Ammeter abnormity identification method and system based on MobileNetV3

Abstract

The invention discloses a MobileNetV3-based ammeter abnormity identification method and system, and the method comprises the steps: obtaining image data of an abnormal ammeter and a normal ammeter, carrying out the preprocessing of the image data, obtaining a sample data set, and randomly dividing the sample data set into a training set and a verification set according to a proportion; training the improved MobileNetV3 network based on the training set, verifying the improved MobileNetV3 network based on the verification set, and taking the trained improved MobileNetV3 network as an abnormal electric energy meter identification model; the improved MobileNetV3 network is a MobileNetV3 network of a bottleneck module with a reverse residual error; and inputting the image data of the to-be-identified electric energy meter into the abnormal electric energy meter identification model, and outputting an electric energy meter abnormity identification result. According to the method, the capability of extracting the appearance characteristics of the electric energy meters is improved, the anomaly recognition accuracy is improved, and exclusive anomaly recognition models of different types of electric energy meters can be generated in a self-adaptive manner.

Description

Technical Field

[0001] The present invention relates to the field of fault identification of electric energy metering devices, and in particular to an electric meter anomaly identification method and system based on MobileNetV3. Background Art

[0002] At present, the abnormal recognition of smart meters mainly relies on traditional image processing methods and relatively basic deep learning networks. In the detection of smart meters, traditional methods use edge detection, color feature extraction, and other methods combined with feature matching methods (such as SIFT or SURF) to detect the appearance of the meter. However, since these methods are easily affected by external conditions such as lighting and angle, their recognition effect is not stable, and feature extraction takes a long time, which is difficult to meet the high efficiency requirements of the actual pipeline.

[0003] In addition, although some existing deep learning-based models (such as VGG16, ResNet50, etc.) perform well in image classification, their structures are relatively complex, the number of model parameters is large, and they are difficult to deploy on devices with limited computing power. Especially in the anomaly detection scenario, the feature extraction effect of these networks is limited, resulting in low accuracy in identifying abnormal appearance of electric meters, and slow detection speed, which is not conducive to the real-time requirements in practical applications. Therefore, the existing technical solutions still have deficiencies in terms of model lightweight and recognition accuracy.

[0004] In view of this, this application is hereby filed. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the accuracy of identifying the abnormal appearance of electric meters in the prior art is not high, and the detection speed is slow, which is not conducive to the real-time requirements in practical applications and cannot adapt to the abnormal identification of electric energy meters of different models. The purpose of the present invention is to provide an electric meter abnormality identification method and system based on MobileNetV3. The present invention realizes the automatic construction of the appearance abnormality identification model of the electric energy meter by improving the MobileNetV3 network, and can generate exclusive identification models to adapt to electric energy meters of different models, thereby improving the generalization ability and adaptability of the model. The present invention improves the ability to extract the appearance features of the electric energy meter, significantly improves the accuracy of abnormality identification, and can adaptively generate exclusive abnormality identification models for electric energy meters of different models, effectively responding to the identification needs in different environments, and has a fast detection speed.

[0006] The present invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides an electric meter anomaly recognition method based on MobileNetV3, the method comprising:

[0008] Obtain image data of abnormal electric energy meters and normal electric energy meters, preprocess the image data to obtain a sample data set, and randomly divide it into a training set and a validation set according to a proportion;

[0009] The improved MobileNetV3 network is trained based on the training set, and the accuracy of the judgment of the improved MobileNetV3 network is verified based on the validation set, and the trained improved MobileNetV3 network is used as the abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module of reverse residual;

[0010] The image data of the electric energy meter to be identified is input into the abnormal electric energy meter identification model, and the abnormal electric energy meter identification result is output.

[0011] The present invention is suitable for identifying abnormalities in the appearance of electric energy meters under automated production conditions of centralized detection and verification of electric energy meters, and can adaptively generate exclusive abnormality recognition models for electric energy meters of different models, which can effectively improve the accuracy of the model while taking into account the recognition rate under assembly line conditions.

[0012] Furthermore, the image data is preprocessed to obtain a sample data set, including:

[0013] The photographed smart energy meters are classified into two categories through manual screening: the first category is normal meters, and the second category is abnormal meters. After the classification is completed, a classified sample data set will be obtained;

[0014] The classification method is to save the normal meter pictures in one folder and the abnormal meter pictures in another folder.

[0015] Furthermore, the improved MobileNetV3 network includes an input layer, a convolutional layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer, and an output layer connected in sequence;

[0016] The bottleneck module with reverse residual reduces the amount of calculation while maintaining performance through the expansion, deep convolution, and compression structure to obtain the data after passing through the bottleneck module with reverse residual.

[0017] Among them, the entire bottleneck module with reverse residual can be expressed as:

[0018] X3=Bneck(X2)

[0019] Where X3 represents the data after passing through the bottleneck module with reverse residual, and Bneck() represents the combination of the triple operations of point-by-point convolution expansion → depth-wise convolution → point-by-point convolution compression.

[0020] Furthermore, the bottleneck module with reverse residual includes a point-wise convolution expansion unit, a depth-wise convolution unit, and a point-wise convolution compression unit;

[0021] a point-by-point convolution expansion unit, configured to expand the data with a lower number of channels through a first point-by-point convolution expansion operation to obtain data with a higher number of channels as expanded data;

[0022] A depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain the depth convolved data;

[0023] The point-by-point convolution compression unit is used to compress the data after the depth convolution back to data with a lower number of channels through a second point-by-point convolution.

[0024] Furthermore, the expression of the first point-by-point convolution is: expand =H_swish(Conv 1×1 (X)), where X expand It is the data expanded by point-by-point convolution. 1×1 () indicates that a convolution operation is performed using a 1×1 convolution kernel, and X is the independent variable input into the operation;

[0025] The expression of the second point-by-point convolution is: compress =Conv 1×1 (X depthwise ), where X compress Represents data after point-by-point convolution compression.

[0026] Furthermore, the operational expression of the depth convolution is:

[0027] X depthwise =H_swish(DepthwiseConv 3×3 (X expand

[0028] Where X depthwise DepthwiseConv is the data after deep convolution. 3×3 () indicates that the 3×3 convolution kernel performs a depthwise convolution operation, which is expressed as:

[0029]

[0030] where Y h,w,c Represents the value of the data at position (h,w), channel c after depth convolution. h+i-1,w+j-1,c Indicates that the data that needs to be convolved is at position (h+i-1,w+j-1), the value of channel c; K i,j,c represents the weight of the 3×3 convolution kernel on the cth channel (learnable); b cRepresents the bias (learnable) of the cth channel.

[0031] Furthermore, improve the loss function of the MobileNetV3 network for:

[0032]

[0033] In the formula, i is the sample, y is i is the true label, is the predicted label.

[0034] In the second aspect, the present invention further provides an electric meter anomaly recognition system based on MobileNetV3, the system comprising:

[0035] An acquisition unit, used for acquiring image data of abnormal electric energy meters and normal electric energy meters;

[0036] A preprocessing and division unit is used to preprocess the image data to obtain a sample data set, and randomly divide it into a training set and a validation set according to a proportion;

[0037] The model building and training unit is used to train the improved MobileNetV3 network based on the training set, verify the accuracy of the improved MobileNetV3 network judgment based on the verification set, and use the trained improved MobileNetV3 network as the abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module of reverse residual;

[0038] The abnormality recognition unit is used to input the image data of the electric energy meter to be recognized into the abnormal electric energy meter recognition model and output the abnormality recognition result of the electric energy meter.

[0039] Furthermore, the improved MobileNetV3 network includes an input layer, a convolutional layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer, and an output layer connected in sequence;

[0040] The bottleneck module with reverse residual reduces the amount of calculation while maintaining performance through the expansion, deep convolution, and compression structure to obtain the data after passing through the bottleneck module with reverse residual;

[0041] The bottleneck module with reverse residual includes point-wise convolution expansion unit, depth-wise convolution unit and point-wise convolution compression unit;

[0042] a point-by-point convolution expansion unit, configured to expand the data with a lower number of channels through a first point-by-point convolution expansion operation to obtain data with a higher number of channels as expanded data;

[0043] A depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain the depth convolved data;

[0044] The point-by-point convolution compression unit is used to compress the data after the depth convolution back to data with a lower number of channels through a second point-by-point convolution.

[0045] In a third aspect, the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned electric meter anomaly identification method based on MobileNetV3 is implemented.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] 1. The present invention is based on the electric meter anomaly recognition method and system of MobileNetV3. The present invention is suitable for the recognition of electric meter appearance anomalies under the conditions of centralized detection and calibration of electric energy meters and automated production. It can adaptively generate exclusive anomaly recognition models for electric energy meters of different models, which can effectively improve the accuracy of the model while taking into account the recognition rate under assembly line conditions.

[0048] 2. Lightweight and efficient: Compared with traditional deep learning models (such as ResNet50, VGG16, etc.), this invention adopts an improved MobileNetV3 network to achieve lightweight model design and reduce the amount of calculation, so that this proposal can achieve efficient operation in a low-power environment and meet the real-time detection requirements of pipeline conditions.

[0049] 3. Higher recognition accuracy: Traditional methods are easily affected by the external environment and the recognition effect is unstable. The proposal of this invention improves the ability to extract the appearance features of the meter through the combination of deep convolution and reverse residual bottleneck modules, significantly improves the accuracy of abnormal recognition, and effectively responds to the recognition needs in different environments.

[0050] 4. Good adaptability: This proposal can perform adaptive modeling for different types of meters and generate exclusive abnormality recognition models. Compared with traditional methods, the solution of the present invention has stronger model generalization ability and can adaptively adjust the model structure according to the appearance characteristics of different meters, ensuring high recognition accuracy under various meter models.

[0051] 5. Deployment convenience: The MobileNetV3-based recognition method of the present invention has good deployment advantages. The traditional model has a large number of parameters and is not suitable for application on embedded devices. However, this proposal can run efficiently on edge devices, which is convenient for wide application and promotion in actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0053] Figure 1 This is a flow chart of the electric meter anomaly recognition method based on MobileNetV3 of the present invention;

[0054] Figure 2 This is a schematic diagram of the improved MobileNetV3 network structure of the present invention;

[0055] Figure 3 This is a comparison diagram of the loss between the present invention and other existing methods;

[0056] Figure 4 This is a comparison diagram of the accuracy of the present invention and other existing methods;

[0057] Figure 5 This is a structural block diagram of the electric meter anomaly recognition system based on MobileNetV3 of the present invention. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, the present invention is based on the electric meter abnormality recognition method of MobileNetV3, the method comprising:

[0061] Step 1, obtaining image data of abnormal electric energy meters and normal electric energy meters, preprocessing the image data to obtain a sample data set, and randomly dividing it into a training set and a validation set according to a proportion;

[0062] In this embodiment, the image data is preprocessed to obtain a sample data set, including:

[0063] The photographed smart energy meters are classified into two categories by manual screening: the first category is normal meters, and the second category is abnormal meters. After the classification is completed, a classified sample data set D is obtained;

[0064] The classification method is to save the normal meter pictures in one folder and the abnormal meter pictures in another folder.

[0065] Step 2: train the improved MobileNetV3 network based on the training set, verify the accuracy of the judgment of the improved MobileNetV3 network based on the validation set, and use the trained improved MobileNetV3 network as the abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module of reverse residual;

[0066] In this embodiment, the classified images are used as input and input into the improved MobileNetV3 neural network. The improved MobileNetV3 neural network structure is as follows: Figure 2 As shown. The input is image X, X is a labeled image in the dataset D, Where H is the height of the image, W is the width of the image, and C is the number of channels of the image. Since all images are RGB images, C = 3. The optimization method used in this network is Adam optimization.

[0067] like Figure 2 As shown in the figure, the improved MobileNetV3 network includes an input layer, a convolutional layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer, and an output layer connected in sequence; the main layer structure is described below:

[0068] (1) Convolutional layer

[0069] First, set the convolution kernel: Where 3 is the height and width of the convolution kernel, 3 is the number of input channels, 16 is the number of output channels, and the stride of the convolution is 1. The operation can be expressed as:

[0070] X1=Conv 3×3 (X,K1)=X*K1+b

[0071] Where X1 is the image after convolution, Conv 3×3 () is a convolution operation, and b is a trainable offset term.

[0072] (2) Activation layer

[0073] The activation layer uses the H_swish activation function to replace the ReLU activation function. Compared with the traditional ReLU activation, H_swish retains negative information and has higher efficiency in lightweight networks. This operation can be expressed as:

[0074]

[0075] Where X2 is the activated data, ReLU6 is the function, and the expression is:

[0076] ReLU6(x)=min(max(0,x),6)

[0077] (3) Bottleneck module with reverse residual

[0078] The bottleneck module with reverse residual is a key structure used in the improved MobileNetV3 network. It enhances the network's feature extraction capabilities while maintaining a lightweight design. The core idea of ​​this module is to reduce the amount of calculation while maintaining performance through an expansion, deep convolution, and compression structure to obtain data after passing through the bottleneck module with reverse residual.

[0079] Specifically, the bottleneck module with reverse residual includes a point-wise convolution expansion unit, a depth-wise convolution unit, and a point-wise convolution compression unit;

[0080] 1) A point-by-point convolution expansion unit is used to expand the data with a lower number of channels through a first point-by-point convolution expansion operation to obtain data with a higher number of channels as the expanded data; so that the data has a higher number of channels to improve the representation capability of the features in the data; the operation expression of the first point-by-point convolution is:

[0081] X expand =H_swish(Conv 1×1 (X)

[0082] Where X expand It is the data expanded by point-by-point convolution. 1×1 () indicates that a convolution operation is performed using a 1×1 convolution kernel, and X is the independent variable input into the operation;

[0083] 2) The depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain the data after depth convolution; the computational complexity of depth convolution is low, and it can extract spatial features while maintaining computational efficiency. The operational expression of depth convolution is:

[0084] X depthwise =H_swish(DepthwiseConv 3×3 (X expand

[0085] Where X depthwise DepthwiseConv is the data after deep convolution. 3×3 () indicates that the 3×3 convolution kernel performs a depthwise convolution operation, which is expressed as:

[0086]

[0087] where Y h,w,c Represents the value of the data at position (h,w), channel c after depth convolution. h+i-1,w+j-1,cIndicates that the data that needs to be convolved is at position (h+i-1,w+j-1), the value of channel c; K i,j,c represents the weight of the 3×3 convolution kernel on the cth channel (learnable); b c Represents the bias (learnable) of the cth channel.

[0088] 3) Point-by-point convolution compression unit, used to compress the data after depth convolution back to data with a lower number of channels through the second point-by-point convolution. This operation speeds up subsequent calculations while ensuring that the feature information is not overly distorted during the compression process. The expression of the second point-by-point convolution is:

[0089] X compress =Conv 1×1 (X depthwise )

[0090] Where X compress Represents data after point-by-point convolution compression.

[0091] After the above triple operations of point-by-point convolution expansion → depth-wise convolution → point-by-point convolution compression, the entire bottleneck module with reverse residual can be expressed as:

[0092] X3=Bneck(X2)

[0093] Where X3 represents the data after passing through the bottleneck module with reverse residual, and Bneck() represents the combination of the triple operations of point-by-point convolution expansion → depth-wise convolution → point-by-point convolution compression.

[0094] (4) Global pooling layer

[0095] Global average pooling compresses the number of features in all spatial dimensions into the average value in the channel dimension. This operation can be expressed as:

[0096]

[0097] Where X pool represents the data after pooling, c represents the number of channels, here it is the number of channels of data X3.

[0098] (5) Fully connected layer and judgment layer

[0099] The output after global average pooling is mapped to two categories through a fully connected layer. Let the weight matrix of the fully connected layer be With bias b, the output is:

[0100]

[0101] Softmax is used to normalize the output into a probability distribution of two classes.

[0102] Specifically, the cross entropy loss function is used to optimize the model. For each sample i, the true label is y i , the predicted label is The loss function is: Improved loss function of MobileNetV3 network for:

[0103]

[0104] Step 3: input the image data of the electric energy meter to be identified into the abnormal electric energy meter identification model, and output the abnormal electric energy meter identification result.

[0105] During the specific implementation, the image data of electric energy meters collected on-site at a batch of single-phase electric energy meters calibration production line is used as sample data for implementation.

[0106] S1: First, the photographed smart meters are classified into two categories by manual screening: one is normal meters (positive samples) and the other is abnormal meters (negative samples). The classification method is to manually save the normal meter pictures in one folder and the abnormal meter pictures in another folder. After the classification is completed, the classified data set D will be obtained. The specific composition of the data D is shown in Table 1.

[0107] Table 1

[0108] Training set Test Set Normal meter 2000 500 Abnormal electric meter 2000 500 total 4000 1000

[0109] S2: Build an improved MobileNetV3 network model, bring the dataset D into the training model, and test the model performance with the test set after the training. The training was carried out for 10 iterations. The specific training and test results are as follows: Figure 3 and Figure 4 shown.

[0110] S3: At the same time, in order to further demonstrate the abnormality identification effect of the electric energy meter of the present invention, the method adopted by the present invention is compared with the test set of data set D of VGG16 and ResNet50 for a control experiment. The comparison effect is shown in Table 2.

[0111] Table 2

[0112]

[0113] The present invention (1) is based on the lightweight model design of MobileNetV3: This proposal adopts the improved MobileNetV3 network structure, which has efficient feature extraction capabilities and lightweight design, and is suitable for deployment on devices with limited computing resources to meet the needs of real-time pipeline detection. (2) Construction of an automated meter anomaly recognition model: This proposal realizes the automated construction of the appearance anomaly recognition model of the electric energy meter by improving the MobileNetV3 network, and can generate exclusive recognition models to adapt to different types of electric energy meters, thereby improving the generalization ability and adaptability of the model. (3) Data-driven adaptive learning strategy: Based on the division of training sets and validation sets, adaptive learning is achieved through MobileNetV3 training, which effectively improves the accuracy of anomaly recognition while retaining the recognition rate under pipeline conditions.

[0114] The present invention is suitable for identifying abnormalities in the appearance of electric energy meters under automated production conditions of centralized detection and verification of electric energy meters, and can adaptively generate exclusive abnormality recognition models for electric energy meters of different models, which can effectively improve the accuracy of the model while taking into account the recognition rate under assembly line conditions.

[0115] Example 2

[0116] like Figure 5 As shown, the difference between this embodiment and embodiment 1 is that this embodiment provides an electric meter anomaly identification system based on MobileNetV3, which corresponds to the electric meter anomaly identification method based on MobileNetV3 in embodiment 1 in a one-to-one manner; the system includes:

[0117] An acquisition unit, used for acquiring image data of abnormal electric energy meters and normal electric energy meters;

[0118] A preprocessing and division unit is used to preprocess the image data to obtain a sample data set, and randomly divide it into a training set and a validation set according to a proportion;

[0119] The model building and training unit is used to train the improved MobileNetV3 network based on the training set, verify the accuracy of the improved MobileNetV3 network judgment based on the verification set, and use the trained improved MobileNetV3 network as the abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module of reverse residual;

[0120] The abnormality recognition unit is used to input the image data of the electric energy meter to be recognized into the abnormal electric energy meter recognition model and output the abnormality recognition result of the electric energy meter.

[0121] As a further implementation, the improved MobileNetV3 network includes an input layer, a convolutional layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer, and an output layer connected in sequence;

[0122] The bottleneck module with reverse residual reduces the amount of calculation while maintaining performance through the expansion, deep convolution, and compression structure to obtain the data after passing through the bottleneck module with reverse residual;

[0123] The bottleneck module with reverse residual includes point-wise convolution expansion unit, depth-wise convolution unit and point-wise convolution compression unit;

[0124] a point-by-point convolution expansion unit, configured to expand the data with a lower number of channels through a first point-by-point convolution expansion operation to obtain data with a higher number of channels as expanded data;

[0125] A depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain the depth convolved data;

[0126] The point-by-point convolution compression unit is used to compress the data after the depth convolution back to data with a lower number of channels through a second point-by-point convolution.

[0127] Among them, the execution process of each unit can be executed according to the process steps of the electric meter abnormality identification method based on MobileNetV3 in Example 1, and will not be repeated one by one in this embodiment.

[0128] At the same time, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned electric meter anomaly identification method based on MobileNetV3 is implemented.

[0129] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0130] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0132] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0133] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The electric meter anomaly recognition method based on MobileNetV3 is characterized by: The method includes: Acquire image data of abnormal electric energy meters and normal electric energy meters, preprocess the image data to obtain a sample data set, and randomly divide the data into a training set and a validation set according to a proportion; An improved MobileNetV3 network is trained based on a training set, and the improved MobileNetV3 network is verified based on a verification set, and the trained improved MobileNetV3 network is used as an abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module with a reverse residual; The image data of the electric energy meter to be identified is input into the abnormal electric energy meter identification model, and the abnormal electric energy meter identification result is output.

2. According to the MobileNetV3-based electric meter anomaly identification method according to claim 1, it is characterized in that: Preprocessing the image data to obtain a sample data set includes: The photographed smart energy meters are classified into two categories through manual screening: the first category is normal meters, and the second category is abnormal meters. After the classification is completed, a classified sample data set will be obtained; The classification method is to save the normal meter pictures in one folder and the abnormal meter pictures in another folder.

3. The electric meter anomaly recognition method based on MobileNetV3 according to claim 1 is characterized in that: The improved MobileNetV3 network includes an input layer, a convolution layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer and an output layer connected in sequence; The bottleneck module with reverse residual reduces the amount of calculation while maintaining performance through an expansion, deep convolution, and compression structure to obtain data after passing through the bottleneck module with reverse residual.

4. According to the MobileNetV3-based electric meter anomaly identification method according to claim 3, it is characterized in that: The bottleneck module with reverse residual includes a point-by-point convolution expansion unit, a depth-wise convolution unit and a point-by-point convolution compression unit; The point-by-point convolution expansion unit is used to expand the data with a low number of channels through a first point-by-point convolution expansion operation to obtain data with a high number of channels as expanded data; The depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain depth-convolved data; The point-by-point convolution compression unit is used to compress the data after the depth convolution back into data with a low number of channels through a second point-by-point convolution.

5. According to the MobileNetV3-based electric meter anomaly identification method according to claim 4, it is characterized in that: The expression of the first point-by-point convolution is: expand =H_swish(Conv 1×1 (X)), where X expand It is the data expanded by point-by-point convolution. 1×1 () indicates that a convolution operation is performed using a 1×1 convolution kernel, and X is the independent variable input into the operation; The expression of the second point-by-point convolution is: compress =Conv 1×1 (X depthwise ), where X compress Represents data after point-by-point convolution compression.

6. The electric meter anomaly recognition method based on MobileNetV3 according to claim 4 is characterized in that: The operational expression of the depth convolution is: X depthwise =H_swish(DepthwiseConv 3×3 (X expand )) Where X depthwise DepthwiseConv is the data after deep convolution. 3×3 () indicates that the 3×3 convolution kernel performs a depthwise convolution operation, which is expressed as: where Y h,w,c Represents the value of channel c at position (h, w) after depth convolution; X h+i-1,w+j-1,c Indicates that the data that needs to be convolved is at position (h+i-1,w+j-1), the value of channel c; K i,j,c represents the weight of the 3×3 convolution kernel on the cth channel; b c Indicates the offset of the cth channel.

7. The method for identifying anomalies of electric meters based on MobileNetV3 according to claim 3 is characterized in that: The loss function of the improved MobileNetV3 network for: In the formula, i is the sample, y is i is the true label, is the predicted label.

8. The electric meter anomaly recognition system based on MobileNetV3 is characterized by: The system includes: An acquisition unit, used for acquiring image data of abnormal electric energy meters and normal electric energy meters; A preprocessing and division unit, used for preprocessing the image data to obtain a sample data set, and randomly dividing it into a training set and a validation set according to a proportion; A model building and training unit, used for training an improved MobileNetV3 network based on a training set, verifying the improved MobileNetV3 network based on a verification set, and using the trained improved MobileNetV3 network as an abnormal electric energy meter recognition model; the improved MobileNetV3 network is a MobileNetV3 network with a bottleneck module of a reverse residual; The abnormality identification unit is used to input the image data of the electric energy meter to be identified into the abnormal electric energy meter identification model and output the abnormality identification result of the electric energy meter.

9. The electric meter anomaly recognition system based on MobileNetV3 according to claim 8 is characterized in that: The improved MobileNetV3 network includes an input layer, a convolution layer, an activation layer, a bottleneck module with a reverse residual, a global pooling layer, an activation layer, a fully connected layer, a judgment layer and an output layer connected in sequence; The bottleneck module with reverse residual reduces the amount of calculation while maintaining performance through an expansion, deep convolution, and compression structure to obtain data after passing through the bottleneck module with reverse residual; The bottleneck module with reverse residual includes a point-by-point convolution expansion unit, a depth-wise convolution unit and a point-by-point convolution compression unit; The point-by-point convolution expansion unit is used to expand the data with a low number of channels through a first point-by-point convolution expansion operation to obtain data with a high number of channels as expanded data; The depth convolution unit is used to perform independent depth convolution on each channel in the expanded data to obtain depth-convolved data; The point-by-point convolution compression unit is used to compress the data after the depth convolution back into data with a low number of channels through a second point-by-point convolution.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the electric meter abnormality identification method based on MobileNetV3 as described in any one of claims 1 to 7 is implemented.

Images