Reinforcement Learning-based five-axis continuous inspection method for complex freeform surface

Abstract

Five-axis continuous inspection is an innovative technology that allows for the efficient and precise inspection of freeform surfaces. Traditional methods for planning the five-axis inspection path rely on manually defined objective functions, which are heavily dependent on the expertise of professionals and often result in suboptimal paths. To overcome these challenges, we have developed a Reinforcement Learning (RL)-based approach for generating inspection paths. This method replaces the explicit objective function with an RL model that incorporates comprehensive geometric metrics of inspection surface, resulting in a high-performing head trajectory for the five-axis inspection path. Additionally, we have introduced a beam search-based method to generate a set of optimal head trajectories that cover the entire inspection surface. Our proposed method enables the automatic generation of short and smooth inspection paths without human intervention. Physical inspection experiments conducted on a five-axis inspection machine have demonstrated that our approach significantly improves inspection efficiency and automation compared to traditional benchmarks.