In real-world network environments, rerouting in response to link failures often requires considerable time, which makes frequent disruptions especially detrimental to overall network performance. To overcome this limitation, we propose PRFR, a reinforcement learning–based framework designed for rapid fault detection and adaptive route computation in the presence of link failures. Central to PRFR is the Path Impact Matrix (PIM) Encoder, which effectively detects link anomalies and integrates this information into the agent’s learning cycle, thereby establishing a direct association between network failures and routing strategies. In addition, action filtering rules are designed based on prior knowledge extracted from the topology. During each action sampling phase, an action mask vector is generated according to these rules, modifying the distribution of the policy network output, which in turn improves training efficiency. Extensive experiments across various real-world network topologies demonstrate that PRFR consistently outperforms conventional methods under failure conditions, achieving more balanced load distribution.