Pixel-Level autonomous identification method for self-explosion defects of aerial insulators based on Deep Q-Network and reinforcement learning

Visible light inspection of insulators by unmanned aerial vehicles (UAVs) often suffers from image degradation caused by interference such as illumination, vegetation and atmospheric conditions. Conventional detection methods can only realize fault detection, and fail to conduct pixel-level quantitative measurement of damaged areas. To address this problem, a pixel-level autonomous identification method for self-explosion defects of UAV aerial insulators based on Deep Q-Network (DQN) and reinforcement learning is proposed. Dark channel dehazing and bilateral edge-preserving denoising are integrated to enhance insulator images. An accurate positioning and segmentation method for insulator strings based on the improved CV model is adopted. The inclination angle is calculated by combining the saliency of the blue channel of insulators and Hough transform, and an adaptive elliptical level set is constructed to extract the target regions of insulators in images. The segmented images are imported into the pixel-level autonomous identification model for insulator self-explosion defects built on DQN and reinforcement learning. Multi-scale features of self-explosion defects are extracted to generate coarse defect probability maps. Pixels with ambiguous boundaries in masks are screened to establish an interactive environment for DQN reinforcement learning. The Intersection over Union (IoU) of segmentation is taken as the reward to dynamically adjust pixel thresholds and optimize defect contours. The corrected probabilities are fused to generate binary defect masks, the damaged area is converted, and pixel-level identification of defects is realized. Experimental results demonstrate that the proposed method can achieve pixel-level autonomous identification of insulator self-explosion defects and output accurate quantified damage sizes under high-interference field working conditions. It can provide reliable quantitative data support for refined fault grading and operation maintenance of transmission lines.