An improved method for fault diagnosis of AC equipment based on neural network

Abstract

The present invention discloses an improved method for diagnosing faults of alternating current equipment based on a neural network, comprising: adding a pre-data sampling point rearrangement module based on the electrical characteristics of the input data and replacing the activation function of the original network on the basis of the MobileNetV3‑small deep network. The pre-sampling point rearrangement module places the sampling points every 0.02 seconds in chronological order from left to right and from top to bottom to form a new area based on the periodic electrical characteristics of the input data. The activation function part replaces h‑swish with h‑tanh and is applied to the entire network. The h‑tanh activation function further accelerates the calculation speed and reduces the computing power required. Compared with existing technologies, the present invention has the characteristics of low cost, high speed, deployment in harsh environments, and closer integration with the power system.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and electrical engineering, and in particular relates to an alternating current equipment fault diagnosis method improved based on a neural network. Background Art

[0002] Power system fault diagnosis technology has always been a research focus in the power system field, with diagnostic speed and accuracy being a key research priority. Rapid fault diagnosis can effectively reduce the scale of subsequent faults and significantly minimize losses caused by faults. Furthermore, improving fault diagnosis accuracy can reduce false trips, which in turn can reduce economic losses and the risk of faults. Among current power system fault diagnosis methods, manual inspection still exists, but this method suffers from slowness and subjectivity. The digitalization of power systems requires the complete replacement of manual labor with equipment. Furthermore, a large portion of power system fault diagnosis still relies on relatively backward methods, such as traditional threshold methods. New methods resulting from the recent development of artificial intelligence have yet to be widely tested in power systems.

[0003] In recent years, the field of artificial intelligence has seen the emergence of numerous mature methods for processing unstructured data, such as images and waveforms, many of which even surpass manual accuracy. Convolutional neural networks (CNNs) have demonstrated excellent performance in image recognition and can even be used in other fields, such as sound signals, through appropriate transformations. However, these methods are relatively large and computationally slow. To address the computational slowness of CNNs and the large size and high-performance computing requirements required for deployment, the MobileNet network was developed. Modifying the convolution kernel, convolution method, and activation function of CNNs reduces computational complexity while barely compromising prediction accuracy. MobileNet has been updated to its third version, MobileNetV3. However, MobileNetV3 was originally designed for image recognition and was not tailored to the electrical characteristics of power system electrical equipment. This leaves significant room for improvement in the application of MobileNetV3 in image and signal recognition for power systems.

[0004] To address the problems in the above two research areas, this paper improves the MobileNetV3-small method, further improving the computing speed and reducing the minimum performance requirements for hardware deployed on the network while adapting the electrical characteristics. Summary of the Invention

[0005] The purpose of the present invention is to provide an improved method for diagnosing faults of AC equipment based on a neural network, so as to improve the accuracy of fault diagnosis.

[0006] To achieve the above object, the present invention provides an improved method for diagnosing AC equipment faults based on a neural network, comprising the following steps:

[0007] Build an AC equipment fault diagnosis model based on the MobileNetV3-small deep network;

[0008] Collecting AC device signal data, and arranging the collected AC device signal data in chronological order;

[0009] Inputting the AC equipment signal data into the AC equipment fault diagnosis model for training until the number of iterations reaches a preset number to obtain a trained AC equipment fault diagnosis model;

[0010] Fault diagnosis is performed based on the trained AC equipment fault diagnosis model.

[0011] Optionally, constructing an AC equipment fault diagnosis model includes: using an activation function based on the MobileNetV3-small model, replacing the number and position of activation functions in the bneck module in the MobileNetV3-small model, using the input of the activation function as the output value of the module on the network, and outputting the calculation result of the dependent variable corresponding to the function input value.

[0012] Optionally, the replacement includes full replacement and partial replacement, and both the full replacement and the partial replacement replace the activation function with the activation function h-tanh.

[0013] Optionally, collecting the signal data of the AC device includes: collecting the waveform digital signal of the AC device and the device, and performing binary bit access to the digital signal.

[0014] Optionally, the AC device signal data is collected by first encoding the AC device signal data, then decoding the AC device signal data, dividing the decoded data or the unencoded original data to obtain sampling points, and rearranging the sampling points in chronological order from left to right and from top to bottom.

[0015] Optionally, the input of the activation function is used as the output value of the module on the network, and the output is the calculation result of the dependent variable corresponding to the function input value. The output value is input into the network for two-dimensional convolution, where the output value is 224×224×3, and the two-dimensional convolution includes three convolution kernels, and the convolution kernel is 3×3.

[0016] Optionally, the bneck module includes: in the first layer, convolving the input data with a 1×1 convolution kernel to obtain higher-dimensional data, and using RELU6 for nonlinear calculation, and connecting the input data to the last layer of the bneck module through a residual edge, performing depth-separable convolution on the higher-dimensional data using a 3×3 convolution kernel, performing nonlinear calculation of the RELU6 activation function, inputting the calculation result into a lightweight attention model, adjusting the weight of each channel, and further convolving the result with a 1×1 convolution kernel to obtain an output, which is input into the second and third layers.

[0017] Optionally, the bneck module also includes: the activation function used by bneck in the fifth to seventh layers is h-tanh, the activation function used by bneck in the eighth to twelfth layers is h-swish, in the thirteenth layer, the output data of the twelfth layer is mean pooled using a 7×7 convolution kernel, in the fourteenth layer, the data of the thirteenth layer is convolved using a 1×1 convolution kernel, the activation function is h-swish, and batch normalization is performed, and in the fifteenth layer, the data of the fourteenth layer is further batch normalized to output the data scale.

[0018] Optionally, fault diagnosis based on the trained AC equipment fault diagnosis model includes: uniquely encoding different types of faults and mapping them to the output nodes of the model. During the training process, the node value mapped to the fault type corresponding to the data is 1, and the values ​​of the remaining nodes are 0; during the diagnosis process of the model, the node with the largest value among the output nodes corresponds to the fault.

[0019] Technical effect of the invention: The present invention discloses an improved AC equipment fault diagnosis method based on a neural network. The present invention improves the MobileNetV3-small model, further improves the calculation speed while adapting the electrical characteristics, and reduces the minimum performance requirements for the hardware deployed in the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0021] Figure 1 Schematic diagram of the framework of the improved AC equipment fault diagnosis method based on neural network according to an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of the pre-data sampling point area according to an embodiment of the present invention;

[0023] Figure 3is the function image of the h-tanh activation function according to an embodiment of the present invention;

[0024] Figure 4 This is a model of the bneck module in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0026] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0027] like Figure 1-4 As shown, this embodiment provides an improved method for diagnosing AC equipment faults based on a neural network, comprising the following steps:

[0028] Build an AC equipment fault diagnosis model based on the MobileNetV3-small deep network;

[0029] Collecting AC device signal data, and arranging the collected AC device signal data in chronological order;

[0030] Inputting the AC equipment signal data into the AC equipment fault diagnosis model for training until the number of iterations reaches a preset number to obtain a trained AC equipment fault diagnosis model;

[0031] Fault diagnosis is performed based on the trained AC equipment fault diagnosis model.

[0032] An improved AC equipment fault diagnosis model based on the MobileNetV3-small network is built on the MobileNetV3-small deep network. Based on the electrical characteristics of the input data, a pre-data sampling point rearrangement module is added, and the activation function of the original network is replaced. The pre-sampling point rearrangement module places the sampling points every 0.02s in chronological order from left to right and from top to bottom to form a new area. The activation function replaces h-swish with h-tanh and is applied to the entire network. The number of sampling points in the relevant area is determined by the time span t of the input data, and its size is in Represents logarithmic value Round down. The value of should be a composite number for the convolution kernel to shift.

[0033] The implementation of the h-tanh activation function is based on the h-sigmoid activation function, which is shifted down by 0.5 units on the basis of the h-sigmoid function to solve the problem that the h-sigmoid activation function has no zero and negative values, further accelerating the convergence speed of model training.

[0034] The specific calculation method and steps of the present invention are as follows:

[0035] Step 1: Collect the waveform digital signal of the AC device, as well as the binary bit depth of the digital signal accessed by the device during sampling. If encoding is performed, the encoding method needs to be recorded.

[0036] Step 2: If the digital signal is encoded during acquisition, it is necessary to decode the digital signal. The decoded data or the unencoded original data is decoded according to The segmentation is performed to obtain the segment PART = {PART1, PART2, ..., PARTN}, and the sampling points of the segments in PART are filled into the rearrangement module in the order of 0.02s from left to right and from top to bottom in chronological order. Figure 1 As shown, sampling point 1 is first added to the position of area 1. Sampling point 4 is obtained 0.02 seconds after sampling point 1 and is therefore filled to the right of sampling point 1 in area 1. Sampling point 7 is obtained 0.02 seconds after sampling point 4 and is filled behind sampling point 4 in area 1. Because the first row is already filled, sampling point 7 is filled to the leftmost side of the second row in area 1. And so on. Sampling points that are not 0.02 seconds away from sampling point 1 and sampling points that are 0.02 seconds away from themselves are filled in other areas according to the above method, as shown in Table 1.

[0037] Table 1

[0038]

[0039] Step 3: Perform convolution operations on the rearrangement module in the convolutional neural network. The size of the convolution kernel is the size of region 1, and the depth is the number of binary bits of the digital signal accessed when the device is sampling: depth. Hyperparameters such as the convolution kernel movement step size, pooling method, and number are determined by the user.

[0040] Step 4: Replace the activation function in MobileNetV3. The MobileNetV3 model is divided into two types: MobileNetV3-large and MobileNetV3-small. Different activation functions are used in different parts of the model. The present invention replaces the h-swish activation function in the bneck part of the model. Because there are many bnecks in the MobileNetV3 model, there are two main replacement methods: 1) Replace all h-swish with h-tanh 2) Partially replace h-swish with h-tanh. The specific number and position of the replacement activation function are determined by the user. The input of the h-tanh activation function is the output value of the previous module on the network, and the output is the calculation result of the dependent variable corresponding to the function input value.

[0041] Step 5: Model calculation steps are as follows Figure 2 As shown in the figure, in this scenario, after rearranging the input data and convolving three convolution kernels, an output of 224×224×3 is obtained. The output is input to the network and a two-dimensional convolution (Conv2d) convolution kernel size of 3×3 is performed. The attention module (SE) is not used and the activation function is h-swish (HS). After the above operations, a data size of 112×112×16 is generated and input to the second layer network. The second layer network inputs the data to the bneck module. The model of the bneck module is shown as follows. Figure 4 As shown, bneck first convolves the input data with a 1×1 convolution kernel to obtain higher-dimensional data, and uses RELU6 (RE) for nonlinear calculations. At the same time, the input data is connected to the last layer of the bneck module through the residual edge, and the higher-dimensional data is subjected to depthwise separable convolution (Dwise) using a 3×3 convolution kernel, and the nonlinear calculation of the RELU6 activation function is still performed. The calculation result is input into the lightweight attention model (SE), and the weight of each channel is adjusted. The result is further convolved with a 1×1 convolution kernel to obtain the output. After passing through the bneck module, 56×56×12-scale data is obtained and input into the third layer. The third-layer data is input into the bneck module of the third layer for calculation. The activation function in bneck still uses RELU6, but the attention model is not used. Similarly, in the fifth to seventh layers, the activation function used by bneck is h-tanh, and the activation function used by bneck in the eighth to twelfth layers is h-swish. The use of the attention module is as shown in the figure. Figure 1As shown in the figure, in the 13th layer, the output data of the 12th layer is mean-pooled using a 7×7 convolution kernel. In the 14th layer, the data of the 13th layer is convolved using a 1×1 convolution kernel with the h-swish activation function, and batch normalization (NBN) is performed. In the 15th layer, the data of the 14th layer is further batch-normalized. The output data scale is determined by the number of fault diagnosis categories. By adjusting the hyperparameters of the convolution kernel in the network, such as the stride and size, the output scale is achieved.

[0042] Step 6: One-hot encode different types of faults and map them to the model's output nodes. During training, the node corresponding to the fault type mapped to the data is assigned a value of 1, while the remaining nodes are assigned a value of 0. During the model's diagnosis process, the fault corresponding to the node with the largest value among the output nodes is the diagnosis result.

[0043] The framework diagram of the AC equipment fault diagnosis model improved based on the MobileNetV3-small network is as follows Figure 1 As shown, based on the traditional MobileNetV3-small, a front-end data sampling point rearrangement module is added, and the rearrangement method is as follows Figure 2 The front sampling point rearrangement module places the sampling points every 0.02s in chronological order from left to right and from top to bottom to form a new area based on the periodic electrical characteristics of the input data. The activation function of the original network is replaced by the new activation function. Figure 3 , based on the existing h-sigmoid activation function, the function is shifted down by 0.5. The model can be used for fault diagnosis of AC equipment in power systems, and is limited to AC equipment with a power frequency of 50Hz. The model needs to be trained using a supervised method. There is no fixed requirement for the size of the training dataset, and 300+ is recommended. The model can be deployed on devices such as Raspberry Pi 4b and STC32 microcontrollers. The model is developed and trained using Pytorch, and deployed using C language. Model training is accelerated using RTX2060, with an Intel(R) Core(TM) i7-5500U CPU@2.40GHz(2) processor and 16GB memory.

[0044] Based on the characteristics of electrical equipment in the power system, the present invention improves the MobileNetV3-small model to further improve the fault diagnosis accuracy while further increasing the diagnosis speed.

[0045] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An improved AC equipment fault diagnosis method based on neural network, characterized in that: The following steps are involved: Build an AC equipment fault diagnosis model based on the MobileNetV3-small deep network; Constructing an AC equipment fault diagnosis model includes: using an activation function based on the MobileNetV3-small model, replacing the number and position of activation functions in the bneck module in the MobileNetV3-small model, using the input of the activation function as the output value of the upper module in the network, and outputting the dependent variable calculation result corresponding to the function input value; The bneck module includes: in the first layer, convolving the input data with a 1×1 convolution kernel to obtain higher-dimensional data, and performing nonlinear calculation using RELU6. At the same time, the input data is connected to the last layer of the bneck module through a residual edge, and the higher-dimensional data is subjected to depthwise separable convolution using a 3×3 convolution kernel, and nonlinear calculation of the RELU6 activation function is performed. The calculation result is input into a lightweight attention model, and the weight of each channel is adjusted. The result is further convolved with a 1×1 convolution kernel to obtain an output, which is input into the second and third layers. The bneck module further includes: in the fifth to seventh layers, the activation function used by bneck is h-tanh, the activation function used by bneck in the eighth to twelfth layers is h-swish, in the thirteenth layer, the output data of the twelfth layer is mean pooled using a 7×7 convolution kernel, in the fourteenth layer, the data of the thirteenth layer is convolved using a 1×1 convolution kernel, the activation function is h-swish, and batch normalization is performed, and in the fifteenth layer, the data of the fourteenth layer is further batch normalized to output data scale; Collecting AC device signal data, and arranging the collected AC device signal data in chronological order; Inputting the AC equipment signal data into the AC equipment fault diagnosis model for training until the number of iterations reaches a preset number to obtain a trained AC equipment fault diagnosis model; Fault diagnosis is performed based on the trained AC equipment fault diagnosis model.

2. The improved AC equipment fault diagnosis method based on neural network according to claim 1, characterized in that: The replacement includes full replacement and partial replacement, and both the full replacement and the partial replacement replace the activation function with the activation function h-tanh.

3. The improved AC equipment fault diagnosis method based on neural network according to claim 1, characterized in that: Collecting the signal data of the AC equipment includes: collecting the waveform digital signal of the AC equipment and the equipment, and accessing the digital signal in binary bits.

4. The improved AC equipment fault diagnosis method based on neural network according to claim 3, characterized in that: To collect AC device signal data, the AC device signal data is first encoded, and then the AC device signal data is decoded. The decoded data or the unencoded original data is divided to obtain sampling points, and the sampling points are rearranged in chronological order from left to right and from top to bottom.

5. The improved AC equipment fault diagnosis method based on neural network according to claim 4, characterized in that: The input of the activation function is used as the output value of the previous module on the network, and the output is the calculation result of the dependent variable corresponding to the function input value. The output value is input into the network for two-dimensional convolution, where the output value is 224×224×3. The two-dimensional convolution includes three convolution kernels, and the convolution kernel is 3×3.

6. The improved AC equipment fault diagnosis method based on neural network according to claim 1, characterized in that: Fault diagnosis based on the trained AC equipment fault diagnosis model includes: uniquely encoding different types of faults and mapping them to the output nodes of the model. During the training process, the node value of the fault type mapped to the data is 1, and the values ​​of the remaining nodes are 0; during the diagnosis process of the model, the node with the largest value among the output nodes corresponds to the fault.

Images