A deep learning-based transmission line icing detection method

Abstract

The application discloses a kind of power transmission line icing detection methods based on deep learning, it is related to power transmission line monitoring image processing field.First, the power transmission line monitoring video image is obtained and the icing detection dataset is constructed, and is labeled according to icing tower body, camera icing / snow / icing and other sheltering etc.Classification;Improved pyramid pooling semantic segmentation network is constructed, and improved attention mechanism is introduced in the main frame segmentation to suppress complex background interference and strengthen small sample icing target representation;Introduce self-supervised contrast learning pre-training strategy to improve model generalization ability;Dynamic weighted compound loss function is constructed to relieve foreground-background extreme imbalance and improve boundary segmentation quality;Test sample is input into model, and the pixel level icing segmentation result and alarm basis output by model are output.The method can significantly improve robustness and real-time performance while maintaining high segmentation accuracy, with a maximum MIoU of 76.83%, and a MIoU of 76.15% and a 2.19 FPS under a lightweight scheme, suitable for power transmission line icing online monitoring and engineering deployment.

Description

Technical Field

[0001] This application relates to the field of power system transmission line status perception and intelligent operation and maintenance technology, specifically to a method for detecting icing on transmission lines based on deep learning semantic segmentation. More specifically, it relates to a pixel-level recognition method for icing areas that combines monitoring image information, attention enhancement mechanisms, contrastive learning pre-training, and a lightweight network design, suitable for online detection and alarm of icing on transmission lines. Background Technology

[0002] Transmission lines are prone to icing under low temperatures, rain, snow, freezing rain, and other meteorological conditions. Icing increases the mechanical load on conductors, alters sag, and degrades insulation performance. In severe cases, it can cause tripping, line breaks, and other faults, affecting the safe and stable operation of the power grid. Therefore, establishing a high-precision, robust, and real-time deployable icing detection technology is of significant engineering importance.

[0003] Existing methods for monitoring icing mainly include mechanistic / empirical model methods based on meteorological and mechanical parameters, online monitoring methods based on field sensors, and visual detection methods based on image processing and machine learning. Mechanistic models typically rely on parameters such as wind speed, temperature, humidity, and liquid water content. However, obtaining some key parameters in the field is difficult, limiting the model's generalization ability. Sensor solutions suffer from high deployment and maintenance costs and limited coverage. Traditional image processing methods (such as edge detection and threshold segmentation) are sensitive to changes in illumination, weather disturbances, and complex backgrounds, resulting in insufficient stability in engineering scenarios.

[0004] In recent years, deep learning methods have achieved good results in the detection of icing on power transmission lines, but the following shortcomings still exist: First, existing studies often clean the monitoring images, weakening or removing samples contaminated by ice, snow, fog droplets, etc., resulting in a distribution difference between the model and the real scene; Second, snow-covered objects, frost layers, and similar textures in complex backgrounds are prone to false detections and false negatives, especially with insufficient ability to identify small samples of iced targets; Third, the number of pixels in the foreground and background of icing is extremely unbalanced, which can easily lead to a decrease in the quality of boundary segmentation; Fourth, the model has a large number of parameters and computational load, making it difficult to balance accuracy and real-time performance when deployed at the edge.

[0005] Therefore, there is an urgent need for a transmission line icing detection method for real-world monitoring scenarios that can reliably identify icing areas even under complex backgrounds and camera contamination conditions, and achieve a balance between accuracy and speed on a limited computing platform, in order to meet the needs of online monitoring and engineering applications. Summary of the Invention

[0006] The purpose of this invention is to achieve highly robust pixel-level recognition of icy areas in complex backgrounds, small sample icy targets, and camera-contaminated scenarios by co-designing the network structure, feature learning mechanism, loss function, and deployment method, without introducing meteorological data, terrain data, or additional sensor information, while also taking into account the real-time requirements of engineering.

[0007] To address the problems of existing transmission line icing detection methods, such as reliance on multi-source external data, complex model structures, and high engineering deployment difficulty, especially in scenarios with complex meteorological conditions, severe camera contamination, and limited edge computing power, existing technologies struggle to achieve stable real-time operation while maintaining detection accuracy. This invention proposes a transmission line icing detection method based solely on a single monitoring image, oriented towards engineering deployment constraints. The invention employs the following technical solution: a deep learning-based transmission line icing detection method, comprising the following steps:

[0008] Step S1: Acquire online monitoring video images of the transmission line and perform frame extraction processing in chronological order; perform pixel-level annotation based on icing morphology and camera obstruction features to construct an icing detection dataset;

[0009] Step S2: Perform preprocessing and multi-strategy data augmentation operations on the monitoring images to enhance the model's adaptability to different weather conditions, shooting angles, and changes in imaging quality;

[0010] Step S3: Construct an improved multi-scale contextual semantic segmentation network as the main model for icing detection. Optimize the pyramid pooling structure in the basic segmentation framework to enhance the ability to model the context of icing regions at different scales.

[0011] Step S4: Introduce an improved attention enhancement mechanism into the main model to recalibrate the channel and spatial dimension features, highlight the response of key icing areas and suppress interference from complex backgrounds;

[0012] Step S5: Construct positive and negative sample pairs using unlabeled transmission line monitoring images for feature representation learning, and use a pre-training-fine-tuning paradigm to transfer the pre-trained parameters to the icing detection model to improve the model's generalization ability under complex meteorological conditions and small sample icing categories.

[0013] Step S: To address the issues of extreme imbalance in the number of pixels between the foreground and background of ice-covered areas and insufficient boundary segmentation quality, an improved joint loss function is constructed. The ice-covered segmentation results are then optimized through a combination of region overlap constraints and hard sample augmentation mechanisms.

[0014] Step S7: To meet the needs of engineering deployment, the model is modified to be lightweight. A lightweight backbone network is adopted and a multi-scale context compensation module based on dilated convolution is designed to significantly reduce the number of model parameters and computational complexity while ensuring detection accuracy.

[0015] Step S8: Input the monitoring image of the transmission line under test into the trained model, output the pixel-level icing segmentation result, and generate icing status alarm information based on the area ratio of the icing area, the characteristics of the connected region, or the preset threshold rule.

[0016] Furthermore, the images obtained in step 1 include at least the following categories: background area, area where the camera is obscured by other foreign objects, area where the camera is fogged or covered by frost, area where the camera is covered by snow, area where the camera is covered by ice, and area where the power transmission tower or conductor is covered by ice.

[0017] Furthermore, the main model for icing detection in step 3 includes a pyramid pooling module, an encoder, and a decoder. The pyramid pooling module is used to extract multi-level semantic features step by step. The encoder is used to extract multi-level semantic features step by step from the input monitoring image. The decoder is used to fuse multi-scale features and restore spatial resolution, and finally outputs a pixel-level category prediction result of the same size as the input image. The encoder and decoder are multi-level structures. Each level of the encoder outputs not only to the next level but also to the same level of the decoder. The final Q level of the decoder is used as the output.

[0018] The pyramid pooling module extracts pooling features from the input image using sizes of 1×1, 2×2, 3×3, and 5×5, respectively, resulting in four branches of pooling features. These four branches of pooling features are then fused and input into the encoder. The encoder's encoding result is then input into the decoder, and the decoder's output is fused. After fusion, the feature map is restored to the same spatial resolution as the input image through convolutional mapping and upsampling operations. Finally, the Softmax function is used to normalize the class probability of each pixel to obtain pixel-level semantic segmentation results.

[0019] Furthermore, step 4 specifically involves introducing an attention enhancement mechanism into the pooling feature branches of the input image extracted at sizes of 2×2, 3×3, and 5×5 in the encoder:

[0020] Step 4.1: Let the encoder input feature map be F∈RH×W×C, where H, W, and C represent the height, width, and number of channels, respectively;

[0021] Each input branch is compressed in the spatial dimension using global average pooling to obtain a channel-level description vector. :

[0022] ;

[0023] =w⊙F=[w1F1,w2F2,...,wCFC], where ⊙ represents channel-by-channel multiplication;

[0024] w=σ(Conv1×1(δ(Conv1×1(z))));

[0025] Where δ represents the ReLU activation function and σ represents the Sigmoid function; the first 1×1 convolution compresses the number of channels to C / r, where r is the reduction rate, and the second 1×1 convolution restores it to C; finally, recalibration is achieved by weighting each channel.

[0026] Step 4.2: Perform channel filtering on the input feature map again to obtain F′′;

[0027] Step 4.3: Perform global average pooling and global max pooling on F′′ along the channel dimension to obtain two two-dimensional feature maps F. vg,F ax∈RH×W×1; then concatenate the two and generate a spatial weight map Ms through a 7×7 convolution:

[0028] Ms=σ(Conv7×7([F vg;F ax]));

[0029] Where [·;·] denotes channel concatenation, and the spatial weight map is multiplied element-wise by F′′ to obtain the final output:

[0030] F′′′=Ms⊙F′′.

[0031] Furthermore, the specific method for step 5 is as follows:

[0032] Step 5.1: Apply two sets of random enhancement operations to the image obtained in Step 1 to generate two different enhanced views:

[0033] xi(1)=T1(xi); xi(2)=T2(xi);

[0034] Where T1(·) and T2(·) are data augmentation functions for random sampling, and the augmentation methods include at least one or more combinations of random cropping, scaling, color jittering, brightness and contrast perturbation, and Gaussian blur, which are used to introduce appearance-level perturbations while maintaining semantic consistency.

[0035] Two enhanced views xi(1) and xi(2) generated from the same original image xi constitute a positive sample pair; enhanced views from different original images in the same training batch constitute a negative sample set.

[0036] Step 5.2: Using the image input shared parameters obtained in Step 5.1, the feature encoder network f(·) constructed in Steps 3 and 4 extracts high-dimensional semantic feature representations.

[0037] hi(1)=f(xi(1)); hi(2)=f(xi(2));

[0038] The feature encoder structure is consistent with the encoder used in the subsequent transmission line icing semantic segmentation model, thereby ensuring that the feature representation learned in the pre-training stage can be seamlessly transferred to the downstream segmentation task.

[0039] To avoid limiting the feature representation ability of the segmentation task by directly imposing contrastive constraints in the high-dimensional semantic space, a projection head network g(·) is introduced after the encoder to map the high-dimensional features to an embedding space dedicated to contrastive learning:

[0040] zi(1)=g(hi(1)); zi(2)=g(hi(2));

[0041] Where zi(1) and zi(2) are feature vectors used for contrastive learning;

[0042] Step 5.3: Define the loss function L IDF :

[0043] L IDF = · +( )· ;

[0044] Where λ(t) is the dynamic weighting coefficient, ;pi∈[0,1] represents the model's predicted probability that the i-th pixel belongs to the ice-covered category,gi∈{0,1} is the true labeled value, N is the total number of pixels, and ϵ is the smoothing coefficient, used to prevent the denominator from being zero and to smooth the gradient;

[0045] ;

[0046] ;

[0047] In the formula, alpha is the class balancing factor, used to alleviate the imbalance between the number of positive and negative samples; γ is the modulation factor, the modulation term. The function is as follows: when the predicted probability pt of a certain pixel is higher than the set threshold, the modulation term approaches 0, reducing the loss contribution of that sample; when pt is less than or equal to the set threshold, the modulation term approaches 1, and the loss contribution is preserved.

[0048] Step 5.4: After completing the learning pre-training, retain the parameter weights learned in the feature encoder f(·), remove the projection head network g(·), and transfer the encoder parameters to the transmission line icing semantic segmentation network as the initial weights of the segmentation model.

[0049] Furthermore, in step 5.2:

[0050] ;

[0051] In the formula, t is the current training round, Tmax is the total training round, λmin and λmax are the preset ranges of the weights, and β is the decay exponent.

[0052] This invention achieves stable identification of icing areas on power transmission lines in complex scenarios without relying on multi-source information such as meteorological and terrain data, thereby reducing system deployment and maintenance costs. Through a self-supervised comparative learning pre-training mechanism, it effectively improves the model's generalization ability under complex meteorological conditions and with a small sample of icing categories. By improving the loss function design, it significantly improves the quality of icing area boundaries and the segmentation of difficult samples. Through a lightweight network structure and deployment-friendly design, the model can perform real-time inference on edge devices with limited computing power, meeting the engineering application requirements for online monitoring of power transmission lines. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a deep learning-based method for detecting icing on power transmission lines, as provided in an embodiment of this application.

[0054] Figure 2 This is an example diagram of the power transmission line icing detection dataset in the embodiments of this application.

[0055] Figure 3 This is a schematic diagram of the pyramid pooling module (PPM) structure in an embodiment of this application.

[0056] Figure 4 This is a schematic diagram of the attention mechanism structure in the embodiments of this application.

[0057] Figure 5 This is a flowchart illustrating the overall process of comparative learning pre-training in the embodiments of this application. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only for illustrating the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the core concept of the present invention, those skilled in the art can make equivalent substitutions and conventional modifications, all of which should fall within the scope of protection of the present invention.

[0059] This embodiment proposes a semantic segmentation detection method for transmission line icing scenarios. The overall process can be summarized as steps S1–S8:

[0060] S1 collects video frames of power line inspections and builds an ice-covered image sample library, completing multi-category pixel-level annotation;

[0061] S2 performs standardized preprocessing and multi-strategy data augmentation on the original samples to form training, validation and testing subsets;

[0062] S3 builds a baseline model based on PSPNet and optimizes the PPM multi-scale context structure to obtain PSPNet-i; S4 introduces improved attention units (SEI and CBAMI) in the backbone network and related pyramid pooling locations to enhance fine-grained recognition.

[0063] S5 superimposes self-supervised contrastive learning pre-training on the optimized segmentation network to improve the discriminativeness and generalization of feature representations;

[0064] S6 constructs an improved loss function (ImprovedDiceFocal) to enhance the segmentation quality of hard samples and boundary regions;

[0065] S7 has undergone lightweight modifications for engineering deployment, adopting the MobileNetV2 backbone and designing PDCM modules;

[0066] The S8 outputs pixel-level icing result images, which, combined with statistics such as category proportion and regional connectivity, form alarm and maintenance assistance information.

[0067] Data Construction and Labeling Standards

[0068] 1) Sample Source and Category System The transmission line icing detection samples used in this embodiment are derived from video data collected by online monitoring cameras of transmission lines. By performing timed frame extraction on the original video sequence, monitoring image samples covering different meteorological conditions, different shooting distances, and different imaging quality states are obtained.

[0069] To address the practical engineering needs of icing detection on transmission lines, a pixel-level semantic segmentation annotation method is adopted to construct a multi-category icing detection annotation system. This system characterizes the icing state of the transmission line itself, the icing or obstruction state of the camera, and complex background interference factors. The category system includes at least the following categories:

[0070] 1. Background (background area);

[0071] 2. Covered Camera (area where the camera is obstructed by other objects);

[0072] 3. IcedCameraFog (Area covered by camera fog or frost);

[0073] 4. Iced Camera Snow (snow-covered area of ​​the camera);

[0074] 5. Iced Camera (area covered in ice for the camera);

[0075] 6. IcedTower (Areas of ice accumulation on transmission towers or conductors).

[0076] Among them, the IcedTower class corresponds to the icing area of ​​the key stress-bearing parts of the transmission line. Its pixel ratio is relatively small, and it has the characteristics of slender structure and complex background overlap. It is a typical small sample and high difficulty segmentation category, which puts higher demands on the model's multi-scale feature expression ability and boundary modeling ability.

[0077] 2) Preprocessing and Data Augmentation Strategies: To improve the model's adaptability to complex weather conditions, different shooting angles, and fluctuations in image quality, multi-strategy preprocessing and data augmentation operations are performed on the original monitoring images during the training phase. The data augmentation strategies include at least: random scaling, random cropping, random rotation, random horizontal or vertical flipping, brightness, contrast, and saturation perturbations, and Gaussian noise or blur perturbations.

[0078] By introducing random perturbations at the appearance level while maintaining semantic consistency of images, the above enhancement methods effectively improve the robustness of the model to complex engineering scenarios such as rain, snow, fog, haze, and changes in lighting, and alleviate the risk of overfitting of small sample classes during training.

[0079] 3) Enhanced Dataset Partitioning Principles: The sample data is divided into training, validation, and test sets, and a fixed random seed is used for partitioning and indexing to ensure the repeatability of model training and performance evaluation.

[0080] In engineering applications, training and validation samples can be constructed by combining information on the regional distribution of transmission lines, seasonal changes, and historical periods of frequent icing, in order to reduce the impact of data distribution offset on the generalization performance of the model.

[0081] Baseline Model and Multiscale Structure Optimization

[0082] 1) An encoder-decoder semantic segmentation network is adopted as the basic model structure for transmission line icing detection. The encoder is used to extract multi-level semantic features from the input monitoring image step by step, and the decoder is used to fuse multi-scale features and restore spatial resolution, finally outputting pixel-level category prediction results of the same size as the input image.

[0083] The baseline model uses the Pyramid Scene Parsing Network (PSPNet) as the basic segmentation framework. It introduces a pyramid pooling module on high-level semantic features to model contextual information at different scales, thereby enhancing the model's overall understanding of complex scenes.

[0084] 2) Improved pyramid pooling module (PPM) structure optimization (1) Problem analysis and improvement motivation

[0085] Determining the optimal scale combination: When the PPM branch extraction size combination was adjusted to 1×1, 2×2, 3×3, and 5×5, the overall MIoU of the model reached 90.97%, which is 1.11 percentage points higher than the baseline combination (1,2,3,6) of 89.86% (see Table 3-5). This improvement has significant practical value in semantic segmentation tasks, proving the effectiveness of scale reconstruction.

[0086] Through controlled variable experiments, this embodiment reveals the functional division of labor among the branches of PPM:

[0087] 1×1 branch: Preserves fine details of the original feature map, making a significant contribution to texture-clear classes such as Covered Camera;

[0088] 2×2 and 3×3 branches: responsible for mesoscale context aggregation, playing a key role in the local structure of the Iced Tower;

[0089] 5×5 branch: Achieve the optimal balance between global semantics and local details, and is specifically optimized for difficult samples such as Iced Camera Fog with blurred boundaries and medium scale.

[0090] This data-driven redistribution of branch functions enables the improved PPM module to adaptively combine multi-scale features based on the scale sensitivity of different categories, achieving an optimal balance between context modeling capability and local structure preservation capability without increasing network depth and parameter scale.

[0091] This embodiment does not simply replace the pooling scale, but rather reconstructs the PPM branch extraction scale based on the scale statistical characteristics, category sensitivity analysis, and systematic experimental verification of the transmission line icing detection task. Experimental results show that the improved scale combination (1,2,3,5) achieves significant improvements in overall accuracy (MIoU improvement of 1.11%) and key fuzzy classes (Iced CameraFog improvement of 3.83%), while maintaining the recognition stability of the slender structure class (Iced Tower). This structural optimization scheme has clear physical meaning and interpretability, reflecting an innovative leap from "general design" to "task customization," and providing a more discriminative feature representation foundation for subsequent introduction of attention mechanisms, contrastive learning pre-training, and other modules.

[0092] 3) Output layer and pixel-level prediction method

[0093] After multi-scale feature fusion, convolutional mapping and upsampling operations are used to restore the feature map to a spatial resolution consistent with the input image. The Softmax function is then used to normalize the class probability of each pixel, resulting in pixel-level semantic segmentation. Through this improved multi-scale contextual modeling structure, the constructed baseline model provides a stable and highly discriminative feature representation foundation for subsequent attention enhancement mechanisms, contrastive learning pre-training, and lightweight structural design.

[0094] Attention Enhancement Modules (SEI and CBAMI)

[0095] To address the challenges of complex backgrounds, large differences in the scale of icing targets, and the tendency for small icing areas to be obscured by background textures in transmission line icing detection scenarios, this embodiment introduces an attention enhancement mechanism based on an improved multi-scale semantic segmentation network. By co-calibrating the channel dimension and spatial dimension features, it strengthens the feature response of key icing areas and suppresses redundant interference in non-icing areas.

[0096] The attention enhancement mechanism includes a channel recalibration unit (SEI) and a spatial-channel collaborative attention unit (CBAMI), which are embedded in the network at different levels and locations to improve the model's ability to distinguish icy areas.

[0097] Let the encoder input feature map be F∈RH×W×C, where H, W, and C represent the height, width, and number of channels, respectively;

[0098] Each input branch is compressed spatially using global average pooling to obtain the channel-level description vector z∈RC:

[0099] ;

[0100] w=σ(Conv1×1(δ(Conv1×1(z))));

[0101] Where δ represents the ReLU activation function and σ represents the Sigmoid function; the first 1×1 convolution compresses the number of channels to C / r, where r is the reduction rate, and the second 1×1 convolution restores it to C; finally, recalibration is achieved through channel-wise weighting.

[0102] =w⊙F=[w1F1,w2F2,...,wCFC], where ⊙ represents channel-by-channel multiplication.

[0103] Based on channel recalibration, the model is guided to focus on the salient location of the icing area in the spatial dimension; CBAMI uses the enhanced feature map F′ output by SEI as input to calculate the improved channel attention and spatial attention respectively, so as to achieve collaborative recalibration.

[0104] Improved channel attention: Using the convolutional channel recalibration of SEI, the input feature map is filtered again to obtain F′′;

[0105] Spatial attention: To capture the precise location of slender structures (such as conductors and towers) and blurred-boundary regions of icy targets, spatial attention employs multi-scale convolution to generate weight maps. First, global average pooling and global max pooling are performed along the channel dimension of F′′ to obtain two two-dimensional feature maps F. vg,F ax∈RH×W×1; then concatenate the two and generate a spatial weight map through a 7×7 convolution (which can be decomposed into 1×7 and 7×1 asymmetric convolutions to adapt to linear targets):

[0106] Ms=σ(Conv7×7([F vg;F ax]));

[0107] Where [·;·] denotes channel concatenation. Finally, the spatial weight map is multiplied element-wise by F′′ to obtain the final output:

[0108] F′′′=Ms⊙F′′;

[0109] 3) Embedding strategy of attention module In this embodiment, the SEI and CBAMI modules are not stacked in a centralized manner, but are embedded in different positions according to the semantic attributes of features at different levels, so as to maximize their efficiency.

[0110] The SEI module is embedded primarily in the high-level semantic feature output stage of the backbone network (ResNet) (e.g., after the last residual block). High-level features have rich semantic information but low spatial resolution. SEI recalibrates the channels here, which can enhance key response channels related to icing texture, edge morphology, and brightness changes, providing cleaner feature inputs for subsequent multi-scale context modeling.

[0111] The CBAMI module is embedded after the improved PPM module (see Example 3), i.e., before and after multi-scale contextual feature fusion. Specifically, CBAMI is placed at the output of the PPM module to simultaneously perform channel filtering and spatial saliency enhancement on the fused multi-scale features, thereby effectively suppressing complex background interference and highlighting the spatial distribution of the icing area. By placing channel attention at a high level of feature extraction and spatial-channel co-attention at the context fusion stage, a progressive feature optimization of "first filtering semantic channels, then accurately locating space" is achieved. This hierarchical design avoids overfitting that may be caused by stacking at a single location (for example, when attention is only placed on the backbone network, the model may overemphasize similar textures in the background, as shown in Table 4, where the accuracy of some attention mechanisms on the IcedCameraFog class decreases significantly), while ensuring computational efficiency.

[0112] Experimental Validation and Analysis: To verify the effectiveness of the SEI and CBAMI modules, this experiment, based on the PSPNet-i model (using improved PPM scales 1, 2, 3, and 5), compared the embedding of different attention mechanisms at different locations. The evaluation metrics used were IoU and MIoU for each category. The experimental results are shown in Tables 6 to 9.

[0113] As can be seen from Tables 6 and 7, embedding a single attention mechanism into the backbone network or PPM module leads to a decrease in overall MIoU, especially in blurry categories such as IcedCameraFog, where the accuracy loss is significant, indicating that stacking at a single location can easily cause overfitting or feature interference.

[0114] Table 8 shows that when the same attention mechanism is embedded in two locations at the same time, the model accuracy is still lower than the baseline. The CA module causes a 7.38% decrease in MIoU, indicating that indiscriminate double embedding actually destroys the existing feature expression.

[0115] Table 9 illustrates the embedding performance of different attention mechanisms. The SE (backbone) + CBAM (PPM) combination (PSPNet-i + SE + CBAM) achieves an MIoU of 90.88%, close to the baseline. The SEI (backbone) + CBAMI (PPM) combination proposed in this patent achieves the best performance: an MIoU of 92.09%, an improvement of 1.12 percentage points over the baseline; the IcedTower class MIoU improves by 2.05% to 76.95%, and the IcedCameraFog class MIoU improves by 1.62% to 94.20%. Meanwhile, the Covered Camera and IcedCamera Snow classes maintain high accuracy.

[0116] The above experiments fully demonstrate that applying a hierarchical strategy of "recalibrating the core high-level embedding channels and coordinating the PPM output embedding space with channels" to SEI and CBAMI can effectively enhance the model's ability to distinguish between blurred boundaries and slender, icy targets, while avoiding misclassification of background textures similar to those of the target. This improvement achieves the effective application of the attention mechanism in complex engineering scenarios without significantly increasing the computational burden, reflecting the algorithm's innovation and practicality.

[0117] Self-supervised contrastive learning pre-training

[0118] This embodiment addresses engineering challenges in transmission line icing detection scenarios, such as high cost of acquiring labeled samples, difficulty in covering complex meteorological conditions, and significant differences in the morphology of icing targets. Building upon the aforementioned improved semantic segmentation network and attention enhancement structure, it further introduces a self-supervised contrastive learning pre-training mechanism for single-modal monitoring images. By performing feature representation learning without manual annotation, the model acquires feature representations that are discriminative of icing-related structures and robust to complex backgrounds and meteorological disturbances before entering supervised segmentation training. This provides high-quality parameter initialization for subsequent transmission line icing semantic segmentation tasks.

[0119] 1) Sample Construction and Enhancement Method In this embodiment, for any unlabeled original transmission line monitoring image sample xi, a multi-view sample is constructed using random data augmentation operators to simulate the appearance differences caused by weather changes, shooting distance variations, and image quality fluctuations in actual engineering scenarios. Specifically, two sets of random enhancement operations are applied to the original image to generate two different enhanced views:

[0120] xi(1)=T1(xi); xi(2)=T2(xi);

[0121] Where T1(·) and T2(·) are data augmentation functions for random sampling, and the augmentation methods include at least one or more combinations of random cropping, scaling, color jittering, brightness and contrast perturbation, and Gaussian blur, which are used to introduce appearance-level perturbations while maintaining semantic consistency.

[0122] Two enhanced views xi(1) and xi(2) generated from the same original image xi constitute a positive sample pair; Enhanced views from different original images in the same training batch constitute a negative sample set; Through the above sample construction method, a training sample set that meets the requirements of contrastive learning constraints can be constructed without manual annotation.

[0123] 2) Feature Encoding and Contrastive Representation Mapping Enhanced View: First, the feature encoder network f(·) constructed in steps 3 and 4 with shared parameters is input to extract high-dimensional semantic feature representations:

[0124] hi(1)=f(xi(1)); hi(2)=f(xi(2));

[0125] The feature encoder structure is consistent with the encoder used in the subsequent transmission line icing semantic segmentation model, thereby ensuring that the feature representation learned in the pre-training stage can be seamlessly transferred to the downstream segmentation task.

[0126] To avoid limiting the feature representation ability of the segmentation task by directly imposing contrastive constraints in the high-dimensional semantic space, a projection head network g(·) is introduced after the encoder to map the high-dimensional features to an embedding space dedicated to contrastive learning:

[0127] zi(1)=g(hi(1)); zi(2)=g(hi(2));

[0128] Where zi(1) and zi(2) are feature vectors used for contrastive learning; the projection head only participates in model training during the pre-training stage and is removed during the subsequent semantic segmentation inference stage to ensure the structural simplicity of the segmentation network.

[0129] 3) Loss Function

[0130] In semantic segmentation tasks, the cross-entropy loss function (CE Loss) tends to bias model predictions towards the majority class (background) when faced with class imbalance, and its ability to handle ambiguous boundaries is weak. To address these issues, this embodiment no longer uses the single CE loss in the traditional PSPNet, but proposes an improved Dice-Focal joint loss function (IDF Loss) that incorporates an adaptive weight adjustment mechanism. Its overall form is as follows:

[0131] L IDF = · +( )· ;

[0132] Here, λ(t) is a dynamic weight coefficient, which is not a fixed constant but adaptively adjusted according to the sample distribution and model learning state as the training epoch t changes. This design aims to use Dice loss to strengthen the overall structural constraints on the foreground region, while using Focal loss to guide the model to focus on difficult pixels with blurred boundaries, achieving complementary advantages of the two at different training stages through dynamic weights.

[0133] To improve the connectivity and integrity of icy regions and mitigate the segmentation issues of slender structures (such as icing on conductors), the loss function directly optimizes the morphological features of the foreground region by constraining the overall overlap ratio between the predicted and ground truth regions; its calculation formula is as follows:

[0134] ;

[0135] In the formula, pi∈[0,1] represents the model's predicted probability that the i-th pixel belongs to the ice-covered category, gi∈{0,1} is the true labeled value, N is the total number of pixels, and ϵ is the smoothing coefficient (set to 1×10-5 in this embodiment) to prevent the denominator from being zero and to smooth the gradient.

[0136] ;

[0137] in:

[0138] ;

[0139] In the formula, alpha is the class balancing factor, used to alleviate the imbalance between positive and negative samples; γ is the modulation factor (γ=2 in this embodiment). Modulation term The function is as follows: when the prediction probability pt of a certain pixel is high (i.e., an easy-to-classify sample), the modulation term approaches 0, reducing the loss contribution of that sample; when pt is small (i.e., a difficult-to-classify sample, such as at the boundary or in a blurry texture), the modulation term approaches 1, and the loss contribution is preserved.

[0140] Innovative principle explanation: In icing detection, a large number of background pixels (such as sky and mountains) are easily classified. Without restrictions, the cumulative effect of their loss will overwhelm the gradient of a small number of iced pixels. Focal loss dynamically reduces the weight of background pixels through modulation terms, allowing the model to focus its attention on difficult icing samples with blurred boundaries, similar textures, or small scales, effectively solving the problem of insufficient learning of difficult samples.

[0141] The simple linear superposition of Dice loss and Focal loss can lead to conflicting optimization directions. To address this issue, this embodiment proposes an adaptive weight adjustment strategy based on training progress, specifically implemented through λ(t):

[0142] ;

[0143] In the formula, t is the current training round, Tmax is the total training round, λmin and λmax are the preset ranges of the weights, and β is the decay exponent.

[0144] Positive sample features of the initial image, {zk This represents the set of features for candidate samples in the same training batch.

[0145] By minimizing the contrast loss function, the feature representations of different augmented views from the same original image remain highly consistent in the embedding space, while constraining the feature representations of different image samples to be distinguishable from each other. This guides the encoder to learn a discriminative feature representation space under unsupervised conditions.

[0146] 4) Pre-training – Fine-tuning transfer method

[0147] After completing the learning pre-training, the parameter weights learned in the feature encoder f(·) are retained, the projection head network g(·) is removed, and the encoder parameters are transferred to the transmission line icing semantic segmentation network as the initial weights of the segmentation model.

[0148] Subsequently, supervised fine-tuning training was performed using labeled data. Since the encoder had already learned, during the pre-training phase, feature representations that exhibit stable responses to icing structure features and the ability to suppress complex backgrounds and meteorological disturbances through contrastive constraints, the model's segmentation robustness under small sample icing categories and complex meteorological conditions could be effectively improved.

[0149] By introducing the aforementioned self-supervised contrastive learning pre-training mechanism, effective constraints on feature distribution are achieved without increasing manual annotation costs. This enables the model to possess strong discriminative ability from the initial training stage, providing a more stable and generalizable parameter initialization foundation for subsequent supervised segmentation training. Simultaneously, after each training epoch, model performance is evaluated based on the mean intersection-over-union (MIoU) metric of the validation set, and the weights of the best-performing model are saved to suppress overfitting and improve the model's generalization ability in unknown scenarios.

[0150] In engineering applications, the overall performance of the model under actual deployment conditions can be evaluated by combining metrics such as frames per second (FPS), inference latency, and memory usage.

[0151] This embodiment addresses the application constraints faced by transmission line icing detection models in actual engineering deployments, such as long-term online operation, limited computing power of edge devices, and strict requirements for inference latency. While ensuring that the icing segmentation accuracy does not decrease excessively, a lightweight design is implemented for the overall structure of the model, and a deployment-friendly icing detection network and online operation scheme that balances detection accuracy and inference efficiency is proposed.

[0152] 1) Lightweight modification of the backbone network In this embodiment, in order to address the problems of large number of parameters, high computational overhead and difficulty in real-time operation on edge devices of traditional heavy backbone networks, the backbone network in the original semantic segmentation model is replaced with the lightweight backbone network MobileNetV2 to significantly reduce the number of model parameters and floating-point operations (FLOPs).

[0153] MobileNetV2 decomposes standard convolution into channel-wise and point-wise convolutions by introducing depthwise separable convolutions and inverse residual structures, significantly reducing model computational complexity while maintaining strong feature extraction capabilities. Through this backbone network replacement, the model's throughput and response speed during inference are significantly improved, providing a structural foundation for real-time deployment on edge or embedded platforms.

[0154] 2) Construction of PDCM Module and Semantic Compensation Mechanism: In view of the potential shortcomings of lightweight backbone networks in high-level semantic modeling capabilities, this embodiment introduces a Pyramid Dilated Convolution Module (PDCM) based on dilated convolution at the multi-scale context modeling position to provide targeted compensation for the decrease in semantic expressive capabilities caused by lightweighting.

[0155] The PDCM module, based on the original pyramid pooling structure, effectively expands the receptive field of features without increasing feature map downsampling by setting up multi-branch dilated convolutional paths in parallel and configuring dilated convolutional kernels with different dilation rates for each branch. Compared with pooling-based contextual modeling, dilated convolution can capture larger-scale contextual semantic information while maintaining spatial resolution, thereby enhancing the model's ability to model the overall morphology of icy areas and complex background relationships.

[0156] By embedding the PDCM module into the multi-scale context modeling location of the lightweight backbone network, the model can simultaneously acquire local detailed features and global semantic information under low computational overhead, thereby achieving semantic expression capability compensation under the lightweight structure and achieving an effective balance between accuracy and inference speed.

[0157] 3) Deployment-friendly online inference workflow design: After the model training is completed, the best-performing icing detection model is converted into a deployment-friendly intermediate representation format, including but not limited to ONNX or TensorRT formats, to adapt to different hardware platforms and inference engines.

[0158] During the online operation phase, the real-time acquired video stream of the transmission line monitoring is processed by frame extraction at fixed time intervals. The extracted image frames are then input into a lightweight icing detection model for inference, outputting corresponding pixel-level icing segmentation mask results. Based on this, the segmentation results are further post-processed to statistically analyze the pixel area ratio, number of connected regions, and spatial distribution characteristics of each icing category, and to generate icing risk levels according to preset rules.

[0159] When the system detects that the icing risk level continuously exceeds the set threshold in multiple consecutive frames of inference results, it automatically triggers an icing alarm and reports it, thus forming an efficient, stable, and real-time deployable icing detection application process for actual power transmission line monitoring scenarios.

[0160] System Implementation and Alternative Solutions

[0161] 1) The system implementation structure is based on the transmission line icing detection method described in the foregoing embodiments. The present invention also provides an icing detection system for performing the method. The system can be deployed on edge computing devices, servers or embedded platforms to meet the engineering application requirements of online monitoring of transmission lines.

[0162] The icing detection system includes at least a processor, a memory, and a computer program stored in the memory. When the computer program is called and executed by the processor, it is used to implement the steps of the deep learning-based transmission line icing detection method in any of the foregoing embodiments, including but not limited to: monitoring image acquisition and preprocessing, feature extraction and semantic segmentation reasoning, icing area discrimination, and icing status output.

[0163] In a specific implementation, the system can be implemented in software or in a hardware-software co-processing manner. The processor can be a general-purpose processor, a graphics processor, or a dedicated acceleration chip. The memory is used to store model parameters, running instructions, and intermediate result data.

[0164] 2) Alternative embodiments: Without departing from the technical concept of the present invention, those skilled in the art can make equivalent substitutions or modifications to some structures or parameters in the above embodiments, and all such substitutions or modifications should fall within the protection scope of the present invention.

[0165] For example, in terms of lightweight network architecture, other backbone networks with low parameter count and low computational complexity can be used to replace MobileNetV2, including but not limited to the ShuffleNet series and GhostNet series, to adapt to deployment requirements under different computing power conditions.

[0166] Regarding the attention enhancement mechanism, while keeping the core idea of ​​"channel recalibration and spatial attention co-modeling" unchanged, other low-computational-cost attention structures can be used to replace the SEI or CBAMI modules to achieve feature enhancement and background suppression in key icing areas.

[0167] In terms of training strategies, the weight coefficients of each component in the joint loss function, the temperature parameters in contrastive learning, the input image resolution, and the inference threshold can all be adjusted according to the climate characteristics of different regions, different route scenarios, and the computing power of the equipment to obtain better engineering adaptation results.

[0168] 3) Engineering Deployment and Operation: In engineering application scenarios, the icing detection system can be linked with the online monitoring equipment of the transmission line to automatically identify and determine the status of the icing area by periodically extracting frames from the real-time video stream.

[0169] To ensure the stability and adaptability of the system under long-term operating conditions, the system can adopt an "offline periodic retraining + online parameter update" operation mode, that is, new samples are periodically introduced in the background to update the model offline, while a stable online inference service is maintained in the front end.

[0170] Meanwhile, the model input parameters and alarm thresholds can be configured regionally by combining the terrain features of the transmission line, the installation angle of the camera, and information on historical high-incidence periods of icing, thereby improving the system's adaptability and reliability in different operating environments.

[0171] Table 1: Model accuracy performance when the extraction size of a certain branch of the PPM module is changed to 4 (unit: %)

[0172] Table 2: Model accuracy performance when the extraction size of a certain branch of the PPM module is changed to 5 (unit: %)

[0173] Table 3: Model accuracy performance when the extraction size of a certain branch of the PPM module is changed to 7 (unit: %)

[0174] Table 4: Model accuracy performance when the extraction size of a certain branch of the PPM module is changed to 8 (unit: %)

[0175] Table 5: Model accuracy performance when the extraction size of a certain branch of the PPM module is changed to 9 (unit: %)

[0176] Table 6: Accuracy performance of models with different attention mechanisms embedded in the backbone network (unit: %)

[0177] Table 7: Accuracy performance of models with different attention mechanisms embedded in the PPM module (unit: %)

[0178] Table 8: Accuracy performance of models with the same attention mechanism embedded in the backbone network and PPM module (unit: %)

[0179] Table 9: Accuracy performance of models embedded in the backbone network and PPM module with different combinations of attention mechanisms (unit: %)

Claims

1. A deep learning-based method for detecting icing on power transmission lines, comprising the following steps: Step S1: Acquire online monitoring video images of the transmission line and perform frame extraction processing in chronological order; Pixel-level annotations are performed based on icing morphology and camera occlusion features to construct an icing detection dataset; Step S2: Perform preprocessing and multi-strategy data augmentation operations on the surveillance video images; Step S3: Construct an improved multi-scale contextual semantic segmentation network as the main model for ice detection, and optimize the pyramid pooling structure in the basic segmentation framework; Step S4: Introduce an improved attention enhancement mechanism into the main model to recalibrate the channel and spatial dimension features; Step S5: Construct positive and negative sample pairs using unlabeled transmission line monitoring images for feature representation learning, and transfer the pre-trained parameters to the icing detection model using a pre-training-fine-tuning paradigm. Step 5.1: Apply two sets of random enhancement operations to the image obtained in Step 1 to generate two different enhanced views: xi(1)=T1(xi); xi(2)=T2(xi); Where T1(·) and T2(·) are data augmentation functions for random sampling, and the augmentation methods include at least one or more combinations of random cropping, scaling, color jittering, brightness and contrast perturbation, and Gaussian blur, which are used to introduce appearance-level perturbations while maintaining semantic consistency. Two enhanced views xi(1) and xi(2) generated from the same original image xi constitute a positive sample pair; enhanced views from different original images in the same training batch constitute a negative sample set. Step 5.2: Using the image input shared parameters obtained in Step 5.1, the feature encoder network f(·) constructed in Steps 3 and 4 extracts high-dimensional semantic feature representations. hi(1)=f(xi(1)); hi(2)=f(xi(2)); The feature encoder structure is consistent with the encoder used in the subsequent transmission line icing semantic segmentation model, thereby ensuring that the feature representation learned in the pre-training stage can be seamlessly transferred to the downstream segmentation task. To avoid limiting the feature representation ability of the segmentation task by directly imposing contrastive constraints in the high-dimensional semantic space, a projection head network g(·) is introduced after the encoder to map the high-dimensional features to an embedding space dedicated to contrastive learning: zi(1)=g(hi(1)); zi(2)=g(hi(2)); Where zi(1) and zi(2) are feature vectors used for contrastive learning; Step 5.3: Defining the loss function L IDF : L IDF = · +( )· ； Where λ(t) is the dynamic weighting coefficient, ;pi∈[0,1] represents the model's predicted probability that the i-th pixel belongs to the ice-covered category,gi∈{0,1} is the true labeled value, N is the total number of pixels, and ϵ is the smoothing coefficient, used to prevent the denominator from being zero and to smooth the gradient; ； ； In the formula, alpha is the class balancing factor, used to alleviate the imbalance between the number of positive and negative samples; γ is the modulation factor, the modulation term. The purpose of this is to: when the predicted probability pt of a pixel is higher than a set threshold, the modulation term approaches 0, reducing the loss contribution of that sample; when pt is less than or equal to the set threshold, the modulation term approaches 1, and the loss contribution is preserved. ； In the formula, t represents the current training round. For the total number of training rounds, and The weights are a preset range, and β is the decay exponent; Step 5.4: After completing the learning pre-training, retain the parameter weights learned in the feature encoder f(·), remove the projection head network g(·), and transfer the encoder parameters to the transmission line icing semantic segmentation network as the initial weights of the segmentation model; Step S6: Construct an improved joint loss function to synergistically optimize the ice-covered segmentation results through region overlap constraints and hard sample augmentation mechanisms; Step S7: Employ a lightweight backbone network and design a multi-scale context compensation module based on dilated convolution; Step S8: Input the monitoring image of the transmission line under test into the trained model, output the pixel-level icing segmentation result, and generate icing status alarm information based on the area ratio of the icing area, the characteristics of the connected region, or the preset threshold rule.

2. The deep learning-based method for detecting icing on transmission lines as described in claim 1, characterized in that, The images obtained in step 1 include at least the following categories: background area, area where the camera is obscured by other foreign objects, area where the camera is fogged or covered by frost, area where the camera is covered by snow, area where the camera is covered by ice, and area where power transmission towers or conductors are covered by ice.

3. The deep learning-based method for detecting icing on transmission lines as described in claim 1, characterized in that, The icing detection main model in step 3 includes a pyramid pooling module, an encoder, and a decoder. The pyramid pooling module is used to extract multi-level semantic features step by step. The encoder is used to extract multi-level semantic features step by step from the input monitoring image. The decoder is used to fuse multi-scale features and restore spatial resolution, and finally outputs a pixel-level category prediction result of the same size as the input image. The encoder and decoder are multi-level structures. Each level of the encoder outputs not only to the next level but also to the same level of the decoder. The final Q level of the decoder is used as the output. The pyramid pooling module extracts pooling features from the input image using sizes of 1×1, 2×2, 3×3, and 5×5, respectively, resulting in pooling features of four branches. The pooling features from the four branches are fused and then input into the encoder. The encoder's encoding result is input into the decoder, and the decoder's output is fused. After fusion, the feature map is restored to the same spatial resolution as the input image through convolutional mapping and upsampling operations. The Softmax function is then used to normalize the class probability of each pixel to obtain pixel-level semantic segmentation results.

4. The deep learning-based method for detecting icing on transmission lines as described in claim 3, characterized in that, Step 4 involves introducing an attention enhancement mechanism into the pooling feature branches of the input image extracted at sizes of 2×2, 3×3, and 5×5 in the encoder. Step 4.1: Let the encoder input feature map be F∈RH×W×C, where H, W, and C represent the height, width, and number of channels, respectively; Each input branch is compressed in the spatial dimension using global average pooling to obtain a channel-level description vector. : ； =w⊙F=[w1F1,w2F2,...,wCFC], where ⊙ represents channel-by-channel multiplication; w=σ(Conv1×1(δ(Conv1×1(z)))); Where δ represents the ReLU activation function and σ represents the Sigmoid function; the first 1×1 convolution compresses the number of channels to C / r, where r is the reduction rate, and the second 1×1 convolution restores it to C; finally, recalibration is achieved by weighting each channel. Step 4.2: Perform channel filtering on the input feature map again to obtain F′′; Step 4.3: Perform global average pooling and global max pooling on F′′ along the channel dimension to obtain two two-dimensional feature maps. , Then, the two are concatenated and a spatial weight map Ms is generated through a 7×7 convolution. ； Where [·;·] denotes channel concatenation, and the spatial weight map is multiplied element-wise by F′′ to obtain the final output: F′′′=Ms⊙F′′.

Images