Research on Detection of Icing Cover Transmission Lines Under Different Weather Conditions Based on Wide-Field Dynamic Convolutional Network LDKA-NET

Abstract

The safety of transmission lines is a crucial guarantee for the operation of the power grid. To address the issue of low detection accuracy for icing transmission line defects with existing models, this paper proposes a defect detection algorithm for icing transmission line defects under different weather conditions based on a Large Dynamic Kernel Aggregation Net (LDKA-NET). First, a wide field of view convolutional network (WFVC Net) is introduced to enhance the network’s perception and generalization capabilities, enabling better adaptation to complex scene targets. Secondly, a full-dimensional dynamic convolutional feature fusion network is proposed, which strengthens the model’s feature extraction ability by learning linear combinations of multiple convolution kernels and their weighted input-related attention. Finally, an Expectation Maximization Dynamic Convolutional Attention (EM-DCA) mechanism is introduced, which focuses on and utilizes important information in the input data to help the model better allocate attention, thereby improving its generalization and robustness. Experimental results show that on the dataset proposed in this paper, the average accuracy of the improved algorithm ([email protected]) reaches 99.01%. The model parameters decreased by 47.7 M compared with the baseline model and 91.26 M compared with the Faster region with CNN feature (Faster R-CNN) model. The accuracy is 4.81% and 3.11% higher than the SSD model and YOLOv5-L model, respectively. Compared with the existing models, our model is smaller and has higher detection accuracy. It can accurately detect ice-covered wires and complete the task of the ice-covered detection of transmission lines.

1. Introduction

With the increasing demand and dependence on electricity in modern society, the safe and stable operation of transmission lines, as important conduits for power transmission, has become particularly crucial [1]. However, a large number of insulators inevitably pass through regions with complex environments such as high altitudes and recurring icing. Moreover, being exposed in the wild year-round, combined with the deteriorating climate conditions in recent years, there are frequent occurrences of haze, rainy weather, and frozen snow, making the weather and geographical conditions affecting the insulators increasingly complex. This complexity easily leads to incidents of icing, damage, flashover, etc., posing severe safety risks to the operation of the power system and causing significant economic losses to society [2].

In view of this, to promptly detect and resolve issues with insulators and improve the safety and reliability of transmission lines, insulator monitoring technology has gradually become one of the focal points in the electricity sector. By monitoring the status and performance of insulators in real time, abnormal conditions can be detected in a timely manner, allowing for appropriate maintenance and replacement measures to ensure the safe operation of transmission lines [3].

Currently, insulator defect detection methods include various technical means such as ultrasonic, infrared thermography, vibration, partial discharge, and visual inspection. Ultrasonic detection requires specialized equipment and personnel, making it costly; infrared thermography is significantly influenced by environmental factors and is insensitive to certain defects, making it unable to directly detect internal problems of insulators; vibration detection is heavily affected by external interference, resulting in limited accuracy; and partial discharge detection requires complex equipment and skilled personnel, which also incurs high costs [4]. In contrast, visual inspection is simple and intuitive, allowing for the quick identification of obvious surface defects in insulators.

Current visual insulator defect detection methods can be mainly categorized into two types: those based on machine learning image processing algorithms and those based on deep learning algorithms [5]. Machine learning image processing algorithms typically involve feature extraction and classification of images. First, image data are collected from the surface of the insulator, followed by the use of techniques such as edge detection, color analysis, and shape matching to extract feature information from the images. Finally, extracted features are classified and identified using methods like support vector machines and random forests to determine whether defects exist in the insulator.

On the other hand, deep learning algorithms utilize deep neural networks for end-to-end learning and processing of raw images. By constructing models such as deep convolutional neural networks (CNNs) and recurrent neural networks (RNNs), features and patterns are learned directly from image data, enabling automatic detection and identification of the insulator [6].

Reference [7] improved the YOLOv7 model by incorporating a coordinate attention mechanism into the multi-scale feature fusion process while also designing a C3GhostNetV2 module to capture long-range dependencies between different spatial pixels and reduce the model’s parameters and floating-point operations. However, this algorithm does not account for the insufficient capture of local details caused by small receptive fields, which might lead to the neglect of global context information. Reference [8] proposed a defect detection algorithm called FasterNet-YOLOv5, which combines FasterNet-tiny and the YOLOv5 model. By designing the DFC-FasterNet module in the backbone feature extraction network, the receptive field is increased during the feature extraction process to enhance the network’s detection accuracy, and the Neck module is redesigned for more accurate localization of insulator defect areas. However, the depth of this network model remains shallow, limiting its ability to extract features from images with particularly complex environmental backgrounds or occluded targets. Reference [9] proposes a hybrid model based on wavelet transform (WT), and Bat algorithm optimized Extreme Learning Machine (BA-ELM) was proposed in the literature. The noise reduction of ice cover thickness and meteorological factor data was carried out through wavelet transform, and segmentation processing was carried out to improve the prediction accuracy. Experimental results show that the accuracy of the model in ice thickness prediction is significantly improved. However, its limitations include a reliance on high-quality initial data, high computational complexity of the model, limited generalization ability, and insufficient ability to predict long-term dynamics. Reference [10] proposes a weakly supervised and phased transfer learning method based on YOLOv5 to identify insulators and different ice types, and improve recognition accuracy and anti-interference ability through multi-dimensional feature fusion. The insulator icing dataset was constructed based on the monitoring data of China Southern Power Grid from 2014 to 2018. The experimental results show that the model recognition accuracy is 86.6%, the recall rate is 91.3%, the speed is 8 ms/image, and the labeling efficiency is significantly improved by pseudo labeling. However, limitations include limited dataset time, inadequate adaptability to extreme environments, and the dependence of labeling intelligence on initial data quality. Reference [11] proposes a segmentation method based on non-subsampled contour wave transform (NSCT) and two-dimensional OSTU is proposed. The NSCT is used to denoise the image, and the improved genetic algorithm and two-dimensional OSTU algorithm are combined to achieve image segmentation. Experiments show that this method can effectively remove the noise of an ice-covered image, accurately extract the target, and reduce the search time, which is suitable for real-time application. However, its limitation is that the segmentation effect in a complex background needs to be verified, the computational complexity of the algorithm may limit its application in resource-constrained devices, and its adaptability to different types of noise needs to be further improved. Lee et al. [12], in order to avoid the impact of black ice on traffic accidents, used the Mask-R-CNN model to detect black ice on the ground, compared different pre-trained models and architectures, analyzed the detected enveloping box data, and determined the degree of dangerous areas based on the number of divided pixels, which was verified by experiments. The detection accuracy reached 92.5%, which played a role in alerting passengers to black ice. However, the dataset of this study failed to fully reflect the diversity of the actual traffic environment, which may affect the generalization ability of the model.

However, the limitation is that the time span of the dataset is limited, which may affect the generalization ability of the model, the robustness in complex environments needs to be verified, and the intelligent labeling is strongly dependent on the quality of the initial data.

In conclusion, this paper proposes an ice-covered insulator defect identification method based on a wide-area visual dynamic convolutional network (LDKA-NET), which solves the problem of low target detection accuracy in complex environments. The main contributions are as follows:

First, it presents a Wide-Domain Vision Convolution Network (a Wide Field View Convolutional Network), which employs larger convolution kernels to enhance the network’s perception and generalization capabilities, making it better suited for the complexities of transmission line scenarios. Additionally, a large kernel CNN can achieve a larger effective receptive field with fewer layers while also providing higher shape variance, enabling better capture of shape information within images.

Secondly, a multi-dimensional dynamic convolution feature fusion network is proposed. Compared with traditional static convolutions, this method significantly improves CNN accuracy by learning a linear combination of multiple convolution kernels and weighted attention related to their input, demonstrating higher parameter efficiency and feature extraction capability.

Finally, the Expectation Maximization Dynamic Convolutional Attention (EM-DCA) mechanism is adopted to focus on and utilize important information within the input data, helping the model better allocate attention to essential targets or scene segments, enabling the learning of more generalized feature representations. This contributes to enhancing the model’s generalization ability and robustness, allowing for more accurate localization of target positions.

2. Transmission Line Ice Cover Monitoring System

The design for the process of collecting images of iced conductors is illustrated below. It is divided into three parts: the image acquisition system, the signal transmission system, and the control system. The main process is as follows: Direct current is supplied by solar panels to power the cameras installed on the transmission lines. These cameras are controlled by a main control board for capturing images. The images are then transmitted to the control center system via a 4G or 5G network and optical fiber. Finally, they are sent to a computer for post-processing. The main principle is shown in Figure 1:

2.1. LDKA-NET Algorithm Model

The detection of insulators on transmission lines presents several challenges: first, the background is complicated, as transmission lines are often located in remote and complex environments, making it difficult for existing visual detection algorithms to distinguish foreground from background [9]; and second, the target is small, as monitoring devices are often installed atop transmission towers, making it difficult to recognize the target from a distance due to its small size [10]. To address these difficulties, this paper makes targeted improvements to the baseline model, the overall structure of which is shown in Figure 2.

As shown in Figure 2, the algorithm presented in this paper consists of four parts: image preprocessing, Backbone, Neck (feature extraction network), and Head (prediction network), with the red boxes indicating the improved modules. The Wide-Domain Vision Convolution Network (WFVC Net) employs larger convolution kernels to enhance the network’s perception and generalization capabilities, making it better suited for the complexities of transmission line scenarios. Secondly, the multi-dimensional dynamic convolution significantly improves the accuracy of the CNN by learning linear combinations of multiple convolution kernels and weighted attention related to their input, demonstrating higher parameter efficiency and feature extraction capability. Finally, through the Expectation Maximization Dynamic Convolutional Attention (EM-DCA) mechanism, which aims to retain information on each channel while reducing computational overhead, some channels are reshaped into batch dimensions, and the channel dimensions are grouped into multiple sub-features, ensuring that spatial semantic features are evenly distributed within each feature group.

2.2. Wide-Domain Vision Convolution Backbone Network

Traditional CNNs can obtain a larger receptive field by stacking multiple small convolution kernels; however, as the number of layers increases, the computational complexity of the network also significantly rises, leading to the potential issues of gradient vanishing or explosion [11]. To address this issue, this paper proposes a Wide-Domain Vision Convolution Backbone Network that utilizes larger convolution kernels to enhance the network’s perception and generalization capabilities, making it better suited for the complex scenarios discussed in this paper. Its overall structure is shown in Figure 3.

As illustrated, the Wide-Domain Vision Convolution Backbone Network consists of four main stages, each including convolution layers such as Stem, Stage, and Transition. Among these, the depthwise Conv represents depthwise separable convolution, which consists of two parts: depthwise (DW) convolution and pointwise (PW) convolution. Each convolution kernel in DW convolution computes only the corresponding feature map, which may weaken the information exchange between pixel points of different feature layers, potentially leading to precision loss. PW convolution is a 1 × 1 convolution that outputs n channels, aimed at compensating for the insufficient feature interaction in DW convolution. Through pointwise multiplication and addition operations, PW convolution can combine information from different feature maps to enhance the interaction between features [12].

Assuming the input feature map size is $D_w\times D_w\times C$, the convolution kernel size is $D_k\times D_k\times C$, and the output feature map size is $D_w\times D_w\times C_{out}$. The parameter count for the standard convolution layer is given as follows:

(1)$P_{S\tan dard}=(D_k\times D_k\times C)\times C_{out}$

where C indicates the number of input feature channels, and $C_{out}$ signifies the number of output feature channels and convolution kernels.

Additionally, in depthwise separable convolution, $D_K$ represents the size of the convolution kernel; the depthwise convolution is responsible for filtering and has a size of $D_k\times D_k\times 1$, with a total of C kernels, applied to each input channel. The pointwise convolution is responsible for channel transformation, with a size of $1\times 1\times C$ and also has $C_{out}$ kernels, applied to the output feature maps of the depthwise convolution. The parameter count for depthwise separable convolution is given by the following formula.

(2)$\begin{array}{cc}\hfill P_{Depthwise}& =(D_k\times D_k\times 1)\times C+1\times 1\times C\times C_{\mathrm{out}}\hfill \\ \hfill & =D_k\times D_k\times C+C\times C_{out}\hfill \end{array}$

The ratio of depthwise convolution parameters to standard convolution parameters is:

(3)$\begin{array}{c}{\displaystyle \frac{P_{Dephwise}}{P_{S\tan dard}}}={\displaystyle \frac{D_k\times D_k\times C+C\times C_{out}}{(D_k\times D_k\times C)\times C_{out}}}\\ ={\displaystyle \frac{1}{C_{out}}}+{\displaystyle \frac{1}{D_k\times D_k}}\end{array}$

Based on this, adopting depthwise separable convolution can resolve the increase in parameter count and floating-point operations brought by using larger convolution kernels. The stem layer consists of a 3 × 3 convolution with a stride of 2, depthwise separable convolution, and a DW convolution with a stride of 2, which is used to upscale the input image and reduce its dimensions. Specifically, it is shown as follows:

(4)$w_{(h/4,w/4,c1)}^{\prime }=D_{s=2}^{3\times 3}\left[f_{s=1}^{1\times 1}(D_{s=2}^{3\times 3}(f_{s=2}^{3\times 3}(w_{h,w,c})))\right]$

As stated in the formula, where w represents feature weights and D denotes depthwise separation. Assuming the input dimensions are H and W, where H is the image height and W is the image width, the output feature size of the stem layer is ${\displaystyle \frac{H}{4}}\times {\displaystyle \frac{W}{4}}\times C1$, where C1 is the number of channels after dimensionality increase.

The Stage layer is composed of stacked RepLK Blocks and ConvFFNs. The RepLK Block includes normalization layers, 1 × 1 convolution, depthwise separable convolution, and crucial residual connections; ConvFFN replaces the fully connected layers with 1 × 1 convolution and employs inter-layer residual connections. The Transition layer mainly utilizes PW convolution and DW convolution with a stride of 2 for image down sampling.

Moreover, utilizing a large kernel backbone network can achieve a greater receptive field with fewer layers while also providing higher shape variance, as shown in the following formula:

(5)$R{F}_n=R{F}_{n-1}\times S_n+(K_n-1)$

The above equation is the receptive field calculation formula, where $R{F}_n$ is the receptive field size at layer n, $R{F}_{n-1}$ is the receptive field size at layer n − 1, $S_n$ is the stride of layer n, and $K_n$ is the kernel size at layer n. It is evident that there is a direct relationship between the convolution kernel size and the receptive field; larger convolution kernels can cover a broader range of the input image, allowing each neuron to perceive more extensive local information. This capability enables better capture of shape information in images, enhancing the model’s representational ability to better handle complex scene detection.

2.3. Multi-Dimensional Dynamic Convolution Feature Fusion Module

In the complex scenarios of transmission lines, the application of traditional static convolutions lacks flexibility, as they remain unchanged during the network training process. This rigidity prevents adjustments according to the varying feature distributions of different input data, leading to a decline in model adaptability and poorer performance in recognizing insulators in complex scenes [13]. Additionally, the baseline model’s use of cross-stage feature fusion strategies may cause feature information to become blurred or overlooked, negatively impacting the model’s performance. Furthermore, it requires learning the weights and biases for feature fusion, increasing the risk of overfitting. Therefore, this paper proposes a multi-dimensional dynamic convolution feature fusion module, as illustrated in Figure 4.

1. Multi-Dimensional Dynamic Convolution

Conventional convolutions use a single static kernel that is independent of the input samples. In contrast, dynamic convolution employs multiple convolution kernels with linear weighting, where the weighting values depend on the input. This input dependency gives dynamic convolution greater flexibility than traditional static convolutions. By learning linear combinations of multiple convolution kernels and input-related attention weights, dynamic convolution can significantly enhance the accuracy of CNNs, providing higher parameter efficiency and feature extraction capability.

As shown in Figure 4, the multi-dimensional dynamic convolution introduces a multi-dimensional attention mechanism using a parallel strategy to learn more flexible attention across the four dimensions of the convolution kernels. For each kernel, attention weights are assigned to the convolution parameters in the spatial domain, the convolution filters of the input channels, the convolution filters of different output channels, and the overall n convolution kernels. Gradually applying different attention weights along the dimensions of position, channel, filter, and kernel will allow the convolution operations to exhibit variability across all dimensions with respect to the input, enhancing the performance in capturing rich contextual information.

Standard convolutional layers have a static kernel that applies to all input samples. In contrast, the dynamic convolution layer employs dynamic weighting using a linear combination of n convolution kernels combined with an attention mechanism, making the convolution operation dependent on the input. The traditional dynamic convolution operation can be defined as follows:

(6)$x_{out}=x{\displaystyle \sum _{i=1}^n{a}_{wi}{W}_{\mathrm{i}}}$

In the above equation, X represents the input features of $(h,w,c_{in})$, and $x_{out}$ represents the output features of $(h,w,c_{out})$. $h$ represents the height of the input feature map, $w$ represents the width of the input feature map, $c_{in}$ represents the number of channels of the input picture, and $c_{out}$ represents the number of channels of the output picture. $W_i$ indicates the i-th output filter kernel, and $a_{wi}$ represents the attention coefficients. However, traditional dynamic convolution only involves convolution kernels and attention weight for the filter kernels. This method overlooks other adjustable parameters that affect convolution and multiplies the convolution parameters by n, resulting in inefficiency. The implementation of multi-dimensional dynamic convolution is described as follows:

(7)$x_{od\_out}=x{\displaystyle \sum _{i=1}^n(a_{wi}\odot a_{fi}\odot a_{ci}\odot a_{si}\odot W_i)}$

In this formula, $a_{wi}$, $a_{fi}$, $a_{ci}$, and $a_{si}$ are the attention coefficients for the spatial, channel, filter, and kernel dimensions, respectively. The four types of attention weights ensure that the convolution operation considers every spatial position for different inputs, all input channels, all filters, and all kernels, thereby providing performance guarantees for capturing rich contextual cues, and where $\odot$ stands for the Hadamard product, which is an element-by-element multiplication operation applicable to vector or matrix operations.

2. Full-Dimensional Feature Fusion ODC2F

In the baseline model, while the C2F module plays a key role in cross-stage feature fusion, it may also blur or overlook some important feature information during this process, negatively impacting the model’s performance. Moreover, it requires the learning of weights and biases for feature fusion, introducing a substantial number of additional parameters, which may lead to risks of overfitting. The formulas for this are as follows:

(8)$F_{out}=f_{s=1}^{1\times 1}(M_{\theta }+\omega F_{in})$

(9)$M_{\theta }=f_{s=1}^{1\times 1}\left\{\lambda \left[\left(f_{s=1}^{3\times 3}(f_{s=1}^{3\times 3}(F_{in}))\right)+\tau F_{in}\right]\right\}$

In these equations, $F_{in}$ is the input feature, and $F_{out}$ is the output feature. When Bottleneck is set to True, $\tau$ is 1; otherwise, it is 0. $\lambda$ represents the number of feature fusion modules, and is the residual weight bias coefficient. Since the information exchange between feature maps is relatively limited, it can result in lower accuracy in object detection. The formula for ODC2F is as follows:

(10)$F_{out}^{″}=f_{s=1}^{1\times 1}(M_{\theta }^{″}+\omega F_{in})$

(11)$M_{\theta }^{″}=f_{s=1}^{1\times 1}\left\{\lambda \left[\left({\sigma }_{s=1}^{3\times 3}({\sigma }_{s=1}^{3\times 3}(F_{in}))\right)+\tau F_{in}\right]\right\}$

In these equations, $F_{out}^{″}$ is the output feature, and ${\sigma }_{s=1}^{3\times 3}$ is the previously described full-dimensional dynamic convolution. By employing full-dimensional dynamic convolution, attention features are learned from the four dimensions of the convolution kernel space in parallel, leading to better performance in capturing rich contextual information.

2.4. Expectation Maximization Dynamic Convolutional Attention Mechanism (EM-DCA)

The scenes of transmission lines are complex, and insulators are often located in areas with color confusion, which poses a significant challenge for the network in distinguishing between foreground and background. The baseline model has not been optimized for such scenarios, resulting in suboptimal performance; the model struggles to adapt to different scenes or environments, making it difficult to accurately identify and localize insulator targets [14]. Furthermore, the baseline model’s features have certain limitations during the information fusion process, as they cannot effectively focus on critical areas or objects. The traditional attention mechanism often employs channel reduction to model cross-channel relationships, which makes it challenging for the network to retain information in each channel, leading to poor model performance and difficulty in achieving the expected results.

In light of this, this paper adopts the Expectation Maximization Attention Mechanism to focus on and utilize important information within the input data. This approach helps the model allocate attention more effectively, concentrating on essential targets or scene segments and learning more generalized feature representations. This, in turn, enhances the model’s generalization ability and robustness.

As shown in the Figure 5, let the input feature be $F\in R^{C\times H\times W}$, where C represents the number of input channels, and H and W represent the height and width of the input features, respectively. The feature F is divided into G sub-features along the channel dimension, where G << C. Thus, F can be expressed as:

(12)$F=\left[F_0,F_i,···,F_{G-1}\right], F_i\in R^{C//G\times H\times W}$

The EM-DCA attention mechanism consists of four branches, with three parallel paths extracting attention weight descriptions for the grouped feature maps. There are two 1 × 1 branches and one 3 × 3 branch. Two global average pooling operations are used in the 1 × 1 branches to encode the channels along two spatial directions, while the 3 × 3 branch includes a multi-dimensional dynamic convolution to capture multi-scale feature representations.

Two parallel 1D feature encoding vectors share a 1 × 1 convolution, allowing the model to capture local cross-channel interactions. The formula for encoding global information along the channel C at a height H is given by:

(13)$P_{\mathrm{c}}^H(H)={\displaystyle \frac{1}{W}}{\displaystyle \sum _{0\le i\le W}{x}_c(H,i)}$

The formula for the pooling output along channel (C) at width (W) is:

(14)$P_{\mathrm{c}}^W(W)={\displaystyle \frac{1}{H}}{\displaystyle \sum _{0\le j\le H}{x}_c(j,W)}$

The 3 × 3 branch captures local cross-channel interactions through convolution to expand the feature space. It not only encodes information between channels to adjust the importance of different channels but also retains precise spatial structural information within the channels.

The cross-spatial learning establishes a dependency relationship between channels and space, achieving richer feature aggregation. Using two-dimensional global average pooling, the output in the 1 × 1 branch encodes global spatial information, and the output of the minimum branch is transformed to the corresponding dimension shape:

(15)$P_c={\displaystyle \frac{1}{H\times W}}{\displaystyle \sum _j^H{\displaystyle \sum _i^W(i,j)}}$

Finally, the outputs of each branch are computed as attention weight values, combined with global semantic information to capture key regions. This module helps the model to more effectively capture critical information in the input data and dynamically adjust the attention distribution based on the importance of different features. At the same time, it aims to retain information on each channel and reduce computational overhead by reshaping some channels to batch dimensions and grouping channel dimensions into multiple sub-features, ensuring spatial semantic features are evenly distributed within each feature group.

2.5. Dynamic Non-Monotonic Focusing Mechanism Bounding Box Loss

The classification loss for the baseline model uses binary cross-entropy loss, while the bounding box regression employs DFL loss and CIOU loss, transforming the regression problem into a classification problem. Although the convergence value is not as good as other models, this study adopts the dynamic non-monotonic focusing mechanism bounding box loss to replace the DFL loss and CIOU loss of the baseline model, as the LDKA-NET maintains a large model complexity while keeping the parameter count low. The dynamic non-monotonic focusing mechanism utilizes “outlierness” instead of IoU to evaluate the quality of anchor boxes and provides a sensible gradient gain allocation strategy. While it reduces the competitiveness of high-quality anchor boxes, it also minimizes the harmful gradient produced by low-quality examples. This allows the model to focus on anchor boxes of ordinary quality, improving the overall performance of the detector.

Since the training data inevitably contain low-quality instances, we constructed a distance attention mechanism with two layers of attention: WIoU.

(16)$L_{WIoUv1}=R_{WIoU}{L}_{\begin{array}{l}IoU\\ \end{array}}$

(17)$R_{WIoU}=\exp ({\displaystyle \frac{{(x-x_{gt})}^2+{(y-y_{gt})}^2}{{(W_g^2+H_{\mathrm{g}}^2)}^{\ast }}})$

where $W_g$, $H_g$ is the size of the minimum enclosing box, and to prevent $R_{WIoU}$ from generating gradients that hinder convergence, ∗ denotes the separation of $W_g, H_g$ from the computation, as shown in the Figure 6 below:

This loss function can better capture the complex features in the data rather than overfitting to the noise in the training data, resulting in optimal model performance and good generalization ability.

3. Experimental Results

3.1. Experimental Environment and Performance Evaluation

In order to better prove the function of the module proposed in Section 2, we adopt the following experiments to verify the practicability of our proposed LDKA-NET network. The experiments were conducted using the PyTorch1.8.1 [15] framework. In the experiments, the batch size was set to eight, the momentum was set to 0.9, the weight decay was set to 0.0005, the learning rate was set to 0.001, and the optimizer used was SGD. The experiments utilized an NVIDIA GeForce RTX 3090 GPU, an Intel i9-11900k processor (NVIDIA Corporation, Santa Clara, CA, USA), and the operating system was Ubuntu 18.04.

To objectively evaluate the advantages of the proposed algorithm, mean average precision (mAP), recall (R), and precision (P) were selected as evaluation metrics [16].

3.2. Dataset Creation

The insulator dataset used in this study consists of two parts: one part includes 816 images of ice-covered transmission lines obtained from the internet, while the other part is generated through data augmentation techniques such as rotation, mirroring, and flipping, resulting in a total of 2448 images. In total, the combined dataset consists of 3264 images. The dataset includes a dataset of real transmission line icing from Qinghai Province, China, covering a total length of about 30 km and monitored through approximately 100 monitoring points placed along the network. These monitoring points are selected to ensure comprehensive coverage of critical areas and to reflect the typical environmental conditions that the transmission lines face, such as high-altitude regions prone to severe icing. The dataset was annotated using the LabelImg software for two types of targets and was randomly divided into training, validation, and test sets in an 8:1:1 ratio.

3.3. Ablation Experiments

To assess the impact of the proposed improvements on the performance of the original algorithm, we integrated the improved mechanisms into the baseline algorithm one by one and performed detailed experimental comparisons on each integrated algorithm configuration using the same test set. This aims to comprehensively understand the contribution of the proposed mechanisms to the overall performance of the algorithm. The experimental results are detailed in Table 1.

From the table, it can be observed that the performance of the original algorithm improves significantly when three different modules are added. The wide field of view convolutional backbone network (WFVC Net) employs larger convolutional kernels to enhance the network’s perception and generalization capabilities. Larger convolutional kernels can cover a broader range of the input image, allowing each neuron to sense a wider array of local information. This enables the model to better capture shape information in images, enhancing its representational ability, and making it more effective in detecting complex scenes. Compared to the baseline model, the missing detection rate has significantly improved, with both recall and mAP values showing enhancements.

The ODC2F module for full-dimensional feature fusion learns attention features across the four dimensions of the convolution kernels through a parallel strategy, providing better performance in capturing rich contextual information, leading to a qualitative improvement in the model’s evaluation metrics. Furthermore, the Expectation Maximization Dynamic Convolutional Attention (EM-DCA) mechanism helps the model to more effectively capture critical information in the input data while suppressing irrelevant complex background information. Compared to the original algorithm, the improvements are evident, with the model’s mAP increasing by 2.78% and the recall value increasing by 3.26%.

From the table data, it is clear that integrating the three different modules into the original algorithm significantly enhances performance. The wide field of view convolutional backbone network strengthens the model’s representational ability by employing larger convolutional kernels, allowing each neuron to sense a broader array of local information. Compared to the baseline model, this approach notably reduces the missing detection rate and achieves improvements in both recall and mean average precision (mAP).

Additionally, the ODC2F module for full-dimensional feature fusion learns across the four dimensions of the convolution kernels using a parallel strategy, enabling the model to more effectively capture rich contextual information, with a 1.87% increase in recall. The EM-DCA mechanism further boosts model performance by more effectively identifying critical information in the input data and suppressing irrelevant complex backgrounds. Compared to the original algorithm, the model integrated with these improvements shows an increase of 2.78% in mAP and 3.26% in recall. This fully demonstrates that enhancing the model’s perception capability and optimizing the information capturing mechanism can significantly improve object detection performance in complex scenes.

The defect detection results for ice-covered insulators in different scenarios using the baseline model and the proposed algorithm are shown in Figure 7. As depicted in the figure, the improved algorithm clearly outperforms the baseline model across various scenes.

3.4. Comparative Experiments

3.4.1. Data Analysis

To compare the performance of the proposed algorithm LDKA-NET with the baseline model and various state-of-the-art (SOTA) algorithms, this experiment trained each different type of SOTA algorithm using the same training set. The algorithms included the two-stage detection network Faster R-CNN [16], the single-stage detection network SSD [17], YOLOv5 [18], the anchor-free single-stage network FCOS [19], and the end-to-end detection network DETR [20,21]. The number of training iterations was uniformly set to 200, and the same test set and hyperparameters were used for validation. The results are shown in Table 2.

The experimental data demonstrate the differences in performance metrics among various object detection algorithms. The baseline model, which uses CSPDarkNet as its backbone, performs exceptionally well in terms of precision and mean average precision ([email protected]), achieving 98.66% and 96.23%, respectively. However, its recall rate is somewhat lacking at only 91.87%. In comparison, while YOLOv5-L [22,23] also employs a similar backbone, it falls slightly short in precision and [email protected], with values of 93.61% and 95.90%, respectively.

Faster R-CNN and FCOS use ResNet50 as their backbone network. However, Faster R-CNN, due to its two-stage feature extraction mechanism, results in a larger computational load and relatively lower precision, achieving only 63.79%. In contrast, FCOS [24] implements object detection through per-pixel prediction, avoiding all computations and hyperparameters associated with anchor boxes, demonstrating excellent detection performance with an [email protected] of 96.19%. Nevertheless, it still lags behind the baseline model and the proposed algorithm, indicating significant room for improvement.

The SSD model, which employs VGG as its backbone, shows good performance in both precision and recall but has a relatively low [email protected] of only 94.02%. This indicates that while the SSD model has certain advantages, its performance is constrained when handling complex detection tasks. Additionally, the DETR [25] network generally performs poorly across all metrics, with an [email protected] of only 71.11%, indicating substantial room for improvement.

The proposed algorithm utilizes WFVC Net as its backbone and demonstrates high performance in precision, recall, and [email protected], achieving 99.42%, 95.13%, and 99.01%, respectively. Compared to other algorithms, the proposed algorithm also has relatively fewer parameters and lower computational demands, showcasing its ability to maintain a lightweight model while still achieving outstanding detection results. This indicates strong applicability and efficiency in practical applications.

3.4.2. Feature Map Analysis

To further validate the principles behind feature extraction in the proposed algorithm, feature maps generated at each stage were visualized and compared with the baseline model, as shown in Figure 8.

In the figure, the feature maps of the baseline model and the proposed algorithm are converted into visual images for comparative analysis of their differences. The feature map in Figure 8b displays more pronounced structures and patterns, with the representation of the target insulator being more prominent and concentrated. This helps to more accurately locate and highlight key information in the image, showing superior performance compared to Figure 8a. In contrast, Figure 8a performs poorly in focusing on the area of the ice-covered insulator.

Furthermore, Figure 8b exhibits higher activation intensity in the mid-to-late stages of the backbone network, indicating that its feature extractor is more effective at capturing the local details of the image. Higher activation intensity is typically associated with better discrimination capability of the network, allowing it to learn more complex feature representations through deeper convolutional layers.

Additionally, compared to the baseline model, the proposed algorithm, which utilizes a more optimized wide field of view convolutional network, enables the feature map to not only capture more critical information but also achieve better focus and analysis of complex scenes in terms of spatial distribution of features. Moreover, observing stages 12–19, the proposed algorithm enhances the model’s capability in processing local details effectively through the use of full-dimensional dynamic convolution and Expectation Maximization Dynamic Convolutional Attention. In contrast, the baseline model, lacking targeted optimization for the scene, shows poorer feature clarity and cannot effectively discern the areas of focus for the target.

In summary, the advantages of the feature map in Figure 8b in terms of structural clarity, activation effect, and stability are directly related to the model’s performance in practical tasks, especially evident in the challenging detection task of ice-covered insulators.

3.4.3. Loss Function Analysis

To clearly describe the trend of loss function changes during the model’s iteration process, the training and validation loss value curves for the baseline model and the aforementioned SOTA algorithms have been plotted, with a total of 200 iterations, as shown in the following Figure 9.

From the figure, it can be observed that the loss curves for all models converge and trend towards smoothness. The YOLOv5 algorithm exhibits the smallest convergent values for both training and validation losses, utilizing binary cross-entropy for confidence and classification losses while employing CIOU loss for localization loss, resulting in higher convergence accuracy. However, considering its relatively low model accuracy, it may be overfitting the training and validation sets, leading to poor performance on the test set. Moreover, due to the simplicity of the model, with a floating-point operation count of only 109.6G, it struggles to effectively capture the key features of the data.

In contrast, the baseline model and the proposed algorithm use binary cross-entropy loss for classification while employing DFL loss and CIOU loss for bounding box regression, transforming the regression problem into a classification problem. Although the convergent values are not as good as those of other models, the proposed LDKA-NET maintains a small parameter count while possessing a greater model complexity, enabling it to better capture the complex features within the data without overfitting to the noise present in the training data. This results in optimal model performance and better generalization capability.

Both the baseline model and the algorithm in this paper use binary cross-entropy loss for classification, while bounding box regression employs DFL loss and CIOU loss, transforming the regression problem into a classification problem. Although the convergence value is not as good as other models, the LDKA-NET in this paper maintains a large model complexity while keeping the parameter count low, allowing it to better capture the complex features in the data instead of overfitting to the noise in the training data, resulting in optimal model performance and good generalization ability.

3.4.4. PR Curve Analysis and Confusion Matrix Analysis

Figure 10a,b display the performance of each class in the model. In Figure 10a, the curve obtained by sequentially adding the improved modules to the baseline model is shown. This curve reflects the trade-off between accuracy and recall of the object detection results at different confidence levels. Generally, the area under the P-R curve (AP, Average Precision) is widely used to evaluate the performance of object detection models, with higher AP values typically indicating better performance. It is clearly observable from Figure 10a that the blue curve effectively integrates the advantages of several other curves, indicating that our model maintains higher precision and recall across different working points, thus improving overall performance. Figure 10b presents a normalized confusion matrix, which provides the accuracy percentage for each class. The results show that the recognition performance is particularly good for insulators and defects.

4. Large Dynamic Kernel Aggregation Net Interface Design

For post-processing of the detected results, we designed the interface as shown in the figure below. On the left side is the original input image, and on the right side is the image showing the detected results. Below the images are detection tabs: Select Image, Image Detection, Select Video, and Video Detection. On the right side are the detection data and the information of the personnel conducting the detection, as shown in the Figure 11 below:

5. Conclusions

To address the low precision issue of existing detection models for ice-covered insulator detection, this paper proposes the LDKA-NET defect detection algorithm based on the wide field of view dynamic convolution network. The main contributions are as follows:

(1) The wide field of view convolution network (WFVC Net) employs larger convolutional kernels to enhance the network’s perceptual and generalization capabilities. It achieves a larger effective receptive field with fewer layers while also exhibiting higher shape bias, enabling it to better capture shape information in images.

(2) The full-dimensional dynamic convolution feature fusion network, as compared to traditional static convolution, significantly improves CNN accuracy by learning linear combinations of multiple convolutional kernels and the weighted attention related to their inputs. This approach offers higher parameter efficiency and enhanced feature extraction capabilities.

(3) The Expectation Maximization Dynamic Convolution Attention (EM-DCA) mechanism focuses on and utilizes important information within the input data, aiding the model in better allocating attention to significant targets or scene components. This results in learning more generalized feature representations.

Results indicate that the proposed algorithm achieves an overall average precision of 99.01%, outperforming comparative algorithms. Future work will focus on enhancing the lightweight universality of this method, deploying lightweight models on the edge to realize integrated applications at the cloud-edge terminal.

Aiming at the problem of insufficient accuracy of existing detection models in the detection of ice-covered insulators, this study proposed the LDKA-NET algorithm and achieved a significant improvement in the average detection accuracy of 99.01%, breaking through the bottleneck of traditional algorithms in terms of the limitation of sensing field and insufficient feature extraction ability, and fully embodying the practical value. By using the full-dimensional dynamic convolution feature fusion network and the expected maximum dynamic convolution attention mechanism, the algorithm achieves high parameter efficiency, a low computation cost, and a stronger feature generalization ability, and is suitable for ice cover defect detection in a variety of complex scenes. In addition, its potential in cloud-edge-to-end lightweight deployment further improves the scalability of practical engineering applications, and is of great significance for real-time monitoring and fault location of power inspection systems.

Author Contributions

Mthodology, X.D.; formal analysis, C.J.; writing—original draft preparation, X.D.; Writing—Reviewing and editing, Y.Z.; Supervise, X.D.; project administration, Y.W. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

No new data were created or analyzed in this study. Data sharing is not applicable to this article.

Conflicts of Interest

Author Xinsheng Dong and Yuanhao Wan were employed by the company State Grid Xinjiang Electric Power Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Transmission lines ice cover monitoring system diagram.

Figure 2. LDKA-NET network structure diagram.

Figure 3. Large dynamic kernel aggregation network.

Figure 4. Full-dimensional dynamic convolutional feature fusion module.

Figure 5. Expectation Maximization Dynamic Convolutional Attention, EM-DCA.

Figure 6. WIoU loss function.

Figure 7. Detection results of baseline model versus algorithm proposed in this paper.

Figure 8. Feature map comparison.

Figure 9. Loss function comparison.

Figure 10. PR curve analysis and normalized confusion matrix display.

Figure 11. Transmission lines ice cover interface design diagram.

Table 1

Ablation study results.

Baseline	WFVC Net	ODC2f	EM-DCA	WIoU	Param	GFLOPs	P	R	mAP50
√					43.608	164.8	98.66	91.87	96.23
√	√				45.452	153.4	96.97	92.46	96.91
√		√			43.887	128.3	98.88	94.60	98.10
√			√		43.887	128.3	95.54	94.44	96.87
√	√	√			45.726	116.9	98.97	93.74	97.91
√	√		√		61.502	246.1	94.72	90.52	94.69
√	√	√	√	√	45.835	117.1	99.42	95.13	99.01

Table 2

Comparison of experiment results.

Method	Backbone	P/%	R/%	mAP0.5/%	Params/M	GFLOPs
Baseline	CSPDarkNet	98.66	91.87	96.23	43.61	164.8
Faster R-CNN	ResNet50	63.79	92.43	90.64	137.10	370.2
SSD	VGG	97.07	92.21	94.02	26.29	62.7
FCOS	ResNet50	96.74	94.88	96.19	32.16	161.9
DETR	ResNet50	74.57	80.91	71.11	36.76	114.2
YOLOv5-L	CSPDarkNet	96.73	93.61	95.90	46.56	109.6
LDKA-NET (Ours)	WFVC Net	99.42	95.13	99.01	45.84	117.1

References

1. Zheng, Y.; Li, Y.; Liu, T.; Mei, L.; Gao, B.; Xiao, X.; Xu, M.; Lin, X. Detection Method and Experiment Research on Icing Thickness for 10 kV Distribution Lines in Plain Area. Guangdong Electr. Power; 2023; 36, pp. 114-120.

2. Liu, B.; Yuan, W. Analysis on Relevance of Surface Discharge Strength to Aging and Pollution of Glass Insulators. Guangdong Electr. Power; 2023; 36, pp. 102-110.

3. Gao, L. Insulator Defect Recognition Based on Lightweight Object Detection. High Volt. Appar.; 2023; 59, pp. 237-244.

4. Deng, C. The method of insulator defect recognition based on group theory. IEEE Access; 2021; 9, pp. 96380-96389. [DOI: https://dx.doi.org/10.1109/ACCESS.2021.3095382]

5. Feng, S.; Chen, Q. Lightweight Glass Insulator Defect Detection Based on Improved YOLOv4. Comput. Simul.; 2024; 41, pp. 116–122+153.

6. Ji, S.; Qu, X.; Song, Q.; Xiao, Y.; Miao, Z.; Li, Y.; Zou, G. Lightweight Insulator Defect Detection Algorithm Based on Improved YOLOv5. Zhejiang Electr. Power; 2023; 42, pp. 64-72.

7. Zhai, Y.; Zhao, X.; Wang, L.; Wang, Y.; Song, X.; Zhu, H. IDD-YOLOv7: A Lightweight Detection Method for Multi-defects of Insulators in Power Transmission Lines. J. Graph.; 2024; 45, pp. 90-101.

8. Wu, K.; Xu, Z.; Shan, H. Rapid Detection Method for Self-explosion Defects of Glass Insulators Based on Improved FasterNet and YOLOv5. High Volt. Technol.; 2024; pp. 1-11.

9. Wang, C. Staged icing forecasting of power transmission lines based on icing cycle and improved extreme learning machine. J. Clean. Prod.; 2022; 208, pp. 1384-1392. [DOI: https://dx.doi.org/10.1016/j.jclepro.2018.10.197]

10. Hao, Y.; Liang, W.; Yang, L.; He, J.; Wu, J. Methods of image recognition of overhead power line insulators and ice types based on deep weakly-supervised and transfer learning. IET Gener. Transm. Distrib.; 2022; 16, pp. 2140-2153. [DOI: https://dx.doi.org/10.1049/gtd2.12428]

11. Qi, L.; Wang, J.; Li, C.D.; Liu, W. Research on the image segmentation of icing line based on NSCT and 2-D OSTU. Proceedings of the International Conference on Modelling; Brittany, France, 2–7 October 2016; [DOI: https://dx.doi.org/10.1109/ICMIC.2015.7409354]

12. Lee, S.Y.; Jeon, J.S.; Le, T.H.M. Feasibility of Automated Black Ice Segmentation in Various Climate Conditions Using Deep Learning. Buildings; 2023; 13, 767. [DOI: https://dx.doi.org/10.3390/buildings13030767]

13. Hu, M.; Wang, Y. Identification and Fault Detection of Transmission Line Insulators Based on Improved YOLOv4. Microcomput. Appl.; 2023; 39, pp. 19-22.

14. Zhang, J.; Wei, X.; Zhang, L.; Chen, Y.; Lu, J. Detection and Localization of Insulators Using Improved YOLO v7. Comput. Eng. Appl.; 2024; 60, pp. 183-191. [DOI: https://dx.doi.org/10.54254/2755-2721/35/20230392]

15. Imambi, S.; Prakash, K.B.; Kanagachidambaresan, G.R. PyTorch. Programming with TensorFlow: Solution for Edge Computing Applications; Springer Nature: Berlin, Germany, 2021.

16. Bai, X.; Xie, Y.; Zhao, M. Defect Detection of Insulators with Small Samples Based on Local Feature Depth Information. Electr. Power Syst. Technol.; 2024; 48, pp. 740-752.

17. Xu, A.; Chen, J.; Zhang, M.; Liu, L. Research on Insulator Target Recognition Algorithm Based on Improved Yolo v4. J. Jilin Univ. (Inf. Sci. Ed.); 2023; 41, pp. 545-551.

18. Zhang, C.; Jin, J. Simulation Study on Insulator Fault Detection Based on YOLOv5s. Electron. Des. Eng.; 2024; 32, pp. 178-182.

19. Fan, M.; Fang, Z.; Yan, Y.; Liu, L.; Huang, P.; Zhong, J. Study on Insulator Defect Detection Method Based on LSKNet. Commun. Inf. Technol; 2024; 1, pp. 49-54.

20. Chi, X.; Zhang, W.; Jia, X.; Xu, Z. Lightweight Insulator Missing Detection Based on Improved YOLOv5s. J. Test. Technol.; 2024; 38, pp. 19-26.

21. Li, Q.; Zhao, F.; Li, B.; Xu, Z.; Yu, J.; Qin, L. Improved Faster R-CNN Algorithm and Its Application in Power Operation Safety Wear Detection. J. Wuhan Univ. (Eng. Sci. Ed.); 2024; 57, pp. 462-469.

22. Li, B.; Huang, H.; Liu, J.; Liu, Z.; Wei, L. Small Target Detection Method for Synthetic Aperture Sonar Images of Interest Based on Improved SSD. Acta Electron. Sin.; 2024; 52, pp. 762-771.

23. Niu, K.; Peng, B.; Yang, X. Lightweight Intelligent Drill Pipe Detection Method Based on Improved YOLOv5s. J. Beijing Univ. Aeronaut. Astronaut.; 2024; pp. 1-19.

24. Chen, Z.; Wang, Y. Research on Airborne Target Detection Based on DATE-FCOS. Comput. Meas. Control; 2024; 32, pp. 78-84.

25. Zhang, X.; Du, X.; Zhang, J.; Tan, P.; Yang, S. Study on Infrared Image Target Detection Method Based on Deformable DETR. Aerosp. Def.; 2024; 7, pp. 16-23.