Mengjie Li

To address the inefficient search caused by high-dimensional action spaces in distribution network fault restoration and the safety limitations of purely data-driven decisions in previously unseen scenarios, this paper proposes a knowledge-enhanced reinforcement learning framework, termed DQN-KG-BNN-IQR-Review. More specifically, a knowledge graph(KG) is first used to retrieve high-quality historical cases and provide heuristic action recommendations, thereby narrowing the search space during early training and alleviating inefficient exploration. On this basis, a Bayesian neural network(BNN) is introduced to quantify the epistemic uncertainty of the DQN in real time. By combining these uncertainty estimates with a sliding-window interquartile range(IQR), a dynamic safety boundary is constructed to govern the transition between autonomous model decisions and expert takeover, which improves decision robustness in cognitively unfamiliar states. In addition, a retrospective expert review mechanism is designed to replay successful expert trajectories after failed episodes. By correcting trajectory deviations and steering parameter updates toward more informative directions, this mechanism further enhances generalization. Experimental results show that the proposed framework not only accelerates training convergence, but also achieves an effective balance between restoration efficiency and operational safety in self-healing decision making through the coordinated interaction of its constituent modules.