Adaptive error compensation in CNC turning based on deep reinforcement learning and genetic algorithm fusion

Abstract

Abstract Precision manufacturing in CNC turning operations faces persistent challenges from complex, time-varying error sources, encompassing thermal deformation, progressive tool wear, and force-induced deflections. In this work we propose an adaptive error compensation framework that brings deep reinforcement learning (DRL) and genetic algorithms (GA) into a coupled, bidirectional loop, rather than treating them as standalone tools. GA performs global hyperparameter and reward-weight optimisation, while DRL handles real-time adaptive control through continuous interaction with the cutting environment. Validation on aerospace-grade Ti-6Al-4 V turning yields a mean absolute error of 2.6 μm—an 86.3% compensation effectiveness—together with 38% faster convergence than standalone DRL and a process capability index of 1.67 that comfortably clears the six-sigma threshold. Beyond ablation against DRL-only and GA-only configurations, we benchmark the fusion algorithm against three mainstream predictive-compensation baselines (BPNN-based, LSTM-based, and PSO-based) under identical machining conditions, and the proposed approach retains a clear margin across MAE, RMSE, recovery time, and convergence. The limits of the method are also reported: compensation accuracy degrades once cutting parameters drift more than 30% beyond the training envelope or when the workpiece material differs substantially from the training distribution, a constraint that motivates the transfer-learning extensions discussed later in the paper.