ABSTRACT The Industrial Internet of Things (IIoT) has enabled large‐scale heterogeneous sensor deployment on ultra‐high‐pressure waterjet (UHPW) equipment, creating urgent demand for on‐device intelligence capable of operating reliably under communication constraints and edge‐node resource limitations. Existing cloud‐centric models are computationally heavy and collapse under IIoT transmission failures or electromagnetic interference that induce sensor missingness and high‐intensity noise. This paper proposes Active Stability Co‐Evolution (ASCE), a TinyML‐oriented perturbation‐aware ensemble framework that reformulates robust edge diagnosis as a closed‐loop game‐theoretic co‐evolutionary process. A CVAE‐based perturbation generator with a learnable Beta‐distribution prior co‐evolves with a compact heterogeneous ensemble (1D‐CNN + BiLSTM + LightGBM)—architectures chosen for minimal parameter counts and suitability for memory‐constrained IIoT edge processors. A supervised contrastive learning module decouples invariant fault signatures from stochastic channel perturbations, while a stability‐regularized meta‐learner ensures consistent evidence synthesis across missing‐data scenarios. Experiments on a 12‐h proprietary UHPW dataset demonstrate that ASCE achieves a Macro F1‐score of 0.951 and a Performance Retention Rate of 97.2% under 50% sensor missingness, significantly outperforming all seven state‐of‐the‐art baselines while maintaining a footprint compatible with representative industrial edge hardware.
Li et al. (Tue,) studied this question.