A Deep Learning-Based Method for Real-Time Monitoring and Anomaly Detection of Electric Pole Force States

This paper proposes a deep learning-based method for real-time monitoring and anomaly detection of electric pole stress states. Traditional methods rely heavily on physical sensors (e.g., strain gauges), which are costly, difficult to maintain, and sensitive to environmental conditions. To address these limitations, we design an Electric Pole Uncertainty State Network (EPUSNet) that integrates multi-stage feature fusion and attention mechanisms to enhance the robustness and accuracy of stress detection from image data. The network incorporates three key modules: CBN-Conv (Convolutional Block Normalization Convolution), GTS-Conv (Graph-based Temporal Spatio Convolution), and TRFS-Conv (Temporal Recursive Feature Sampling Convolution). Experimental results demonstrate that our method achieves 93.4% accuracy in stress state detection, significantly outperforming traditional sensor-based and image-processing approaches. This approach reduces dependency on physical sensors, improves adaptability to environmental variations, and enables efficient real-time monitoring of electric poles in power transmission systems.