Adaptive Optimization of Fault Diagnosis Models Based on Reinforcement Learning and Robustness Quantification Feedback

Spacecraft operate in a highly complex in-orbit environment. Telemetry data from power systems is often plagued by multi-source disturbances such as plasma noise, electromagnetic interference, and communication packet loss, which places very strict demands on fault diagnosis model reliability. Using dynamic disturbance test manifolds and the comprehensive robustness quantification index (RQI) for evaluation, we found a critical vulnerability: baseline models such as the joint loss enhanced global sparse attention transformer (JLE-GSA-Transformer) perform well in resisting continuous signal disturbances, but their performance drops sharply when temporal structures are disrupted due to data missingness. This finding highlights that static diagnostic models have fragile robustness boundaries when operating under extreme non-stationary conditions. To solve this problem, this paper puts forward an adaptive optimization method that combines deep reinforcement learning (RL) with quantitative robustness feedback. By making use of the dynamic disturbance manifold, this method embeds the RQI into the reward function of the RL algorithm. Through continuous updates to the agent’s policy, it drives the adaptive adjustment of key parameters in the JLE-GSA-Transformer, mainly including the coding structure of global sparse attention (GSA) and the discriminative boundaries of temporal cosine-center joint loss. This closed-loop mechanism can significantly improve the stability, adaptability, and long-term reliability of the model when it works in complex environments with multi-source disturbances.