Dynamic Class-Conditional Diffusion Adaptation for Class-Imbalanced Cross-Machine Machining State Recognition

Abstract Machining chatter is a kind of unstable self-excited vibration phenomenon that degrades surface quality, accelerates tool wear, and reduces machine tool service life. Deploying intelligent chatter recognition across different machines faces two coupled obstacles: domain shift caused by differences in structural dynamics and operating conditions, and severe class imbalance in which weak chatter is critically scarce relative to stable cutting. Unlike prior approaches that address these challenges in isolation or apply generative augmentation in a static, offline manner, this paper proposes a dynamic class-conditional diffusion adaptation (DCDA) framework, whose central idea is to embed a single-domain conditional denoising diffusion probabilistic model (DDPM) directly into the unsupervised domain adaptation (UDA) training loop, allowing minority-class augmentation to be steered in real time by per-class accuracy feedback and screened by confidence-based filtering. Specifically, DCDA dynamically allocates generation resources toward underperforming classes based on real-time classifier feedback and employs dual-criterion confidence filtering to ensure the quality of injected samples. This dynamic, in-loop coupling of generation and adaptation distinguishes DCDA from prior static augmentation pipelines. Bidirectional cross-machine transfer experiments on two real computer numerical control (CNC) turning machines demonstrate that DCDA achieves a mean accuracy of 90.40% and a mean macro F 1 -score of 80.63%, outperforming eight UDA baselines relative to the multikernel maximum mean discrepancy baseline.