Cognitive-cycle learning framework for small-sample industrial process quality prediction

Abstract Quality prediction of industrial process under small-sample conditions is fundamentally constrained by insufficient data availability, which severely limits the performance and robustness of existing prediction methods. To address this challenge, a cognitive-cycle learning framework (CCLF) is proposed in this paper, in which a newly designed trend-aware sampler and a multi-criteria validator (MV) are developed and integrated to generate high-confidence data for augmenting the training set, thereby improving prediction accuracy in data-scarce industrial scenarios. The trend-aware Bayesian sampler generates candidate samples by capturing the macro-level evolutionary patterns of industrial process data. Subsequently, the MV evaluates and filters these candidate samples from five complementary perspectives: causal consistency, distribution density, trend coherence, initial distribution distance, and predictive uncertainty, thereby ensuring the reliability and industrial relevance of the augmented samples. Meanwhile, the cognitive-cycle mechanism continuously updates the knowledge base across iterations, enabling the joint improvement of sample quality and prediction accuracy. Experimental studies conducted on industrial datasets from semiconductor manufacturing and penicillin fermentation demonstrate that CCLF consistently achieves superior prediction performance across varying sample sizes. In particular, under small-sample conditions, CCLF significantly outperforms partial least square, support vector regression, Random Forest, GP-UCB, TuRBO, BOCK, and representative active learning baselines.