Power transmission line icing thickness detection method based on deep convolutional neural network

Abstract

The invention discloses a power transmission line icing thickness detection method based on a deep convolutional neural network, and the method specifically comprises the steps: carrying out the image data collection of a power transmission line, and carrying out the manual marking of an image; preprocessing the manually labeled image to generate a data set, and dividing the data set into a training set, a verification set and a test set according to a proportion; selecting a deep convolutional neural network model; training the deep convolutional neural network model by using a training set, evaluating by using a verification set, and testing by using a test set to obtain a power transmission line icing thickness detection model; and deploying the power transmission line icing thickness detection model to a cloud server or an edge computing device, and carrying out icing thickness detection on the power transmission line. According to the transmission line icing thickness detection method based on the deep convolutional neural network, the problems of low detection efficiency, poor accuracy and high cost in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transmission line detection, and specifically relates to a method for detecting ice thickness of power transmission lines based on a deep convolutional neural network. Background Art

[0002] As my country's infrastructure continues to improve, the length and span of transmission lines are also increasing. A series of severe problems caused by the ice and snow climate have gradually emerged, posing a severe challenge to the power system that cannot be ignored. It can be seen that the study of intelligent detection of ice thickness on transmission lines can help power personnel quickly and accurately judge the ice condition and prevent ice accidents, which is of vital importance to the intelligent monitoring of ice on transmission lines.

[0003] Traditional methods for detecting ice thickness on transmission lines mainly include manual inspection and mechanical model calculation, but these methods have problems such as low efficiency, poor accuracy and high cost. With the continuous development of image interpretation technology, non-contact ice detection technology has gradually occupied a dominant position in the field of ice measurement. With the continuous advancement of information technology and artificial intelligence technology, the application of computer deep learning in ice thickness detection has become more and more extensive. Summary of the invention

[0004] The purpose of the present invention is to provide a method for detecting ice thickness on power transmission lines based on a deep convolutional neural network, which solves the problems of low detection efficiency, poor accuracy and high cost in the prior art.

[0005] The technical solution adopted by the present invention is a method for detecting ice thickness of a power transmission line based on a deep convolutional neural network, which specifically includes the following steps: S1: Collect image data of the transmission line and manually annotate the image; S2: Preprocess the manually annotated images to generate a data set, which is then divided into a training set, a validation set, and a test set in proportion; S3: Select a deep convolutional neural network model; train the deep convolutional neural network model using a training set, evaluate it using a validation set, and test it using a test set to obtain a transmission line ice thickness detection model; S4: Deploy the transmission line ice thickness detection model to the cloud server or edge computing device to detect the ice thickness of the transmission line.

[0006] The present invention is also characterized in that: S1 specifically includes: installing image acquisition equipment on the transmission line, and regularly collecting images of the transmission line with different degrees of ice coverage; the collected images cover various weather conditions and ice coverage conditions; and manually annotating the collected images, including the ice thickness level, which is specifically divided into light ice coverage, moderate ice coverage, and heavy ice coverage.

[0007] In S2, the manually annotated images are preprocessed, including image enhancement, denoising, and normalization; Image enhancement, including: random cropping, multi-angle rotation, resizing and flipping of images; adjusting the brightness, hue, saturation and contrast of images; adding random noise to images; Denoising includes the use of mean filtering, median filtering, and Gaussian filtering methods to remove noise interference in the image and enhance image details.

[0008] In S2, the dataset is divided into training set, validation set and test set according to the ratio of 2:1:1.

[0009] In S3, the EfficientNet-B4 architecture is selected as the deep convolutional neural network model. The EfficientNet-B4 architecture includes multiple MBConv modules, multiple convolutional layers, pooling layers, and fully connected layers. The MBConv module consists of two 1×1 convolutional layers, a Dropout layer, a depth-separable convolutional layer in the middle, and an SE attention module. The SE attention module consists of a global average pooling and two fully connected layers. The EfficientNet-B4 architecture extracts features from the input image. It extracts the same feature information of the input image through a combination of multiple convolutional layers and pooling layers. Specifically, the input image is first dimensionally increased through a 1×1 convolutional layer, and then a feature map is obtained through batch normalization and a Swish activation function. Next, the feature map is resized through a depthwise separable convolutional layer. The resized feature map enters the SE attention module to focus on key features. The SE attention module includes a Squeeze part and an Excitation part. The Squeeze part uses global average pooling to globally describe the features along the channel direction, reducing the original feature map of size H×W×C to 1×1×C, and converting the spatial features into global features, which can be expressed as:

[0010] In the formula, W , H Represents the dimension of the feature map, F sq represents global average pooling; u c Represents the output of channel c; u c ( i , j ) is the c channel ( i , j )'s position element; The Excitation part analyzes global features, reveals the dependencies between channels, optimizes the model's response to different channel features, and strengthens the model's ability to identify key features. Subsequently, the feature map after focusing on the key features through the SE attention module is again reduced in dimension through a 1×1 convolutional layer, and a BN layer and a Dropout layer are introduced to weaken the dependency between convolutional layers; The feature map processed by the MBConv module is input into the fully connected layer again. Based on the extracted feature information, the ice thickness level of the transmission line in the image is judged and the ice level result is output.

[0011] In S3, the deep convolutional neural network model is trained using the training set and the cross entropy loss function is adopted; the number of iterations is set to 30; the parameter optimizer is set to SGD; the gradient distribution of the sample is calculated in each training iteration, and the parameters are adjusted inversely according to the gradient distribution; the learning rate is set to 0.02, the learning rate momentum is set to 0.9, and the weight decay coefficient is set to 0.0001; During the training process, the validation set is used to evaluate the deep convolutional neural network model, and the model parameters are adjusted according to the evaluation results. When the performance of the model on the validation set no longer improves, the training is stopped; The test set is used to evaluate the accuracy, loss rate, and recall rate of the trained deep convolutional neural network model, and a transmission line ice thickness detection model is obtained.

[0012] S4 specifically includes the following steps: S4.1: Deploy the model to a cloud server or edge computing device and configure the operating environment; S4.2: The image is transmitted to the cloud server via a wireless network or to the edge computing device via a wired and / or wireless network to complete the ice thickness detection of the line image and output the ice thickness level.

[0013] The beneficial effects of the present invention are: The present invention is based on a transmission line ice thickness detection method based on a deep convolutional neural network, which adopts the EfficientNet-B4 network architecture. The MBConv module has many repetitions, and the model can process and transform features more deeply. Through multiple feature extraction and conversion, the model can better learn the essential features of the image, thereby improving the accuracy of classification. The MBConv module also incorporates the SE attention module, which enables the model to focus on important feature information by dynamically adjusting the dependencies between channels, helping the model to focus on key areas in the image and suppressing the interference of irrelevant information, thereby improving the quality of features and the accuracy of classification; and the EfficientNet-B4 network pioneered the combination of an efficient model parameter composite scaling strategy. By using advanced neural structure search technology, it can accurately find the optimal parameter ratio between depth, width and resolution, thereby maximizing model performance and optimizing processing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of a method for detecting ice thickness on a power transmission line based on a deep convolutional neural network of the present invention; Figure 2 It is a sample diagram of ice thickness classification of transmission lines in the method for detecting ice thickness of transmission lines based on a deep convolutional neural network of the present invention; Figure 3 It is a schematic diagram of the overall structure of the EfficientNet-B4 network in the power transmission line ice thickness detection method based on a deep convolutional neural network of the present invention; Figure 4 It is a structural schematic diagram of the MBConv module in the EfficientNet-B4 network in the power transmission line ice thickness detection method based on deep convolutional neural network of the present invention; Figure 5 It is a structural diagram of the SE attention mechanism module in the power transmission line ice thickness detection method based on deep convolutional neural network of the present invention; Figure 6 This is a graph of the detection results of ice thickness on power lines using a method for detecting ice thickness on power lines based on a deep convolutional neural network in the present invention. DETAILED DESCRIPTION

[0015] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] The present invention provides a method for detecting ice thickness on power transmission lines based on a deep convolutional neural network. Figure 1 As shown, the specific steps include: S1: Collect image data of the transmission line and manually annotate the image.

[0017] Given that power transmission lines are located in a changeable and complex natural environment, the commonly used image acquisition method is to install image acquisition equipment on the transmission lines, such as industrial cameras, high-definition cameras, etc., to regularly collect images of power transmission lines with different degrees of ice coverage; the collected images cover various weather conditions and ice coverage to ensure data diversity; the collected images are manually annotated, and the annotation content includes ice thickness levels, which are specifically divided into light ice coverage, moderate ice coverage, and heavy ice coverage. The sample picture of ice thickness level is as follows Figure 2 shown.

[0018] S2: Preprocess the manually annotated images to generate a data set, and divide it into training set, validation set, and test set according to the proportion.

[0019] Preprocess the collected images to improve image quality and model training effect. Preprocessing operations include: Image enhancement: In order to solve the problem of reduced model generalization ability due to the small size of the data set, the present invention uses data enhancement technology to expand the data dimension of the sample by adjusting the specific pixel values ​​in the original ice-covered image, thereby effectively improving the diversity of the sample. Specifically, it includes random cropping, multi-angle rotation, resizing and flipping operations on the image to increase the deformation diversity of the image; fine-tuning the brightness, hue, saturation and contrast of the image to simulate image changes under different lighting and color conditions; adding random noise to the image to simulate interference with the transmission signal. These comprehensive measures effectively expand the original data set, thereby improving the generalization performance and anti-interference ability of the model.

[0020] Denoising: Considering that images acquired in reality are often polluted by various noises, such as sensor noise, transmission noise, etc. These noises will make the image blurry and distorted, affecting the visual effect of the image. To this end, the present invention adopts the methods of mean filtering, median filtering, and Gaussian filtering to remove noise interference and make the image clearer and more natural; enhance image details to retain the details of the image as much as possible, making the details of the image more prominent. Provide a better foundation for subsequent image processing tasks and reduce the amount of calculation and storage space.

[0021] The processed data set is divided into training set, validation set and test set according to the ratio of 2:1:1.

[0022] S3: Select a deep convolutional neural network model; train the deep convolutional neural network model using a training set, evaluate it using a validation set, and test it using a test set to obtain a transmission line ice thickness detection model.

[0023] In the case of limited resources, the design and development of convolutional neural networks are usually subject to certain constraints. The availability of resources is usually proportional to the training accuracy of the model. In order to increase the availability of resources, the EfficientNet network pioneered a high-efficiency model parameter compound scaling strategy. It cleverly introduced a high-performance compound optimization coefficient and made detailed structural optimizations for the three key dimensions of the network: depth, width, and resolution.

[0024] By using advanced neural architecture search technology, we can accurately find the optimal parameter ratio between depth, width and resolution. In this way, we can maximize the model performance and optimize the processing speed. The high-performance composite optimization coefficient is as follows:

[0025]

[0026] in, d is the network depth; w is the network width; r is the input feature resolution; α is the depth allocation coefficient; β is the width allocation coefficient; γ is the resolution allocation coefficient, ϕ Indicates the size of consumed resources.

[0027] The size of the model mainly depends on the scaling parameters of three dimensions: width, depth and resolution. These three dimensions are interrelated and not independent of each other. For example, when processing higher resolution input images, a deeper network structure is required to ensure sufficient receptive field. Similarly, for the processing of high-resolution images, it is also required to increase the number of channels to capture more accurate feature information. This interdependence must be fully considered when selecting and optimizing the model. The EfficientNet series of models are most suitable for the intelligent identification of ice thickness on transmission lines, and can achieve higher accuracy and efficiency by scaling the multi-dimensional parameters of the network.

[0028] In the EfficientNet series of models, the accuracy of the model has gradually improved from EfficientNet-B0 to EfficientNet-B7, and its scale has been continuously expanded. The evolution of EfficientNet-B0 to EfficientNet-B7 versions is accompanied by different scaling between three dimensions. The present invention uses a total of eight models from the EfficientNet-B0 to EfficientNet-B7 series to compare the model accuracy and its loss value on the transmission line ice thickness test set. The specific evaluation indicators of the same series of models are shown in Table 1.

[0029] Table 1 Comparison of evaluation indicators of the same series of models

[0030] Through a comprehensive comparative analysis of different series of models in terms of Accuracy and Loss, the EfficientNet-B4 model has the best comprehensive performance evaluation index in the transmission line icing scenario. Specifically, EfficientNet-B4 is not only significantly higher than other models in terms of accuracy, but also performs better in terms of optimizing the loss function. This means that the model can not only more accurately identify and classify icing scenarios, but also more effectively reduce prediction errors, thereby improving the reliability and stability of the overall system.

[0031] Therefore, the present invention selects the EfficientNet-B4 architecture as the deep convolutional neural network model. Figure 3 As shown in the figure, it includes stem, 7 blocks and the final output layer Final layers. The EfficientNet-B4 model, combined with compound scaling technology, can effectively and intelligently detect the ice thickness of transmission lines. Compared with the EfficientNet-B0 baseline model, the EfficientNet-B4 network has more channels. More channels allow the model to process more information in each layer, enhancing the model's ability to extract different features. In terms of resolution adaptability, EfficientNet-B4 can better utilize the fine-grained information in high-resolution images when processing images of different resolutions, while also maintaining good performance on lower-resolution images.

[0032] The EfficientNet-B4 architecture includes multiple MBConv modules, multiple convolutional layers, pooling layers, and fully connected layers; Figure 4 As shown in the figure, the MBConv module consists of two 1×1 convolutional layers, a Dropout layer, a depth-separable convolutional layer in the middle, and an SE attention module; the SE attention module consists of a global average pooling and two fully connected layers.

[0033] The SE attention module includes a Squeeze part and an Excitation part, such as Figure 5 As shown in the figure, the squeeze part uses global average pooling to describe the features globally along the channel direction, reducing the original feature map of size H×W×C to 1×1×C, and converting spatial features into global features, which can be expressed as:

[0034] In the formula,W , H Represents the dimension of the feature map, F sq represents global average pooling; u c Represents the output of channel c; u c ( i , j ) is the c channel ( i , j )'s position element; Secondly, the Excitation part analyzes global features, reveals the dependencies between channels, optimizes the model's response to different channel features, and strengthens the model's ability to identify key features. By integrating Squeeze and Excitation, the SE module effectively improves the network's deep understanding of channel relationships and significantly enhances the overall performance of the model.

[0035] Subsequently, the feature map after focusing on the key features through the SE attention module is again reduced in dimension through a 1×1 convolutional layer, and a BN layer and a Dropout layer are introduced to weaken the dependency between convolutional layers; The feature map processed by the MBConv module is input into the fully connected layer again. Based on the extracted feature information, the ice thickness level of the transmission line in the image is judged and the ice level result is output.

[0036] Use the training set to train the deep convolutional neural network model, using the cross entropy loss function; set the number of iterations to 30; set the parameter optimizer to SGD; calculate the gradient distribution of the sample in each training iteration, and adjust the parameters inversely according to the gradient distribution; set the learning rate to 0.02, the learning rate momentum to 0.9, and the weight decay coefficient to 0.0001; During the training process, the validation set is used to evaluate the deep convolutional neural network model, and the model parameters are adjusted according to the evaluation results. When the performance of the model on the validation set no longer improves, the training is stopped; The test set is used to evaluate the accuracy, loss rate, and recall rate of the trained deep convolutional neural network model, and a transmission line ice thickness detection model is obtained.

[0037] The accuracy rate refers to the ratio of the number of samples correctly predicted by the model to the total number of samples, which is specifically the ratio of the number of images with correctly detected ice thickness of power transmission lines to the total number of images in the entire data set. The Loss function plays a vital role in deep learning. It is a key indicator for optimizing model performance and improving prediction accuracy. If the Loss value is too high, it may be that the model is overfitting or the original data is wrong, so the lower the Loss value of the model, the better. The recall rate refers to the ratio of the number of positive samples correctly predicted by the model to the actual number of positive samples.

[0038] S4: Deploy the transmission line ice thickness detection model to the cloud server or edge computing device to detect the ice thickness of the transmission line.

[0039] S4.1: Deploy the model to a cloud server or edge computing device and configure the operating environment; In practical applications, the model can be deployed on a cloud server or an edge computing device. The edge computing device can use NVIDIA Jetson ORIN NX CLB 256 [16G].

[0040] S4.2: The image is transmitted to the cloud server via a wireless network or to the edge computing device via a wired and / or wireless network to complete the ice thickness detection of the line image and output the ice thickness level.

[0041] If deployed on a cloud server, the image is transmitted to the model on the cloud server via a wireless network for detection. The image will be encrypted and compressed during the transmission process to ensure data security and transmission efficiency. The model on the cloud server will quickly analyze and process the image and return the detection results to the user.

[0042] If deployed on edge computing devices, the edge computing devices can be directly installed at monitoring sites near the transmission lines. Images can be transmitted to the model on the edge computing device through a local wireless network or a wired network for detection. Edge computing devices have low latency and high computing performance, and can achieve real-time monitoring of the ice coverage of transmission lines. When it is detected that the ice coverage thickness exceeds a certain threshold, the edge computing device can immediately issue an alarm to notify relevant personnel to take appropriate measures.

[0043] In this way, whether deployed on a cloud server or an edge computing device, real-time monitoring of the icing condition of the transmission line can be achieved. Figure 6 To provide a strong guarantee for the safe operation of the power system.

[0044] Example 1 This embodiment provides a method for detecting ice thickness on a power transmission line based on a deep convolutional neural network. Figure 1 As shown, the specific steps include: S1: Collect image data of the transmission line and manually annotate the image; S2: Preprocess the manually annotated images to generate a data set, which is then divided into a training set, a validation set, and a test set in proportion; S3: Select a deep convolutional neural network model; train the deep convolutional neural network model using a training set, evaluate it using a validation set, and test it using a test set to obtain a transmission line ice thickness detection model; S4: Deploy the transmission line ice thickness detection model to the cloud server or edge computing device to detect the ice thickness of the transmission line.

[0045] Example 2 Based on Example 1, S1 specifically comprises: installing image acquisition equipment on the transmission line, and regularly acquiring images of the transmission line with different degrees of ice coverage; the acquired images cover various weather conditions and ice coverage conditions; and manually annotating the acquired images, wherein the annotated contents include ice coverage thickness levels, which are specifically divided into light ice coverage, moderate ice coverage, and heavy ice coverage.

[0046] In S2, the manually annotated images are preprocessed, including image enhancement, denoising, and normalization; Image enhancement, including: random cropping, multi-angle rotation, resizing and flipping of images; adjusting the brightness, hue, saturation and contrast of images; adding random noise to images; Denoising includes the use of mean filtering, median filtering, and Gaussian filtering methods to remove noise interference in the image and enhance image details.

[0047] The dataset is divided into training set, validation set and test set according to the ratio of 2:1:1.

[0048] In S3, the EfficientNet-B4 architecture is selected as the deep convolutional neural network model. The EfficientNet-B4 architecture includes multiple MBConv modules, multiple convolutional layers, pooling layers, and fully connected layers. The MBConv module consists of two 1×1 convolutional layers, a Dropout layer, a depth-separable convolutional layer in the middle, and an SE attention module. The SE attention module consists of a global average pooling and two fully connected layers. The EfficientNet-B4 architecture extracts features from the input image. It extracts the same feature information of the input image through a combination of multiple convolutional layers and pooling layers. Specifically, the input image is first dimensionally increased through a 1×1 convolutional layer, and then a feature map is obtained through batch normalization and a Swish activation function. Next, the feature map is resized through a depthwise separable convolutional layer. The resized feature map enters the SE attention module to focus on key features. The SE attention module includes a Squeeze part and an Excitation part. The Squeeze part uses global average pooling to globally describe the features along the channel direction, reducing the original feature map of size H×W×C to 1×1×C, and converting the spatial features into global features, which can be expressed as:

[0049] In the formula, W , H Represents the dimension of the feature map, F sq represents global average pooling; u c Represents the output of channel c; u c ( i , j ) is the c channel ( i , j )'s position element; The Excitation part analyzes global features, reveals the dependencies between channels, optimizes the model's response to different channel features, and strengthens the model's ability to identify key features. Subsequently, the feature map after focusing on the key features through the SE attention module is again reduced in dimension through a 1×1 convolutional layer, and a BN layer and a Dropout layer are introduced to weaken the dependency between convolutional layers; The feature map processed by the MBConv module is input into the fully connected layer again. Based on the extracted feature information, the ice thickness level of the transmission line in the image is judged and the ice level result is output.

[0050] Example 3 On the basis of Example 2, in S3, the deep convolutional neural network model is trained using the training set and the cross entropy loss function is adopted; the number of iterations is set to 30; the parameter optimizer is set to SGD; the gradient distribution of the sample is calculated in each training iteration, and the parameters are adjusted inversely according to the gradient distribution; the learning rate is set to 0.02, the learning rate momentum is set to 0.9, and the weight decay coefficient is set to 0.0001; During the training process, the validation set is used to evaluate the deep convolutional neural network model, and the model parameters are adjusted according to the evaluation results. When the performance of the model on the validation set no longer improves, the training is stopped; The test set is used to evaluate the accuracy, loss rate, and recall rate of the trained deep convolutional neural network model, and a transmission line ice thickness detection model is obtained.

[0051] S4 specifically includes the following steps: S4.1: Deploy the model to a cloud server or edge computing device and configure the operating environment; S4.2: The image is transmitted to the cloud server via a wireless network or to the edge computing device via a wired and / or wireless network to complete the ice thickness detection of the line image and output the ice thickness level.

Claims

1. A method for detecting ice thickness on power transmission lines based on deep convolutional neural networks, characterized in that: The specific steps include: S1: Collect image data of the transmission line and manually annotate the image; S2: Preprocess the manually annotated images to generate a data set, which is then divided into a training set, a validation set, and a test set in proportion; S3: Select a deep convolutional neural network model; and train the deep convolutional neural network model using the training set, evaluate it using the validation set, and test it using the test set to obtain a transmission line ice thickness detection model; S4: Deploy the transmission line ice thickness detection model to the cloud server or edge computing device to detect the ice thickness of the transmission line.

2. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: S1 specifically includes: installing image acquisition equipment on the transmission line, and regularly collecting images of the transmission line with different degrees of ice coverage; the collected images cover various weather conditions and ice coverage conditions; and manually annotating the collected images, including the ice thickness level, which is specifically divided into light ice coverage, moderate ice coverage, and heavy ice coverage.

3. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: In S2, the manually annotated image is preprocessed, and the preprocessing includes image enhancement, denoising, and normalization; The image enhancement specifically includes: performing random cropping, multi-angle rotation, resizing and flipping operations on the image; adjusting the brightness, hue, saturation and contrast of the image; Add random noise to the image; The denoising includes adopting the methods of mean filtering, median filtering and Gaussian filtering to remove the noise interference in the image and enhance the image details.

4. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: In S2, the dataset is divided into training set, validation set and test set according to the ratio of 2:1:

1.

5. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: In S3, the EfficientNet-B4 architecture is selected as the deep convolutional neural network model. The EfficientNet-B4 architecture includes multiple MBConv modules, multiple convolutional layers, pooling layers, and fully connected layers. The MBConv module consists of two 1×1 convolutional layers, a Dropout layer, a depth-separable convolutional layer in the middle, and an SE attention module. The SE attention module consists of a global average pooling and two fully connected layers. The EfficientNet-B4 architecture extracts features from the input image. It extracts the same feature information of the input image through a combination of multiple convolutional layers and pooling layers. Specifically, the input image is first dimensionally increased through a 1×1 convolutional layer, and then a feature map is obtained through batch normalization and a Swish activation function. Next, the feature map is resized through a depthwise separable convolutional layer. The resized feature map enters the SE attention module to focus on key features. The SE attention module includes a Squeeze part and an Excitation part. The Squeeze part uses global average pooling to globally describe the features along the channel direction, reducing the original feature map of size H×W×C to 1×1×C, and converting the spatial features into global features, which can be expressed as: In the formula, W , H Represents the dimension of the feature map, F sq represents global average pooling; u c Represents the output of channel c; u c ( i , j ) is the c channel ( i , j )'s position element; The Excitation part analyzes global features, reveals the dependencies between channels, optimizes the model's response to different channel features, and strengthens the model's ability to identify key features. Subsequently, the feature map after focusing on the key features through the SE attention module is again reduced in dimension through a 1×1 convolutional layer, and a BN layer and a Dropout layer are introduced to weaken the dependency between convolutional layers; The feature map processed by the MBConv module is input into the fully connected layer again. Based on the extracted feature information, the ice thickness level of the transmission line in the image is judged and the ice level result is output.

6. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: In S3, the deep convolutional neural network model is trained using the training set and the cross entropy loss function. The number of iterations is set to 30. The parameter optimizer is set to SGD. The gradient distribution of the sample is calculated in each training iteration, and the parameters are adjusted inversely according to the gradient distribution. Set the learning rate to 0.02, the learning rate momentum to 0.9, and the weight decay coefficient to 0.0001; During the training process, the validation set is used to evaluate the deep convolutional neural network model, and the model parameters are adjusted according to the evaluation results. When the performance of the model on the validation set no longer improves, the training is stopped; The test set is used to evaluate the accuracy, loss rate, and recall rate of the trained deep convolutional neural network model, and a transmission line ice thickness detection model is obtained.

7. The method for detecting ice thickness of power transmission lines based on deep convolutional neural network according to claim 1 is characterized in that: S4 specifically includes the following steps: S4.1: Deploy the model to a cloud server or edge computing device and configure the operating environment; S4.2: The image is transmitted to the cloud server via a wireless network or to the edge computing device via a wired and / or wireless network to complete the ice thickness detection of the line image and output the ice thickness level.

Images