Deep reinforcement learning-based optimization for machining process: Cutting parameters dynamic optimization considering tool condition degradation

Xuandong Mo, Mingyuan Xia, Shenping Mei, Xiaofeng Hu,
Deep reinforcement learning-based optimization for machining process: Cutting parameters dynamic optimization considering tool condition degradation,
Journal of Manufacturing Processes,
Volume 156, Part A,
2025,
Pages 1006-1031,
ISSN 1526-6125,
https://doi.org/10.1016/j.jmapro.2025.11.031.
(https://www.sciencedirect.com/science/article/pii/S152661252501240X)
Abstract: In modern manufacturing, optimizing cutting parameters is crucial for enhancing efficiency and quality, yet dynamic tool condition degradation during machining process presents significant challenges to the cutting parameters optimization. Current research often neglects the tool condition degradation when optimizing cutting parameters, treating it as a static problem, limiting adaptability and real-time performance. To overcome these limitations, this article proposes a reinforcement learning (RL)-based dynamic optimization method for cutting parameters that incorporates tool condition degradation. Firstly, a multi-task learning teacher-student architecture is developed to comprehensively monitor tool condition using limited labeled data, which can provide accurate tool condition information for the subsequent cutting parameter optimization. Then the dynamic, high-dimensional, and coupled cutting parameter optimization problem is formulated as a Markov decision process (MDP). Finally, considering that the decision variables in this paper include a mixed action space of tool change and cutting parameter adjustment, we design a Proximal Policy Optimization (PPO)-based optimization algorithm. In this approach, the agent selects process parameters or decides whether to change tool for each cutting pass based on real-time machining state, thereby achieving efficient and high-quality manufacturing. Experimental validation was conducted on the actual manufacturing site of the areo-engine casing. The results show that the proposed tool condition monitoring (TCM) method can achieve a monitoring performance of R2 0.9484 for tool wear value and an f1-score of 0.8975 for breakage condition. Compared to ten multi-objective optimization algorithms, the proposed approach improves the optimization effect and computational efficiency by an average of 19.7185 % and 99.2615 %, respectively.
Keywords: Tool condition monitoring; Deep reinforcement learning; Cutting parameters optimization; Surface roughness; Production efficiency