The vision task of intelligent inspection of transmission lines is crucial to safety and stability of the power system. Although deep learning networks perform well on uniformly distributed training and test datasets， deviations in data distribution often degrade model performance in real-world applications. To solve this problem， a Training Method based on Contrastive Learning （TMCL） was proposed， aiming to enhance robustness of the model. Firstly， a benchmark test set， TLD-C （Transmission Line Dataset-Corruption）， specially designed for transmission line scenario was constructed to evaluate the model’s robustness facing image corruption. Secondly， the model’s ability to distinguish different categories of features was improved by constructing positive and negative sample pairs that are sensitive to category features. Thirdly， a joint optimization strategy combining contrastive loss and cross-entropy loss was used to impose additional constraints on the feature extraction process， so as to optimize representation of the feature vectors. Finally， a Non-local Feature Denoising network （NFD） was introduced to extract features closely related to categories. Experimental results show that compared to the original method， the improved training method achieves an average precision improved by 3.40 percentage points on Transmission Line Dataset （TLD）， and a relative Corruption Precision （rCP） increased by 4.69 percentage points on TLD-C dataset.

In transmission line inspection tasks， utilizing deep learning technology to track the movement of construction machinery effectively is crucial for smart grid construction. To address the issue of significant performance degradation in multi-object tracking caused by occlusion among targets and false or missed detections， a multi-object tracking algorithm combining improved YOLOv5s and optimized ByteTrack was proposed. In the object detection section： firstly， lightweight Ghost convolution and SimAM were used to construct the SGC3 （SimAM and Ghost convolution with C3） module， thereby improving feature utilization and reducing redundant computations in the algorithm. Secondly， in deeper layers of the backbone network， a convolution-guided triplet attention module R-Triplet （RFAConv with Triplet attention） was proposed， thereby using a multi-branch structure to enhance cross-dimensional information interaction of the algorithm and suppress irrelevant background information to improve object association capability. Finally， in the feature fusion stage， a Multi-branch Receptive Block （MRB） was added， thereby utilizing dilated convolution to expand the receptive field of the object and enhancing reuse of multi-scale global feature information of the object. In the object tracking section： based on ByteTrack algorithm， according to motion characteristics of construction machinery， an NSA （Noise Scale Adaptively） Kalman filter algorithm with adaptive noise scale computation was proposed to decrease the influence of low-quality detection boxes on filtering performance. At the same time， Gaussian Smoothing Interpolation （GSI） algorithm was introduced into the data association process to further optimize multi-object tracking performance. Experimental results indicate that compared to the baseline algorithm YOLOv5s， the proposed CRM-YOLOv5s algorithm achieves mean Average Precision （mAP） of 97.4%， which is improved by 3.8 percentage points with the of parameters and floating-point operations reduced by 0.28×10⁶ and 1.8 GFLOPs， respectively， demonstrating stronger generalization capability in various application scenarios. Additionally， compared to the original YOLOv5s+ByteTrack tracking algorithm， after combining with improved ByteTrack， the proposed CRM-YOLOv5s algorithm has the Multiple Object Tracking Accuracy （MOTA） increased by 4.5 percentage points， the number of Identity switches （IDs） decreased by 15， and higher inference speed， demonstrating that the algorithm is suitable for multi-object tracking task of construction machinery in transmission line scenarios.