and voltage data;

The data preprocessing module performs sliding division of the arc current and voltage data in fixed time windows;

The two-dimensional interpretation extension module embeds the features of the current and voltage data processed by the data preprocessing module, and then converts the current and voltage data into two-dimensional data corresponding to k different frequencies to obtain two-dimensional data with periodic features;

The feature extraction module cyclically inputs the two-dimensional data with periodic features output by the two-dimensional interpretation and expansion module into the multi-scale convolution module to obtain multiple multi-scale features with attention representation;

The feature fusion module based on adaptive weights fuses multiple multi-scale features with attention performance into multi-period features based on the frequency corresponding period;

The linear classifier performs layer normalization and GELU activation on the fusion results of multi-cycle features in sequence, and finally flattens them and inputs them to the fully connected layer to obtain the classification result.

In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

In a fourth aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

The beneficial effects of the present invention are:

1. The present invention performs position embedding and label embedding on the original signal data to avoid losing partial phase information of the time point relative to the entire time sample during the subsequent two-dimensional explanatory expansion, thereby improving the robustness of the current and voltage data with respect to periodic changes during the two-dimensional explanatory expansion, thereby achieving the accuracy of fault identification.

2. The present invention proposes to analyze the frequency domain data of current and voltage through FFT, and extract the corresponding period in the frequency domain as a two-dimensional explanatory extension of the time series data. This is because the features in the frequency domain are usually considered effective in the electrical field, but feature engineering is cumbersome. While avoiding feature engineering, the present invention introduces abstract frequency domain features from the perspective of two-dimensional interpretation, which is reflected in the selection of the period and the final fusion based on the periodic features.

3. The present invention selects a fixed-size convolution kernel in the design of the parallel convolution block, namely, a 1×3 convolution block and a 3×1 convolution block, and ensures the acquisition of multi-scale features by using dilated convolution. The multi-scale features allow the model to understand signals with a larger time span and obtain larger local and deeper detail features. In order to further reduce the parameters of the model, spatially separable convolution is used, which greatly reduces the amount of parameter calculation. At the same time, a convolution attention module is added to the multi-scale convolution module, and attention calculations are performed on the space and the channel respectively, which not only allows the model to focus more on the sampling information at the time point during the suspected fault period, but also enhances the sensitivity of the effective arc features.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of the structure of the system of the present invention.

Figure 2 shows a spatially separable dilated convolution method (taking r = 2 as an example);

FIG3 is a schematic diagram of the structure of a channel attention module;

FIG4 is a schematic diagram of the structure of a spatial attention module;

FIG5 is a loss variation curve of the present invention during the training process on a validation set of high-resolution collected data;

FIG6 is a curve showing the accuracy change of the training process of the present invention on the validation set of high-resolution collected data;

FIG7 is an ROC curve drawn after the present invention performs an ablation experiment on the cycle selection part and the attention convolution module of the model.

Detailed ways

Further analysis is given below in conjunction with specific embodiments and drawings.

In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

In the process of arc fault detection, due to the increasing number of nonlinear loads and the existence of multiple loads running at the same time, it is difficult to ensure the recognition accuracy of voltage and current feature extraction. Therefore, in order to better improve the accuracy of arc fault recognition, the following problems need to be solved: how to avoid losing data cycle changes during feature extraction, and how to mine data features that are not affected by nonlinearity and multiple loads as much as possible.

Based on this, an embodiment of the present application provides an arc fault identification method based on multi-cycle features and multi-scale convolution. The present invention performs a two-dimensional interpretation and expansion on the time series data, which facilitates the subsequent data input while ensuring that the increase in the size of the multi-cycle data is linear, rather than exponential, compared to the original data. Through adaptive feature fusion based on multi-scale convolution and multi-cycle, effective arc features are extracted in conjunction with multi-scale convolution based on the attention mechanism. For the purpose of streamlining the model size, the convolution kernel size is restricted to ensure recognition accuracy. The maximum convolution kernel size is 3. At the same time, the depth of the neural network is controlled within 3 layers. Finally, the fully connected layer is used to realize the recognition of arc faults.

The arc fault recognition method based on multi-cycle features and multi-scale convolution, as shown in Figure 1, includes the following steps:

Step (1), two-dimensional interpretation and expansion: The original collected data is input in the form of a window, and some features are embedded. Then, different frequency periods are selected through FFT and the corresponding amplitude mean values are retained. The input is converted into two-dimensional data at different frequency periods so that it can be input into the subsequent multi-scale convolution layer. The specific steps are as follows:

Step 1: The current and voltage data collected by the sensor are divided into sliding windows with a certain time window size. For the data collected with high resolution (10KHz), the window size is set to 1000, that is, it takes 0.1s to identify whether there is a fault. For the data collected with low resolution (6400Hz), the window size is set to 1024, that is, it takes 0.16s to identify whether there is a fault.

Step 2: Position embedding is performed on the data in Step 1. This is because some global phase information in the original waveform will be lost during the two-dimensional interpretation process. Here, the position embedding adopts the sine-cosine position encoding method to generate a continuous vector for each time point in the sequence. This vector can capture the relative relationship between positions.

Among them, pos represents the position of the data, d represents the dimension of the position encoding, and i represents the index of the dimension.

Step 3: Similarly, the data in Step 1 is labeled and embedded. This is because important information may be lost when dividing the window, and the data on the edge may not be completely covered by the convolution kernel, so the circular filling method is used for convolution. At the same time, the discrete collection values can be represented by a set of continuous vectors.

Step 4: Add the position embedding result obtained in Step 2 and the tag embedding result obtained in Step 3, and then output them for subsequent cycle selection.

Step 5: Perform a fast Fourier transform (FFT) on the output of Step 4, then compare the amplitude mean of the signal at different frequencies on all channels, select the three highest frequencies and form a frequency list. And store the corresponding amplitude mean in a weight matrix. Specifically:

First, the kth element of the complex sequence _Xi on the original signal channel i is calculated by FFT:

Where N represents the length of the signal, k represents the index of the converted frequency, x _i [n] represents the nth element of the original sequence on the i-th signal channel after processing in Step 2 and Step 3, and j represents the imaginary unit (j ² = -1);

Secondly, the complex amplitude X[k] _{i is} used to calculate the amplitude corresponding to f _k and the average of the absolute values of the amplitudes on all channels is calculated:

Finally, the channel amplitude means AMP _k corresponding to different f _k are compared, the frequencies corresponding to the highest k means are selected and returned in the form of a list, and the corresponding amplitudes are stored in the weight matrix W = {AMP ₁ , AMP ₂ , ..., AMP _k }.

Step 6: Perform a two-dimensional interpretation of the original signal after feature embedding according to the frequency list returned in Step 4, and rearrange the original signal with T _k = 1/f _k as the period length; if N is not divisible by T _k , it is necessary to pad the signal with zeros at the end so that the signal length becomes N' _k , ensuring that N' _k is divisible by T _k , and obtain three two-dimensional data with periodic features.

Step (2), multi-scale convolution based on attention mechanism:

Three two-dimensional data with periodic features are cyclically input into the multi-scale convolution module. Each multi-scale convolution module contains two multi-scale convolution operations and one convolution attention module, and the GELU activation function is applied after each multi-scale convolution operation. Finally, multiple output data with multi-scale features of attention performance are obtained, and each output corresponds to a selected period. The specific steps are as follows (taking a convolution block as an example):

Step 1: Input the three two-dimensional data with periodic features of the current cycle, and pass through a multi-scale convolution module to obtain three scale features Feature ₁ , Feature ₂ , and Feature ₃ ; as shown in Figure 2, the multi-scale convolution module includes convolution kernels of 1×3, 3×1, and 1×1. The traditional 3×3 convolution kernel is split into 1×3 convolution kernels and 3×1 convolution kernels to reduce the amount of model calculation. At the same time, in order to capture multi-scale features, different expansion rates r＝{1,2,3} are introduced. In this module, a total of 2×3+1＝7 convolution operations are performed in parallel. At the same time, in order to ensure that the shapes of the parallel convolution outputs are exactly the same, it is necessary to design corresponding padding during convolution.

Step 2: After all the parallel convolutions in Step 1 are completed, the multiple outputs are stacked along a new dimension to generate a new dimension, and then the average is calculated on this new dimension to obtain the output result of the multi-scale convolution, and the GELU activation function is used on the output result to obtain the final result X _H*W*C ; this dimension represents the number of convolution kernels designed in the multi-scale convolution module.

X _H*W*C =GELU(Avg(Stack(Feature ₁ ,Feature ₂ ,Feature ₃ ))

Step 3: After the multi-scale convolution modules and activation functions of Step 1 and Step 2, the spatial attention and channel attention are calculated on the output respectively.

As shown in FIG3 , the calculation method of channel attention is to first perform average pooling AvgPooling and maximum pooling MaxPooling along the channel on the final result X _H*W*C after step 3-2, and obtain two channel weights A _c and _Mc . Then, these two channel weights are passed through a shared MLP to obtain their respective new attention weights A' _c and Mc ' _c , and then the final channel attention ChannelAttention(X _H*W*C ) is obtained by addition and sigmoid function mapping.

ChannelAttention(X _H*W*C )＝Sigmoid(A' _c +M' _c )

As shown in Figure 4, the calculation method of spatial attention is to perform average pooling and maximum pooling on the spatial feature map along the channel X _H*W*C , then stack the two feature outputs along a new dimension, pass the stacked data through a point convolution and finally pass through sigmoid to obtain the final spatial attention.

Step 4: Multiply the final output X _H*W*C obtained in Step 2 by the channel attention ChannelAttention(X _H*W*C ) calculated in Step 3, and then multiply it by the spatial attention SpatialAttention(X _H*W*C ) calculated in Step 3 to obtain the final attention-based feature output X2.

X1＝X _H*W*C ×ChannelAttention(X _H*W*C )

X2＝X1×SpatialAttention(X _H*W*C )

Step 5: Use the result X2 of Step 4 as the input of the multi-scale convolution module, repeat Step 1 and Step 2 to get the final output

Step (3), feature fusion based on adaptive weights: The fusion of multi-periodic features is achieved based on the amplitude of the corresponding period returned by FFT in the two-dimensional interpretation extension module.

Step 1: The cycle-based two-dimensional data corresponding to each selected cycle T _k The data is restored to one-dimensional data X ^k and cut to the original window size. The corresponding values at time points beyond the window are discarded.

Step 2: The one-dimensional feature map ^Xk corresponding to each selection cycle _Tk is stacked along one dimension and then multiplied by the weight matrix W retained in the previous work to finally form a feature output based on multi-cycle fusion.

Step 3: Add the output of Step 2 to the original Step 4 in the two-dimensional expansion module to obtain the fusion result of multi-period features, which is used as the output. Here, the idea of a residual design block is adopted.

Step (4), linear classification: The fusion results of multi-cycle features are layer normalized and GELU activated in turn. The dropout parameter is set to 0.2 to prevent the model from overfitting, and finally the flattened input is sent to the fully connected layer. The output dimension of the fully connected layer is 2, corresponding to the scores of the two states. The state is whether there is a fault arc in the sample.

The above model is trained. The loss function used in the present invention is the cross entropy loss function. The optimization method designed is the Adam optimizer. The batch size is 128, the initial size of the learning rate is 0.001, and a learning rate scheduler is designed to update the learning rate with step_size=5 and reduce the learning rate with the set gamma=0.2. The early stopping method is used to monitor the training process. It is found that the loss value of the validation set in 5 epochs does not drop a certain threshold, so the training is terminated and the best three model output results are retained. The data set used in the experiment is collected at a resolution of 10kHZ, and then divided into two data sets, HR-single and HR-mix. The HR-single data set contains independent collection data of six different electrical appliances, namely electric kettles, dishwashers, refrigerators, microwave ovens, rice cookers and induction cookers, and HR-mix is nine different combinations of these appliances. The loss rate and accuracy change curves of the training process on the validation set of high-resolution collection data are shown in Figures 5 and 6. In order to prove the effectiveness of the module, ablation experiments were carried out on the cycle selection part and the attention convolution module of the model and the ROC curve was plotted, as shown in Figure 7, where w/a means removing the attention convolution part, w/p means removing the cycle selection part, and w/a+p means removing both.

An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a memory and a processor. The memory stores executable code. When the processor executes the executable code, the method described in any one of the implementation modes is implemented.

The memory may include high-speed random access memory (RAM), and may also include non-volatile memory (Non-volatile Memory), such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

The bus may be an ISA bus, a PCI bus or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc.

The memory is used to store the program, and the processor executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any of the embodiments of the present invention can be applied to the processor or implemented by the processor.

The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly embodied as a hardware decoding processor to be executed, or a combination of hardware and software modules in the decoding processor to be executed. The software module may be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can be referred to the previous method embodiment, which will not be repeated here.

If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. Arc fault identification method based on multicycle features and multiscale convolution, characterized in that the method comprises the following steps:

step (1), collecting current and voltage data of an electric arc, and carrying out sliding division on the current and voltage data in a fixed time window;

step (2), two-dimensional interpretation expansion:

Firstly, embedding characteristics of current and voltage data processed in the step (1); secondly, selecting the first k different frequency corresponding periods with the highest channel amplitude mean value through fast Fourier transformation, wherein k is more than or equal to 1; finally, converting the current and voltage data processed in the step (1) into two-dimensional data under the corresponding periods of k different frequencies to obtain two-dimensional data with periodic characteristics;

step (3), multiscale convolution based on an attention mechanism:

Circularly inputting the two-dimensional data with the periodic characteristics obtained in the step (2) into a multi-scale convolution module to obtain a plurality of multi-scale characteristics with attention expression; wherein the multi-scale convolution module comprises two multi-scale convolution operations and a convolution attention module which are arranged in parallel, and a GELU activation function is applied after each multi-scale convolution operation is finished;

step (4), feature fusion based on self-adaptive weight:

based on the amplitude mean value of the corresponding period of the frequency, carrying out multi-period feature fusion on a plurality of multi-scale features with attention expression;

Step (5), linear classification:

and (3) sequentially carrying out layer normalization and GELU activation on the fusion result of the multicycle features in the step (4), and finally flattening and inputting the fusion result into a full-connection layer to obtain a classification result.

2. The method according to claim 1, wherein step (2) is specifically:

2-1, carrying out position embedding on the data processed in the step (1);

2-2, carrying out mark embedding on the data processed in the step (1);

2-3, adding the position embedding result obtained in the step 2-1 and the mark embedding result obtained in the step 2-2;

2-4 performing fast Fourier transform on the addition result in the step 2-3, further comparing the amplitude averages of the signals on all channels under different frequencies, selecting the frequencies corresponding to the k highest amplitude averages, forming a frequency list, and simultaneously storing the corresponding channel amplitude averages in a weight matrix W;

2-5, sequentially performing two-dimensional interpretation on the original signals subjected to the feature embedding processing according to the frequency list returned in the step 2-4, and rearranging the original signals by taking T _k＝1/f_k as a period length; if N is not divisible by T _k, the signal is required to be subjected to tail zero padding so that the signal length is changed into N '_k, N' _k is ensured to be divisible by T _k, and k two-dimensional data with periodic characteristics are obtained Wherein each column represents a time point with the same phase among N'/T _k periods divided in step 2-4, each row/>Representing data within a T _k size period.

3. The method of claim 2, wherein the embedding of the positions in step 2-1 uses a sine-cosine position coding method to generate a continuous vector for each time point in the current and voltage data, the vector capturing the relative relationship between the positions;

where pos represents the position of the data, d represents the dimension of the position code, and i represents the index of the dimension.

4. The method of claim 2, wherein the label embedding in step 2-2 is convolved with a cyclic filling scheme to achieve a representation of discrete acquisition values using a set of continuous vectors.

5. The method according to claim 2, wherein in step 2-4 is specifically:

First, the kth element on the complex sequence X _i on the original signal path i is calculated by fast Fourier transform, resulting in complex amplitude X [ k ] _i:

Wherein N represents the length of the signal, k represents the index of the converted frequency, x _i [ N ] represents the nth element of the original signal sequence after feature embedding processing on the ith signal channel, and j represents the imaginary unit;

Next, the amplitude corresponding to the frequency f _k is calculated using the complex amplitude X [ k ] _i And calculates the amplitude mean AMP _k over all channels:

Wherein Re (-) represents the real part and Im (-) represents the imaginary part;

Finally, channel amplitude mean values AMP _k corresponding to different frequencies f _k are compared, frequencies corresponding to the highest k channel amplitude mean values are selected and returned in a list mode, and meanwhile, the corresponding channel amplitude mean values are stored in a weight matrix W= { AMP ₁,AMP₂,...,AMP_k }.

6. The method of claim 5, wherein step (3) is specifically:

3-1, the two-dimensional data with periodic characteristics obtained in the step (2) is subjected to a multi-scale convolution module to obtain a plurality of scale Feature vectors { Feature ₁,Feature₂,...,Feature_n };

The multi-scale convolution module comprises 1X 3 convolution blocks, 3X 1 convolution blocks and 1X 1 convolution blocks which are arranged in parallel; wherein the 1 x 3 convolution blocks, the 3 x 1 convolution blocks differ in expansion rate;

3-2 stacking the plurality of scale Feature vectors { Feature ₁,Feature₂,...,Feature_n } obtained in the step 3-1 to generate a new dimension, then averaging the new dimension to obtain an output result of the multi-scale convolution, and obtaining a final result by using a GELU activation function on the output result:

x _H*W*C＝GELU(Avg(Stack(Feature₁,Feature₂,...,Feature_n)) type (6)

Where Stack (·) represents a stacking operation, avg (·) represents an averaging; H. w, C denotes the length, width and height dimensions of the X _H*W*C feature map;

3-3 performing channel attention and spatial attention calculations on the final result X _H*W*C of step 3-2, respectively;

the method for calculating the channel attention comprises the steps of firstly carrying out average pooling AvgPooling and maximum pooling MaxPooling on a final result X _H*W*C in the step 3-2 along a channel to obtain two channel weights A _c and M _c, then obtaining new attention weights A '_c and M' _c respectively through a shared MLP, and then obtaining a final channel attention ChannelAttention (X _H*W*C) through addition and sigmoid function mapping;

ChannelAttention (X _H*W*C)＝Sigmoid(A′_c+M′_c) A piece of cloth (7)

The calculation method of the spatial attention comprises the steps of carrying out average pooling and maximum pooling on a spatial feature map along a channel through a final result X _H*W*C in the step 3-2, then stacking two feature outputs, and then carrying out point convolution on the stacked data and carrying out sigmoid to obtain the final spatial attention;

3-4 multiplying the final result X _H*W*C obtained in step 3-2 by the channel attention ChannelAttention (X _H*W*C) calculated in step 3-3 to obtain feature X1; then the feature X1 is multiplied by the spatial attention SpatialAttention (X _H*W*C) calculated in step 3-3 to obtain the final attention-based feature output X2;

x1=x _H*W*C×ChannelAttention(X_H*W*C) (8)

X2=x1× SpatialAttention (X _H*W*C) formula (9)

3-5 Taking the attention-based characteristic output X2 of the step 3-4 as the input of the multi-scale convolution module, repeating the operations of the steps 3-1 to 3-2 to obtain the final output

7. The method of claim 6, wherein step (4) is specifically:

4-1 outputting the two-dimensional data corresponding to T _k obtained in step (3) Restoring to one-dimensional data X ^k, then cutting the length of the one-dimensional data X ^k to the window length of the step (1), namely discarding the corresponding data at the time point exceeding the original signal length;

4-2, stacking the one-dimensional data corresponding to each T _k along one dimension, and multiplying the one-dimensional data by the weight matrix W obtained in the step 2-4 to finally form a characteristic output based on multi-period fusion;

4-3 based on the idea of a residual design block, adding the output obtained in the step 4-2 with k two-dimensional data with periodic characteristics obtained in the step 2-5 to obtain a fusion result of the multicycle characteristics.

8. An arc fault identification system implementing the method of any one of claims 1-7, comprising:

The data acquisition module is used for acquiring current and voltage data of the electric arc;

the data preprocessing module is used for carrying out sliding division on current and voltage data of the electric arc in a fixed time window;

the two-dimensional interpretation expansion module is used for performing characteristic embedding on the current and voltage data processed by the data preprocessing module, and then converting the current and voltage data into two-dimensional data under k periods corresponding to different frequencies to obtain two-dimensional data with periodic characteristics;

The feature extraction module circularly inputs the two-dimensional data with periodic features output by the two-dimensional interpretation and expansion module into the multi-scale convolution module to obtain a plurality of multi-scale features with attention expression;

the characteristic fusion module based on the self-adaptive weight fuses a plurality of multi-scale characteristics with attention expression into multi-period characteristics based on the frequency corresponding period;

And the linear classifier sequentially performs layer normalization and GELU activation on the fusion result of the multicycle characteristics, and finally flattens and inputs the fusion result to the full-connection layer to obtain a classification result.

9. A computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of any of claims 1-7.

10. A computing device comprising a memory having executable code stored therein and a processor, which when executing the executable code, implements the method of any of claims 1-7.