Multi-pass cutting parameters optimisation with causal reinforcement learning for deformation control of thin-walled parts

Cutting parameters are pivotal for ensuring the machining quality of thin-walled parts. To this end, this paper proposes a causal reinforcement learning framework of dynamic cutting parameters optimisation for thin-walled parts. Firstly, the optimisation task is formulated as a Markov Decision Process. Secondly, to capture effects of cutting parameters on workpiece quality, machining physical states are introduced as intervening nodes, and their causality is identified by the NOTEARS-nonlinear algorithm, thereby constructing a quality-oriented causal graph. Based on this, a shared causal graph attention module is designed in Soft Actor Critic’s framework. Specifically, to capture dynamic causality arising from fluctuations in machining states, critic network performs self-attention over the quality-states subgraph and generates a more accurate value for action evaluation. Shared by the critic, the actor finds the relevant states via backtracking from the quality node on the subgraph, and executes cross attention to assign dynamic weights to each cutting parameter for modulating action policies. In case studies, the proposed method reduces the maximum deformation error by 26.68% and 17.79% and improves machining efficiency by 1.57% and 3.02% compared with two benchmarks, offering a practical paradigm for thin-walled part milling and related manufacturing scenarios with high precision requirements.