EDPNet: A Transmission Line Ice-Thickness Recognition End-Side Network Based on Efficient Dynamic Perception

Abstract

Ice-covered transmission lines seriously affect the stable operation of the power system. Deploying a recognition network for measuring the ice thickness on transmission lines within a deicing robot, and controlling the robot to perform resonant deicing, is an effective solution. In order to solve the problem that the existing recognition network is not suitable for an edge device, an ice-thickness recognition network for transmission lines based on efficient dynamic perception (EDPNet) is proposed. Firstly, a lightweight multidimensional recombination convolution (LMRC) is designed to split the ordinary convolution for lightweight design and extract feature information of different scales for reorganization. Then, a lightweight deep fusion module (LDFM) is designed, which combines the attention mechanism with different features to enhance the information interaction between the encoder and decoder. Then, a new dynamic loss function is adopted in the training process to guide the model to perform refined detection of ice-covered boundaries. Finally, we count the ice pixels and calculate the ice thickness. The model is deployed on an OrangePi5 Plus edge computing board. Compared with the baseline model, the maximum ice-thickness detection error is 4.2%, the model parameters are reduced by 86.1%, and the detection speed is increased by 74.6%. Experimental results show that EDPNet can efficiently complete the task of identifying ice-covered transmission lines and has certain engineering application value.

1. Introduction

In winter, transmission lines are affected by severe weather such as rain and snow. This can cause ice to form on them, which may lead to tower collapse and disrupt the normal operation of the power system [1,2]. Since most transmission lines are built in remote, wild environments, power operation and maintenance personnel are unable to understand the ice status in real time, which brings great inconvenience to deicing work. Deicing robots on transmission lines can perform automated deicing [3,4]. The robot’s camera captures images of ice-covered transmission lines and uses a deep learning algorithm to calculate the thickness of the ice, thereby controlling the robot to perform resonant deicing [5]. The calculation of ice thickness is key to the entire deicing process. Since the network coverage in the field environment where the transmission lines are installed is poor, cloud data transmission and calculation cannot be achieved. Therefore, it is necessary to identify the ice thickness on the robot’s side [6].

Among the methods of ice-thickness recognition, the use of image detection technology is currently one of the mainstream methods [7]. The principle is to segment the ice-covered area from the non-ice-covered area in the ice-covered image through a deep learning algorithm and then count the pixels in the ice-covered area to calculate the ice thickness [8]. Image segmentation is a computationally intensive computer vision task [9], which requires more computing resources than image classification [10] and object detection algorithms [11]. Currently, commonly used image segmentation algorithms such as U-Net [12] and Deeplab-V3 [13] have complex network models, while the computing power of edge devices is limited. Such algorithms are not suitable for computing on edge devices [14]. Therefore, how to design an ice-thickness detection algorithm suitable for edge devices is the focus of this article.

We propose a lightweight transmission line ice-thickness recognition network based on efficient dynamic perception. The main framework of EDPNet is the encoding and decoding architecture, which has been redesigned into the following three parts: the encoding part, decoding part, and loss function. The LMRC block is designed in the encoding part. By splitting the ordinary convolution and combining separable asymmetric convolutions, the number of parameters is greatly reduced without compromising the feature extraction capability. In the decoding part, we propose LDFM to dynamically fuse the encoding and decoding features through a lightweight attention module. Finally, in the loss function part, a dynamic loss function is designed to train the model in stages. After the model optimization becomes stable, we introduce boundary refinement loss for more accurate model training. Therefore, our contributions can be summarized as follows:

(1)

A lightweight multidimensional recombination convolution (LMRC) block is designed to efficiently extract ice image features.

(2)

A lightweight deep fusion module (LDFM) is proposed to perform lightweight dynamic fusion of encoding and decoding features.

(3)

A new dynamic loss function is designed, and a boundary refinement loss function is introduced to refine model training.

(4)

It was trained on a dataset of ice-covered transmission lines and deployed on an OrangePi5 Plus edge device, with a thickness recognition error of 4.2% and a detection speed of 10 FPS.

The structure of this paper is as follows: Chapter 1 introduces the research background and significance of identifying the thickness of ice cover on transmission lines. Chapter 2 introduces related work. Chapter 3 elaborates on the composition of EDPNet. Chapter 4 presents experimental results and analysis. And finally, the conclusion is drawn.

2. Related Work

Currently, there are three types of algorithms for identifying the ice thickness of a transmission line: manual measurement method [15], physical model method [16], and image detection method [17]. The manual measurement method involves the power grid staff climbing up the transmission tower to measure the ice thickness of the transmission line. This method has the most accurate measurement results, but it is dangerous. Moreover, the manual measurement method cannot meet the needs of real-time detection. Usually, the measurement and deicing operation can only be carried out after the transmission line is covered with ice. Therefore, this method is not widely used at present. The physical model method collects temperature, humidity, and meteorological data through sensors, then builds an ice-thickness prediction model and sends the collected data into the model for prediction. Veerakumar et al. [18] studied and investigated the mechanism of icing on transmission lines and summarized the environmental factors under different icing conditions. Zarnani et al. [19] constructed an ice-covered transmission line prediction model based on meteorological characteristics and achieved good results in experiments on different datasets. This method reduces the danger of on-site operations, but it needs to consider many influencing factors and relies heavily on expert knowledge and experience, so the prediction results are often not accurate enough. Moreover, when the input variables change, the prediction model also needs to be redesigned [20]. The most popular method for identifying the ice thickness on transmission lines is the image detection method, which uses a camera to obtain an image of the ice-covered transmission lines and then uses an image algorithm to calculate the ice thickness [21]. The image detection method can obtain the ice thickness on transmission lines in real time and intuitively.

Image detection methods can be divided into traditional edge detection methods [22] and image segmentation methods. Traditional edge detection methods use edge algorithms to extract the ice edge of a transmission line, then calculate the pixel difference between the upper and lower edges and obtain the ice thickness by comparing the pixel difference with the normal transmission line. Chang et al. [23] used the Ratio edge detection operator to perform edge detection on an ice-covered transmission line image and then calculated the ice thickness. Liang et al. [24] used morphological processing methods to preprocess the ice-covered image, then used the Line Segment Detector (LSD) algorithm for edge detection, and finally compared it with the diameter of the transmission line to calculate the ice thickness. The limitation of this method is that irregular ice will form on the surface of the transmission line, and the pixel difference between the upper and lower edges is obvious, which is easily affected by errors. Secondly, most current studies use traditional edge detection operators such as Canny [25] and Sobel [26]. These operators cannot accurately identify the ice edge when faced with complex scenes, resulting in the inability to detect the ice thickness.

The image segmentation method distinguishes the ice-covered area from non-ice-covered area and then counts the number of pixels in the ice-covered area and compares it with the number of pixels in the non-ice-covered area to obtain the ice thickness on transmission lines. This type of method often uses deep learning technology to segment ice-covered areas, such as U-Net. Hu et al. [27] combined U-Net with a generative adversarial network and used Swin Transformer and Convolutional Block Attention Module (CBAM) to perform image segmentation of ice-covered transmission lines. He et al. [28] denoised the ice-covered transmission line image, then segmented the ice-covered area and calculated the ice thickness using regional pixel statistics. Pan et al. [29] proposed a Soft Label Mask Optimization (SLMO) network, in which a pixel shuffling module was introduced to replace the traditional upsampling module, and used an ice-covered transmission line dataset collected by a drone to segment the ice-covered area.

The currently popular deep learning models are not suitable for edge computing devices. The computing power of edge devices is weak and cannot support the computing requirements of these models, so the models need to be lightweight. Currently, common lightweight models include MobileNet [30], ShuffleNet [31], SqueezeNet [32], etc. Abbas et al. [33] proposed a lightweight framework for accurately segmenting road cracks, achieving real-time processing capability. CamoFocus [34] introduces feature segmentation and modulation modules, as well as context refinement modules, to improve the performance of camouflage object detection. In addition, there are also lightweight networks designed for image segmentation, such as ENet [35], DABNet [36], etc. Wang et al. [37] proposed a compression model for ice-thickness recognition. They extracted image features through a lightweight MobileNet and then used channel pruning to compress the model, making it suitable for end-side deployment. Xu et al. [38] proposed an ice detection method suitable for edge devices. They used Ascend 310 to perform recognition model inference, classified the ice thickness into different levels, and determined the severity of the ice. These models do not optimize the design for the ice-covered transmission lines. This paper not only takes lightweight design into account, but also takes into account the high-precision ice segmentation effect.

3. Efficient Dynamic Perception Net

In view of the existing problems in the identification of ice thickness on transmission lines, this paper proposes an efficient dynamic perception end-side recognition network. EDPNet reduces the number of model parameters and improves the segmentation accuracy of the model by redesigning the encoder, decoder, and loss function.

The following is the pseudocode implementation of Algorithm 1, namely EDPNet:

Algorithm 1: EDPNet

1: Encoder Feature = LMRC(Input)

2: Decoder Feature = DeConv(Encoder Feature)

3: Fusion Feature = LDFM(Encoder Feature, Decoder Feature)

4: Ice Thickness = Ice Compute(Fusion Feature)

Figure 1 is a schematic diagram of the EDPNet proposed in this paper. Initially, we preprocess the ice-covered transmission line image and then direct it to the encoder to extract multi-dimensional image features. The encoder part is composed of LMRC blocks, which use the idea of lightweight segmentation and reorganization to reduce model parameters and improve the feature extraction capability of the LMRC block. The function of the LMRC block is to extract features of ice-cover images at different levels. Then, in the decoder section, the encoder features and decoder features are jointly processed through a lightweight deep fusion LDFM module. The function of this module is to integrate the image features extracted by the LMRC block with the DeConv upsampling features, enhance the information exchange between the two parts, and facilitate the generation of ice segmentation images. After multiple LDFM modules and upsampling modules, the image is restored to the size of the input image. Finally, the ice thickness is obtained using the pixel statistics method for the ice-covered area. During the training process, we employ a dynamic loss function to refine the segmentation image further.

3.1. Image Preprocess

The robot cannot use the collected images directly for ice detection. The size of the original image is too large; therefore, direct detection will affect the computing speed of the edge device. Therefore, it is necessary to normalize the size and uniformly scale it to 512 × 512. We can send the original images to EDPNet in batches for model training only after they have been uniformly resized. When the robot collects images, it is affected by sunlight, and the ice-covered images obtained will be too bright. Therefore, we use the histogram equalization method to adjust the image’s brightness and contrast, thereby reducing the impact of lighting on thickness detection.

3.2. Image Segmentation

EDPNet segments the ice-covered area and the non-ice-covered area of the pre-processed ice-covered image. The encoder incorporates a lightweight design that involves the decomposition and reorganization of standard convolutional operations, and the decoder performs lightweight deep fusion of the encoding and decoding features, which improves the generation capability of ice-covered segmentation images. In the second half of the training, the dynamic loss function enhances the model’s ability to recognize the edge of ice-covered areas through the boundary refinement loss function, making the generated ice-covered areas more refined.

3.2.1. LMRC

The encoder extracts feature information of different levels from the ice-covered transmission line image, which is related to the quality of subsequent ice-covered segmentation image generation. Therefore, the lightweight design of the encoder part is the focus of this paper. Currently, depthwise separable convolution is widely used in lightweight networks. Depthwise separable convolution splits ordinary convolution into depth convolution and point convolution. Each convolution kernel of the depth convolution is only responsible for the calculation of one channel, so its output channel is the same as the input channel number. Since the depth convolution does not change the number of feature channels, resulting in too little feature information of the image, the number of channels is increased by using a 1 × 1 convolution kernel of the point convolution. The depth convolution is responsible for extracting spatial information, and the point convolution is responsible for extracting channel information. The number of parameters for ordinary convolution is

$$Par{a}_{conv}=D_K\times D_K\times M\times N$$

(1)

$D_K$ is the convolution kernel size, $M$ is the number of input channels, and $N$ is the number of output channels. The number of parameters of separable convolution is

$$Par{a}_{Deconv}=D_K\times D_K\times M+M\times N$$

(2)

The parameter ratio is

$${\displaystyle \frac{Par{a}_{conv}}{Par{a}_{Deconv}}}={\displaystyle \frac{D_K\times D_K\times M\times N}{D_K\times D_K\times M+M\times N}}=N+{D_K}^2$$

(3)

Therefore, depthwise separable convolution reduces the number of parameters of ordinary convolution, enabling the model to run on edge devices. In addition, asymmetric convolution can also reduce the number of convolution calculations and is more suitable for bar-shaped objects such as transmission lines. The convolution kernel of ordinary convolution is a square such as 3 × 3, while the asymmetric convolution kernel splits it into 1 × 3 and 3 × 1. This asymmetric convolution improves the model’s ability to detect bar-shaped objects and reduces the number of model parameters.

The surface of the ice-covered transmission line is irregular in shape, and convolution kernels of different sizes can extract multi-scale image feature information. This paper uses the ideas of depthwise separable convolution, asymmetric convolution, and multi-scale to design a LMRC block. Figure 2 is a schematic diagram of the LMRC block.

The following is the pseudocode implementation of Algorithm 2, namely LMRC:

Algorithm 2: LMRC

1: Feature = Conv_3×3(Input)

2: Feature = PointConv_1×1(Feature)

3: for dil ∈ {1,2,5,8} do

4:  Feature1 = DepthConv_1×3(Feature, dil = 1)

5:  Feature1 = DepthConv_3×1(Feature1, dil = 1)

6: Output = Concat(Feature1,∙∙∙, Feature4)

The LMRC block consists of two parts. Firstly, the channel is converted through a 3 × 3 ordinary convolution, doubling the number of input channels to enhance the richness of features. Then, the features are extracted using a lightweight multidimensional reorganization convolution. A 1 × 1 convolution is used to decompose the input channel into four branches with the same number of channels, which then use point convolution to reduce the number of parameters. The asymmetric convolutions of different scales are used to extract multi-dimensional spatial information, and the change of scale is controlled by the dilation rate. Dilated convolution inserts gaps in the intervals of regular convolutions, expanding the receptive field range without increasing the number of convolution parameters. Finally, these decomposed branches are recombined to obtain the output features. The aggregation of feature information at different scales enhances the expressive power of the features.

The number of parameters for lightweight multidimensional reorganization convolution is

$$Par{a}_{LMRC}={\displaystyle \frac{\left(M+8\times D_K\right)\times N}{4}}$$

(4)

The parameter ratio with ordinary convolution is

$${\displaystyle \frac{Par{a}_{conv}}{Par{a}_{LMRC}}}=4{D_K}^{2}N+2D_{K}M$$

(5)

3.2.2. LDFM

We feed the ice-covered transmission line image to the encoder to extract image features at different levels and then fuse these features with the decoder features through cross-layer connections. In U-Net, feature fusion is achieved by cascading encoder and decoder features. This approach of only concatenating two features is relatively simple, resulting in incomplete fusion between encoder and decoder features. We propose a lightweight deep fusion module that uses lightweight attention to deeply fuse the features of the encoder and decoder.

ECA-Net is an efficient and lightweight attention module proposed by Wang et al. [39] in 2020. ECA-Net uses a local cross-channel interaction strategy to significantly reduce the complexity of the attention module while maintaining performance. Figure 3 is a schematic diagram of ECA-Net. First, after using Global Average Pooling (GAP) without dimensionality reduction to aggregate convolution features, the one-dimensional convolution kernel size K is adaptively selected. After one-dimensional convolution, the channel attention is calculated using the Sigmoid function.

Figure 4 is a schematic diagram of the lightweight deep fusion module. The following is the pseudocode implementation of Algorithm 3, namely LDFM:

Algorithm 3: LDFM

1: Feature = Add(Encoder Input, Decoder Input)

2: Attention = ECA-Net(Feature):

3:  Feature = AvgPool(Feature)

4:  Attention = Conv_1×1(Feature)

5:  Attention = Sigmoid(Attention)

6: Encoder Output = Attention × Encoder Input

7: Decoder Output = Attention × Decoder Input

The lightweight deep fusion module first adds the encoder and decoder features to fully fuse the feature information, then sends the fused features to ECA-Net to extract the fused attention information. The fused features undergo average pooling along the spatial dimensions to compress spatial information. Then, 1 × 1 convolution is used to extract channel attention, and the attention weights are multiplied with the encoder and decoder features to obtain the attention-weighted features. Finally, the cascaded encoder and decoder features are sent to the decoder to generate the ice segmentation image.

The attention weights extracted after adding the encoder and decoder are calculated from the fused features and contain information about the encoding and decoding features. Assigning attention weights to both helps the decoder pay more attention to the important features.

3.2.3. Dynamic Loss Function

During the EDPNet training process, we use a dynamic loss function. This loss function consists of two parts. The first half is the cross-entropy and Dice loss functions, which are used for rapid convergence of the model. The second half is the boundary refinement loss function, which is used to refine the output ice-covered area.

The loss function in the first half is composed of the cross entropy loss function and the Dice loss function. The cross entropy loss function is used to calculate the pixel-level loss, and the Dice loss function calculates the overall global loss. The calculation formula of the cross entropy loss function is

$$L_{BCE}=-\left[ylog\widehat{y}+\left(1-y\right)log\left(1-\widehat{y}\right)\right]$$

(6)

where $y$ represents the true pixel value, taking the value of 0 or 1.$\widehat{y}$ represents the predicted probability of the label. During model training, the cross-entropy loss function compares the difference between the true value and the predicted value of a single pixel, so it can be seen as optimizing the model from a microscopic perspective. The Dice coefficient is a function used to measure the similarity of a set, which can calculate the similarity between two different samples. The calculation formula for the Dice coefficient is

$$Dice={\displaystyle \frac{2\left|X\cap Y\right|}{\left|X\right|+\left|Y\right|}}$$

(7)

where $X$ and $Y$, respectively, represent the set of real pixels and the set of predicted pixels for image segmentation. $\left|X\cap Y\right|$ represents the intersection of the predicted and actual segmentation images, which is the set of correctly predicted pixels. $\left|X\right|+\left|Y\right|$ represents the total set of pixels in the predicted and actual segmentation images. The Dice coefficient focuses on the degree of overlap between the real segmentation image and the predicted segmentation image, emphasizing the optimization of the model from a holistic perspective. The Dice loss function is derived from the Dice coefficient, and its calculation formula is

$$L_{Dice}=1-{\displaystyle \frac{2\left|X\cap Y\right|}{\left|X\right|+\left|Y\right|}}$$

(8)

Therefore, we combine the cross-entropy loss function with the Dice loss function for the early training of the model and quickly optimize the model from both local and global levels. The cross-entropy loss function focuses on the difference between a single pixel prediction and the true value, while the Dice loss function measures the overall similarity between the predicted segmentation image and the true segmentation image. When the model is trained to a certain extent, the overall difference between the predicted image and the real label gradually decreases, but there is still a considerable gap in the details of the ice boundary. In the second half of model training, we introduce the boundary refinement loss function DouLoss [40]. This loss function optimizes the model by comparing the boundary difference between the real label of the ice-covered image and the predicted mask, guiding the model to generate a more refined ice-covered segmentation image. The calculation formula of DouLoss is as follows:

$$L_{DoU}={\displaystyle \frac{G\cup P-G\cap P}{G\cup P-\alpha \times G\cap P}}$$

(9)

$G$ represents the true value, $P$ represents the predicted value, and $\alpha$ is the adjustment factor. $G\cup P$ represents the union of the true value and the predicted value, $G\cup P$ represents the intersection of the true value and the predicted value, and $G\cup P-G\cap P$ is the difference set between the true label and the predicted value, which includes all misclassified pixels. $G\cup P-\alpha \times G\cap P$ is a part of the area where the boundary between the true value and the predicted value overlaps. By calculating the difference set and the ratio of this part, the closeness between the predicted boundary and the true boundary is represented. DouLoss encourages the model to fine tune the boundary regions, thereby improving the overall segmentation performance. Therefore, in the latter half of the model training, DouLoss is added to refine the segmentation of ice-covered areas. The dynamic loss function can be expressed as

$$Loss=\left\{\begin{array}{c}{L}_{BCE}+L_{Dice}, Training<50\%\\ L_{BCE}+L_{Dice}+L_{DoU}, Training\ge 50\%\end{array}\right.$$

(10)

In the first 50% of training, the cross-entropy loss function and Dice loss function quickly optimize the output of the model. In the second half of the stage, a boundary refinement loss function is added to optimize the boundaries of the ice-covered segmentation image and further refine the output.

3.3. Ice Thickness Calculation

After obtaining the segmented image of ice-covered transmission lines, the ice-covered area and the background area are filled with black and white pixels of different pixel values; the ice-covered area is marked as white pixels, and the background area is marked as black pixels, achieving binary processing of the image. Then, by traversing the number of pixels in the white and black areas, the ice-covered area and background area can be calculated. By calculating the proportion of white pixels in the overall image and then the proportion of the ice-covered area to the overall area, the resulting ratio can be calculated.

In order to accurately identify the thickness of ice, it is necessary to calculate the pixel ratio of the transmission line itself to the overall image when there is no ice cover. By obtaining clear images of the same transmission line in an ice-free state and repeating the aforementioned segmentation, binarization, and pixel statistics processes, the proportion of ice-free transmission lines can be determined. Comparing the actual pixel area under ice cover with the theoretical pixel area without ice cover, infer the additional pixel area due to ice cover and combine it with the known radius of the transmission line to calculate the ice-cover thickness. The calculation formula for the ice thickness of the transmission line is as follows:

$$h={\displaystyle \frac{1}{2}}\left({\displaystyle \frac{D_1{\mathrm{d}}_2}{d_1}}-D_1\right)$$

(11)

$D_1$ is the diameter of the actual transmission line, $d_1$ is the number of pixels of the actual line image transmission line, $d_2$ is the number of pixels of the transmission line in the ice-covered line image, and $h$ is the ice thickness. By utilizing the proportional relationship between the actual size of the image and the pixel area, the proportion of pixels in the image can be converted into the actual ice thickness.

3.4. Model Deployment

In order to realize remote and autonomous identification of the thickness of ice-covered transmission lines, it is necessary to configure edge computing devices inside the deicing robot. Therefore, this article selects OrangePi5 Plus (Shenzhen Xunlong Software Co., Ltd., Shenzhen, China) as the deployment end device of EDPNet. The CPU of OrangePi5 Plus is RK3588, which integrates the neural network computing unit NPU with a computing power of up to 6 Tops. EDPNet is deployed on the device to implement the edge-side reasoning function. Figure 5 is a flowchart of model deployment.

Since EDPNet is trained on the PyTorch framework, the final trained model format is a PyTorch model and cannot be directly run on edge devices. The NPU inside RK3588 can only support the RKNN model for accelerated reasoning, so the PyTorch model needs to be converted to the ONNX model and then converted to the RKNN model.

4. Experimental Results and Analysis

The ice-covered transmission line dataset used in this experiment consists of two parts. One part comes from the ice-covered transmission line images collected by ourselves in winter, and the other part is a dataset collected on the Internet. Among these, the self-collected dataset uses the transmission lines and cables built outdoors when snow falls in winter to restore the actual ice-covered transmission lines. When the cable is ice-covered, a camera is used to take pictures and collect ice-covered images. The other part is to search for ice-covered transmission line images on the Internet, download them, and manually screen them to remove poor quality images. Finally, after sorting, 939 images of ice on transmission lines were obtained and labeled using the Labelme annotation tool to obtain authentic label images. During training, the dataset is expanded through online data augmentation methods and divided into training and testing sets in an 8:2 ratio.

The experiment uses the PyTorch deep learning framework for model training, the GPU used for training is an RTX4090, and the model inference device is an OrangePi5 Plus. During training, the BatchSize is set to 24, the number of iterations is 1600, and the Stochastic Gradient Descent (SGD) optimizer is selected for gradient optimization. The evaluation indicators of the final segmentation effect are mean intersection over union (mIoU) and balanced F-Score (F1-Score). IoU represents the intersection-over-union ratio of the real mask to the predicted mask. mIoU averages the IoUs of different categories to get the average intersection-over-union ratio. The calculation formula is

$$mIoU={\displaystyle \frac{1}{K}}\mathsf{\Sigma }{\displaystyle \frac{TP}{TP+FP+FN}}$$

(12)

where K represents the classification category. TP means that the predicted and true labels are consistent and both are true. FN means that the predicted result is wrong; that is, the predicted value is false and the true label is true. FP means that the true label is false and the predicted value is true. F1-Score is the harmonic mean of precision and recall, which is used to comprehensively evaluate the performance of the model. The calculation formula is

$$F1={\displaystyle \frac{Precision\times Recall\times 2}{Precision+Recall}}$$

(13)

4.1. Ablation Experiment

In order to verify the effectiveness of the module proposed in this paper, a model ablation experiment is conducted. In this section, four groups of control experiments are set up to compare the segmentation accuracy and detection speed of different groups. Experiment 1 is based on the basic U-Net. Experiment 2 replaces the encoder convolution block with the LMRC block, and the decoder remains unchanged. Experiment 3 keeps the encoder unchanged and adds a lightweight deep fusion module to the decoder. In experiment 4, the encoder and decoder are replaced by LMRC blocks and lightweight deep fusion modules, respectively, but the loss function does not use the dynamic loss function. Finally, it is compared with the EDPNet proposed in this paper. The experimental results are shown in

Table 1. Model ablation experiment results.

Method	mIoU/%	F1-Score/%	Parameter/M	Time/ms
Exp1	90.6	95.0	4.32	564
Exp2	89.9	94.6	0.59	94
Exp3	91.3	95.4	1.09	330
Exp4	92.7	96.1	0.60	96
EDPNet	93.4	96.5	0.60	96

.

As can be seen from the table, in Experiment 2, the encoder’s convolutional block was redesigned into an LMRC block, and the number of convolutional channels in each layer was halved, which greatly reduced the number of parameters and increased the inference speed of a single image by six times. However, the segmentation accuracy mIoU and F1-Score were both reduced. The results of Experiment 3 prove that the decoder’s lightweight deep fusion module effectively compensates for the loss of accuracy caused by halving the convolution channel, and the segmentation accuracy has increased. Experiment 4 tests the LMRC block and the LDFM; the accuracy mIoU is improved by 2.1 percentage points compared with the basic network, and the F1-Score is improved by 1.1%. In addition, we also calculated the 95% confidence intervals and standard deviations of mIoU and F1- Score for EDPNet. The standard deviation of mIoU is 0.38, with confidence intervals of (92.49, 93.42). The standard deviation of F1-Score is 0.19, and the confidence intervals are (96.06, 96.53). This further proves the stability and reliability of the model. Finally, the EDPNet proposed in this paper achieved the highest segmentation accuracy, and the inference speed was further improved to 96 ms. Compared with Experiment 4, it proves that the dynamic loss function can refine the ice segmentation image.

4.2. Dilation Rate Combination Experiment

In the encoder, the LMRC convolution block splits the ordinary convolution into four branches with the same number of channels, three of which use different dilation rates to enhance the asymmetric convolution receptive field. The combination of different dilation rates directly affects the extraction of ice-covered transmission line image features. Three groups of different combinations of dilation rates were selected for experiments. The first group is (2, 3, 5), the second group is (2, 5, 8), and the third group is (5, 8, 11). The experimental results are shown in Figure 6.

As shown in Figure 6, the combination (2, 3, 5) consists of convolutions with small dilation rates, and its receptive field is limited, resulting in a lack of attention to large-scale features. Although the combination (5, 8, 11) has a large dilation rate, it also loses attention to smaller features, resulting in the lowest segmentation accuracy among the three groups. When the dilation rate combination is (2, 5, 8), the mIoU and F1-Score are the highest. This combination takes into account both large and small receptive fields, so it can dynamically perceive different ice-covered transmission line characteristics.

4.3. Loss Stage Experiment

In the dynamic loss function, the loss function is divided into two stages. In the latter stage, a boundary refinement loss function is proposed to refine the ice segmentation output. This section of the experiment explores where to start the boundary refinement training. The experiment is divided into the first 25% of the training, 50% of the training, and 75% of the training stage. These percentage points represent different training stages and can observe the performance of the model at different levels of maturity. Thus, 25% represents the early training stage, 50% represents the mid stage, and 75% is close to training completion. We switch losses at different stages to explore the optimal timing for fine-grained segmentation output. The experimental results are shown in Figure 7.

From the analysis of Figure 7, we can see that when the next stage of refined training is started in the first 25% of the training, the mIoU value of the blue curve grows significantly faster, proving that the boundary refinement loss function can accelerate the convergence of the model. However, since the ice-covered image generated at this time is not accurate enough, the refinement effect is not good. In the 75% stage of training, its mIoU and F1-Score are not the best choices, and the benefit of the boundary refinement loss function is not great. Therefore, this paper chooses to change the loss function halfway through the training. As shown in Figure 7, when the training reaches 50%, the mIoU value of the purple curve has increased to a certain extent.

4.4. Comparative Experiment with Other Algorithms

This paper designs lightweight networks for edge devices based on U-Net. Therefore, this section compares it with other lightweight algorithms, including using common lightweight networks such as MobileNet and ShuffleNet to replace the encoder backbone, and using neural networks such as ENet that are specifically designed for image segmentation. In addition, a comparison was made between the non-lightweight segmentation model U-Net++ and TransBTS. UNet++ represents a pure convolutional neural network, whereas TransBTS belongs to the Transformer series of architectures. The experimental results are shown in

Table 2. Comparison results with other algorithms.

Method	mIoU/%	F1-Score/%	Precision/%	Recall/%	Parameter/M	GFLOPS	Time/ms
MobileNet	90.2	94.7	95.1	94.3	0.82	7.85	256
ShuffleNet	89.9	94.5	94.0	95.0	0.66	7.41	250
SqueezeNet	87.5	93.1	93.5	92.6	0.58	7.27	125
ENet	88.5	93.7	93.2	94.3	0.36	2.35	90
DABNet	91.7	95.6	95.8	95.4	0.75	7.94	176
U-Net	90.6	95.0	94.6	95.3	4.32	40.37	564
U-Net++	94.2	96.8	97.1	96.5	11.48	142.85	1519
TransBTS	94.7	97.2	97.5	96.9	13.73	66.54	1672
EDPNet	93.4	96.5	96.7	96.2	0.60	6.50	96

:

As can be seen from the above figure, ENet has the fastest detection speed and the smallest model size, but its detection accuracy is very poor. EDPNet not only leads other lightweight models in segmentation accuracy, but also has a much faster detection speed. Compared to non-lightweight segmentation networks, EDPNet lags behind by 1% in segmentation accuracy, but its detection speed significantly outpaces that of other such networks. Figure 8 is a schematic diagram of segmentation comparison of different algorithms.

As shown in Figure 8, the left side is a comparison of the self-made dataset, and the right side is a comparison of the network collected images. As can be seen from the figure, lightweight networks such as MobileNet and ENet can identify the ice-covered area, but the details are not processed finely enough, and a small number of false detections and missed detections occur. EDPNet performs well on the self-made data set and can perfectly identify the ice-covered area, but in the ice-covered images of network transmission lines with more complex backgrounds, there are also some areas that are misidentified. Overall, its recognition effect is better than other lightweight networks.

4.5. Energy Consumption Experiment

To verify the inference energy consumption of the EDPNet model, energy consumption tests were conducted on the OrangePi5 Plus device. Different CPU frequencies were locked separately to obtain runtime power and inference time. The experimental results are shown in the

Table 3. Energy consumption experiment results.

CPU	Frequency	Consumption/w	Time/ms
RK3588	2.3 GHz	4.25	91
	1.8 GHz	3.92	93
	1.0 GHz	3.20	96
	0.4 GHz	2.75	101

.

It can be seen from the above table that the power of EDPNet in operation is very low, only less than 4 w, so it is enough to meet the needs of edge computing. In addition, we conducted multiple high- and low-temperature tests on the OrangePi5 Plus to simulate real-world usage scenarios. The device was able to operate normally in extreme environments ranging from −40 °C to 55 °C.

4.6. Optimize Experiment

In addition to designing lighter network structures, quantitative methods can also be used to improve model recognition speed. Model quantization converts floating-point parameters in the model from FLOAT16 to INT8 type, which can reduce memory usage and improve inference speed, but this method can cause a decrease in accuracy.

Table 4. Quantitative experimental results.

Performance Parameter	Before Quantification	After Quantification
mIoU/%	93.4	82.9
F1-Score/%	96.5	88.4
Time/ms	96	80

shows the performance comparison before and after quantization. It can be seen from the table that the inference speed of the quantized model has been improved, but its detection accuracy has significantly decreased, with a loss of about 10%.

4.7. Ice Thickness Comparison Experiment

This section conducts a comparative experiment on the calculation of ice thickness on transmission lines and verifies the advantages and disadvantages of the thickness identification algorithm by comparing DABNet and EDPNet. The experimental results are shown in the

Table 5. Comparison results of ice thickness.

Actual Thickness/mm	Test Method	Predicted Thickness/mm	Error/%
4.52	DABNet	4.29	5.0
	EDPNet	4.33	4.2
7.48	DABNet	7.64	2.1
	EDPNet	7.62	1.9
12.54	DABNet	12.28	2.0
	EDPNet	12.37	1.3

.

As can be seen from the table, the recognition error of EDPNet is below 5%, with the highest being 4.2%. The thickness recognition error of DABNet is higher than that of EDPNet. The experimental results prove that EDPNet can accurately identify the ice thickness on transmission lines, and its predicted thickness is not much different from the actual thickness.

5. Conclusions

Aiming at the task of recognizing the ice thickness on transmission lines on the end-side equipment, this paper proposes EDPNet, a network for recognizing ice thickness on transmission lines based on efficient dynamic perception. The LMRC lightweight convolution block is first proposed to accelerate the encoder feature extraction speed. Then, a lightweight deep fusion module is added to the decoder to improve the accuracy of ice segmentation of transmission lines with minimal parameters. Then, a dynamic loss function is proposed, and in the second half of model training, the output of ice segmentation images is optimized using the boundary refinement loss function. Experiments were conducted on the transmission line ice dataset, and the segmentation accuracy mIoU reached 93.4% and the F1-Score was 96.5%. The model was deployed on OrangePi5 Plus; the number of model parameters was reduced by six times. The experimental results show that the EDPNet proposed in this paper can complete the task of identifying the ice thickness on transmission lines on edge devices. This is of great significance for the automation operation and maintenance of the power system, improving the efficiency of monitoring the status of transmission lines and preventing the harm caused by ice cover on transmission lines.

Since image segmentation is a computationally intensive task in computer vision, its recognition speed on edge devices is not fast enough. Due to extreme weather conditions, it is difficult to obtain the thickness of ice cover using visual methods, which can lead to image blurring and inaccurate recognition. Therefore, the next step is to reduce the number of parameters and computational complexity of the model through pruning, optimizing network models, and knowledge distillation techniques, thereby reducing the model’s demand for computing resources and improving detection speed on edge devices. On the other hand, attempts are being made to explore multimodal learning that combines visual and other meteorological data to provide more accurate recognition of ice cover thickness.