Fault arc detection method based on CNN and channel attention mechanism

Abstract

The invention discloses a fault arc detection method based on a CNN and a channel attention mechanism, and the method comprises the steps: inputting an original current signal into a deep learning network, enabling the network to autonomously mine features implied behind current signal data, and achieving the recognition of a series fault arc. According to the method, on the basis of a channel attention mechanism, under the condition that local spatial data are ignored, global average pooling is carried out on channel information of the feature map, the importance degree of each channel of the feature map is obtained, and the weight is assigned, so that the neural network pays attention to the feature channel with the high weight, and the fault arc detection and recognition rate is improved. In order to solve the problem of over-fitting of the deep learning network, batch standardization Batchnorm and dropout are introduced, so that the neural network is trained more quickly and stably, and the over-fitting of the model is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of arc fault detection in power systems, and in particular relates to an arc fault detection method based on CNN and a channel attention mechanism. Background Art

[0002] According to the fire statistics published by the National Fire and Rescue Bureau, the number of electrical fires is still high, accounting for more than one-third of the number of various fires. More than 40% of electrical fires are caused by fault arcs, which have become the main cause of electrical fire accidents in low-voltage distribution systems. In actual low-voltage distribution networks, parallel fault arcs and ground fault arcs are usually caused by short circuits between power lines, and they have large currents. Therefore, traditional short-circuit or overcurrent protection systems can detect and make protective actions more effectively. The current size of the series fault arc is limited by the load in the circuit, and the current of the series fault arc is generally the same as the normal working current, or even slightly smaller. When a series fault arc occurs, it is equivalent to an impedance, which makes the actual current in the circuit loop slightly smaller than the normal current, and the circuit breaker cannot operate. Therefore, traditional short-circuit, overcurrent, overload, and leakage protection technologies cannot effectively detect series fault arcs in the circuit. The occurrence of series fault arcs is concealed and sudden, but once it occurs, it is easy to cause disasters. Research on low-voltage system series fault arc detection technology is of great significance to improving power safety and reducing damage to people's lives and property. At present, the fault arc detection methods are mainly based on the physical characteristics and time-frequency characteristics of the arc. Among them, the fault arc detection algorithm based on physical quantities, such as arc sound, arc light, arc voltage and arc temperature, has great limitations due to the unknown location of the fault arc. It is often used to detect the fault arc of specific equipment such as switch cabinets and distribution cabinets. The fault arc detection algorithm based on the time-frequency characteristics of the arc current uses tools such as Fourier transform and wavelet transform to extract the time domain, frequency domain and time-frequency domain characteristics of the arc current, and detects the fault arc by setting empirical thresholds or classification algorithms. They require manual extraction of some features and are effective in some applications, but the diversity of loads makes them lack reliability and generalization ability. Summary of the invention

[0003] In order to solve the problems of difficult threshold setting, high computational complexity, poor generalization ability, etc. in traditional detection methods, the purpose of the present invention is to provide a fault arc detection method based on CNN and channel attention mechanism.

[0004] The object of the present invention is to achieve the following technical solution: a fault arc detection method based on CNN and channel attention mechanism, comprising the following steps:

[0005] Collecting current data of various loads in arc state and normal state, and normalizing the current data to enhance the stability of the neural network and the accuracy of the model;

[0006] Automatically establish a fault arc database using timestamps provided by the acoustic sensor and the optical sensor, and perform one-hot encoding on the database;

[0007] A network model based on CNN and channel attention mechanism is constructed; the model consists of 5 convolutional layers, 2 pooling layers, 2 channel attention mechanism layers and 1 fully connected layer; the convolutional layer uses the Relu activation function, a normalized Batchnorm is added after each convolutional layer, the pooling layer selects average pooling, the fully connected layer uses the Softmax activation function, and a dropout operation is introduced before the fully connected layer;

[0008] Training and verifying the network model according to the data in the database, adjusting the network hyperparameters, and saving the network model with the highest accuracy during the verification process;

[0009] The network model is tested according to the data in the database to obtain a fault arc detection result.

[0010] Furthermore, during the normalization process, a maximum-minimum normalization method is used to normalize the current data to the (-1, 1) interval.

[0011] Furthermore, the fault arc database is randomly divided, wherein 70% of the data is used for training, 15% of the data is used for verification, and 15% of the data is used for testing.

[0012] Furthermore, the structure of the network model includes in sequence: convolution layer 1: convolution kernel size is 8, step size is 2, and the number of convolution kernels is 16; convolution layer 2: convolution kernel size is 8, step size is 2, and the number of convolution kernels is 16; pooling layer 1: pooling kernel size is 2, step size is 2; convolution layer 3: convolution kernel size is 4, step size is 2, and the number of convolution kernels is 64; channel attention mechanism layer 1; convolution layer 4: convolution kernel size is 4, step size is 2, and the number of convolution kernels is 256; pooling layer 2: pooling kernel size is 2, step size is 2; convolution layer 5: convolution kernel size is 2, step size is 1, and the number of convolution kernels is 512; channel attention mechanism layer 2; and flatten layer and fully connected layer.

[0013] Furthermore, when constructing a network model based on CNN and channel attention mechanism, a loss function is used to quantify the error between the predicted value and the true value. The loss function of the network model is a cross entropy loss function, specifically:

[0014]

[0015] Among them, y i Indicates the true labels of normal and fault arcs, Represents the predicted label, and N is the total number of samples.

[0016] Furthermore, when training the model, the Adam optimizer is selected to adjust the model parameters, wherein the exponential decay rate β1 of the first-order moment estimate is 0.9, the exponential decay rate β2 of the second-order moment estimate is 0.999, and the learning rate α is 0.002.

[0017] The beneficial effect of the present invention is that: the present invention proposes a fault arc detection method based on CNN and channel attention mechanism. A large number of data sets are automatically constructed using sound wave and light sensors and one-hot encoded. Based on the channel attention mechanism, the neural network can select more critical channels to identify cross-channel fault arcs, reducing the redundancy of the model.

[0018] The present invention directly uses the original data of the current, does not consider the changes in the current characteristics before and after the fault occurs, and does not need to set a threshold. The various current data collected are directly input into the network model based on CNN and channel attention mechanism for fault data detection. The network is used to autonomously mine the features hidden in the current signal to realize the detection of series fault arcs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

[0020] Figure 2 It is a data collection schematic diagram of the method of the present invention;

[0021] Figure 3 A network structure diagram of the method of the present invention;

[0022] Figure 4 This is a diagram of the channel attention mechanism of the present invention. DETAILED DESCRIPTION

[0023] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0024] Reference Figure 1 As shown, an arc fault detection method based on CNN and channel attention mechanism provided by an embodiment of the present invention comprises the following steps:

[0025] Step 1. Collect current data of various loads in arc state and normal state, and normalize these data; specifically including:

[0026] Normalizing the data samples can effectively improve the accuracy of prediction. The maximum and minimum normalization method adopted in the present invention normalizes the sample data to the interval (-1,1), and the minimum and maximum normalization calculation formulas are:

[0027]

[0028] Among them, x represents sample data, x * is the normalized sample data, x max and x min are the maximum and minimum values ​​in the sample sequence, respectively, max and min are the right and left endpoints of the normalized interval respectively.

[0029] Step 2. Since actual fault arcs are always random and difficult to capture or reproduce, a specific test platform must be used to generate arcs with different types of actual loads. According to the IEC62606 standard, the present invention designs a series fault arc experimental platform. A large number of fault arc data sets are automatically constructed using acoustic and optical sensors, and the fault arc data are automatically labeled with the help of computer programs and timestamps provided by acoustic and optical sensors. The data set is randomly divided, 70% of the data is used for training, 15% of the data is used for verification, and 15% of the data is used for testing. The data acquisition experimental platform is as follows: Figure 2 shown.

[0030] Step 3. Build a network model based on CNN and channel attention mechanism; specifically include:

[0031] The network structure of the network model based on CNN and channel attention mechanism is as follows Figure 3 As shown in Figure 2, the model consists of five convolutional layers, two pooling layers, two channel attention mechanism layers, and one fully connected layer.

[0032] The original current data (5000, 1) collected in step 1 is passed through the first convolution layer, with a convolution kernel size of 8, a step size of 2, and a number of convolution kernels of 16, to obtain a matrix of size (2497, 16). After the second convolution layer, the convolution kernel size is 8, the step size is 2, and the number of convolution kernels is 16, to obtain a matrix of size (1246, 16). The matrix of size (1246, 16) is average pooled, with a pooling kernel size of 2 and a step size of 2, to obtain a matrix of size (623, 16). After the third convolution layer, the convolution kernel size is 4, the step size is 2, and the number of convolution kernels is 64, to obtain a matrix of size (311, 64). Then, the channel attention mechanism layer is used to perform global average pooling on the information of each channel of the feature map, obtain the importance of each channel of the feature map, and assign weights. After the fourth convolution layer, the convolution kernel size is 4, the stride is 2, and the number of convolution kernels is 256, resulting in a matrix of size (155, 256). The matrix of size (155, 256) is average pooled, with a pooling kernel size of 2 and a stride of 2, resulting in a matrix of size (78, 256). After the fifth convolution layer, the convolution kernel size is 2, the stride is 1, and the number of convolution kernels is 512, resulting in a matrix of size (79, 512). Then, the channel attention mechanism layer is used to re-weight each channel of the feature map, so that the neural network pays attention to more important channel features. Then, through the Flatten layer, the input is "flattened", that is, the multi-dimensional input is one-dimensionalized, that is, the (79, 512) matrix is ​​processed by this layer and the output is a one-dimensional data of 79*512. Then, through a fully connected layer, dropout is added before the full connection to reduce the overfitting of the model, and finally the softmax function is used to output the category.

[0033] Step 4. After each convolutional layer, a batch normalization operation Batchnorm1d is performed to normalize the feature map, thereby accelerating network learning. The batch normalization process is: calculate the mean and variance of the elements in a minibatch, then subtract the mean from x and divide it by the standard deviation, and finally use γ and β to perform an affine transformation to obtain the final BN output. The specific process is as follows:

[0034]

[0035] Among them, x i represents the sample data after convolution, m represents the batch size, μ represents the mean of the small batch sample data, σ represents the variance of the small batch sample data, Represents the data after standardization, y i Express The final output after translation and scaling, ε is a constant, usually set to 1e-5 or 1e-6 to ensure numerical stability, γ and β are learnable hyperparameters.

[0036] Among them, the first formula is for finding the mean, the second formula is for finding the variance, the third formula is for normalization, and the fourth formula is for affine transformation.

[0037] The channel attention mechanism can adaptively learn the importance of different channels, thereby enhancing the ability to characterize features. By dynamically adjusting the channel weights, the network's attention to key features can be increased, thereby improving the detection and recognition rate of fault arcs. Figure 4 As shown in the figure: First, the input data is globally averaged pooled in the channel direction, and then the correlation between channels is analyzed through two full connections to obtain the importance of different channels. Finally, the obtained channel attention score is multiplied by the original input through the broadcast mechanism to re-weight different channels.

[0038] Step 5. Use a large number of self-built data sets to train and verify the network model, adjust the network hyperparameters, select the network model with the highest accuracy during the training process, test the network model based on the data in the data set, and obtain the fault arc detection results.

[0039] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A fault arc detection method based on CNN and channel attention mechanism, characterized in that: The steps include: Collecting current data of various loads in arc state and normal state, and normalizing the current data to enhance the stability of the neural network and the accuracy of the model; Automatically establish a fault arc database using timestamps provided by the acoustic sensor and the optical sensor, and perform one-hot encoding on the database; A network model based on CNN and channel attention mechanism is constructed; the network model consists of 5 convolutional layers, 2 pooling layers, 2 channel attention mechanism layers and 1 fully connected layer; the convolutional layer uses the Relu activation function, a normalized Batchnorm is added after each convolutional layer, the pooling layer selects average pooling, the fully connected layer uses the Softmax activation function, and a dropout operation is introduced before the fully connected layer; Training and verifying the network model according to the data in the database, adjusting the network hyperparameters, and saving the network model with the highest accuracy during the verification process; The network model is tested according to the data in the database to obtain a fault arc detection result.

2. The arc fault detection method based on CNN and channel attention mechanism according to claim 1, characterized in that: During the normalization process, the maximum and minimum normalization method is used to normalize the current data to the (-1, 1) interval.

3. The arc fault detection method based on CNN and channel attention mechanism according to claim 1, characterized in that: The fault arc database is randomly divided, wherein 70% of the data is used for training, 15% of the data is used for verification, and 15% of the data is used for testing.

4. The arc fault detection method based on CNN and channel attention mechanism according to claim 1, characterized in that: The structure of the network model includes: convolution layer 1: convolution kernel size is 8, step size is 2, and the number of convolution kernels is 16; convolution layer 2: convolution kernel size is 8, step size is 2, and the number of convolution kernels is 16; pooling layer 1: pooling kernel size is 2, step size is 2; convolution layer 3: convolution kernel size is 4, step size is 2, and the number of convolution kernels is 64; channel attention mechanism layer 1; convolution layer 4: convolution kernel size is 4, step size is 2, and the number of convolution kernels is 256; pooling layer 2: pooling kernel size is 2, step size is 2; convolution layer 5: convolution kernel size is 2, step size is 1, and the number of convolution kernels is 512; channel attention mechanism layer 2; and flatten layer and fully connected layer.

5. The arc fault detection method based on CNN and channel attention mechanism according to claim 1, characterized in that: When constructing a network model based on CNN and channel attention mechanism, a loss function is used to quantify the error between the predicted value and the true value. The loss function of the network model is a cross entropy loss function, specifically: Among them, y i Indicates the true labels of normal and fault arcs, Represents the predicted label, and N is the total number of samples.

6. The arc fault detection method based on CNN and channel attention mechanism according to claim 1, characterized in that: When training the network model, the Adam optimizer is used to adjust the model parameters, wherein the exponential decay rate β1 of the first-order moment estimate is 0.9, the exponential decay rate β2 of the second-order moment estimate is 0.999, and the learning rate α is 0.002.

Images