A web development method for transmission line O&M components based on multi-source data sensing and deep reinforcement learning algorithms

Transmission corridors in high-altitude and micro-topographic regions are chronically subjected to strong electromagnetic pulse interference, multi-physics field coupling, and extreme climatic stressors. This results in temporal asynchrony and spatial mismatch in multi-source monitoring data, while environmental disturbances cause confidence drift in single data sources. Furthermore, traditional O&M systems lack topological awareness and dynamic decision-making capabilities, leading to disruptions in the O&M process chain. To address this, we propose a Web development method for power transmission line O&M components based on multi-source data perception and deep reinforcement learning. The proposed method features a B/S architecture tailored for power transmission O&M, where the front end utilizes Vue 3 and Mapbox GL to construct a real-time interactive interface, and the back end integrates data fusion and decision-making modules via a microservices engine. We establish a tri-modal asynchronous spatiotemporal alignment mechanism that integrates UAV vision, micro-meteorological data, and online monitoring. By employing cubic spline interpolation and graph convolutional networks, we extract topological relationships and deep health states among components. We propose an adaptive weighted fusion strategy based on the exponential decay of historical variance to dynamically suppress confidence drift. The O&M decision-making process is formulated as a Markov Decision Process within a large-scale discrete action space, and we introduce a Deep Deterministic Policy Gradient (DDPG) algorithm with Gumbel-Softmax relaxation to handle action discretization. Using O&M costs and failure risks as a multi-objective reward function, we train the system to generate Pareto-optimal policies. Through WebSocket full-duplex communication and ECharts-based dynamic visualization, the system achieves millisecond-level push of DDPG decision commands and enables human-machine collaborative execution. Experimental results demonstrate that in typical scenarios including conductor temperature rise, sudden wind speed increases, and abnormal leakage currents, the temporal causal consistency of multi-source perception exceeds 92%, and the cosine similarity of the health state vector remains between 55% and 85% under comprehensive noise interference; the cost-risk Pareto front exhibits a convex distribution toward the origin, confirming the near-optimality of the decision within the scope of our experimental settings of our experimental settings; the full-link closed-loop integrity rate reaches 100% (mean value across 10 repeated trials, n = 10, with zero variance, and the 95% confidence interval of [99.2%, 100%] was calculated using the Wilson score interval based on 10 independent repeated experiments), with no data reordering or command loss.