PCBCover：面向电路图自动化生成的覆盖率引导强化学习框架

Defects in printed circuit board (PCB) design tool chain can cause functional errors in the resulting PCBs. If these faulty PCBs are used in safety-critical domains such as aerospace or healthcare, the consequences can be severe. Therefore, it is essential to ensure the reliability of PCB design tool chain. However, current testing methods often rely on manually created test cases (e. g., schematics) or randomly generated schematics from schematic generators with fixed configurations. These methods lack effective guidance for generating high-quality test inputs and have limited capability in triggering hidden defects. To improve the effectiveness of schematic generation, two key challenges must be addressed: designing suitable metrics to evaluate parameter configurations, and efficiently searching the large configuration space to find combinations that can trigger defects. To address these challenges, this paper proposes PCBCover, an automated testing approach designed to enhance defect detection in PCB design tool chain. PCBCover combines coverage feedback with an A2C-based reinforcement learning algorithm to guide the generation of schematics that are more likely to trigger defects. PCBCover includes three main components: a coverage extraction module, an A2C-based dynamic configuration search module, and a differential testing module for detecting defects. It defines both static (i. e., netlist level) and dynamic (i. e., simulation level) coverage metrics to evaluate the effectiveness of configurations and uses them as reward signals during the search process. Experiments on real-world tools such as Ngspice and KiCad show that PCBCover significantly outperforms existing methods in detecting defects. It successfully detects 13 real bugs confirmed by tool developers, demonstrating strong detection capability and practical value.