Traditional empirical selection of CNC machining parameters often neglects the non-linear dynamics and stochastic disturbances present in dynamic cutting environments, leading to inefficiencies and tool failures. This paper introduces an adaptive optimisation framework utilising Deep Reinforcement Learning (DRL) within a Constrained Markov Decision Process (CMDP) to enhance machining performance and ensure operational safety. The Soft Actor-Critic (SAC) algorithm optimises continuous parameters such as spindle speed, feed rate, and depth of cut, while a multi-sensor fusion strategy processes data from vibrations and acoustic emissions through Wavelet Packet Decomposition. A key innovation is the incorporation of a physical‑constraint safety layer that prevents potential spindle overload. Experimental results demonstrated an average Material Removal Rate (MRR) of 1556 mm3/min and an 18.2% reduction in tool wear, outperforming conventional fixed-parameter machining, Genetic Algorithm (GA), and Particle Swarm Optimisation (PSO)–based optimisation methods while maintaining zero constraint violations with real-time inference latency below 15 ms.
Zhao et al. (Mon,) studied this question.