A Deep Reinforcement Learning-Based Defect Detection Method for Differential Bevel Gears

The manufacturing quality of differential bevel gears, a critical component in automotive power transmission systems, directly impacts vehicle safety and mechanical stability. Traditional manual inspection methods often fall short under the pressure of modern industrial automation demands, as they are susceptible to operator subjectivity, have limited reliability on reflective surfaces, and struggle to detect minor defects such as cracks or edge peeling. To address these challenges, this paper proposes a defect detection method based on deep reinforcement learning, combining a multi-view observation strategy with an enhanced Faster R-CNN framework. The policy network dynamically optimizes camera angles to achieve comprehensive surface coverage, while the channel-space attention mechanism enhances feature representation for precise defect localization. Additionally, a sequential confidence accumulation strategy is introduced to suppress false positives by leveraging the rotational redundancy of gear teeth. Experiments conducted on over $\mathbf{6, 0 0 0}$ images from real production lines demonstrate that the proposed method achieves excellent accuracy and recall rates in both frame-level and overall gear inspections. The system exhibits strong scalability and industrial applicability, providing a high-performance automated solution for intelligent gear quality assurance.