A quality-relevant deep rule-based system with complementary lifelong learning for adaptive quality prediction in industrial semi-supervised process data streams

Abstract

Deep learning techniques have been widely applied for industrial quality prediction. However, industrial process data are often generated as data streams, which typically exhibit characteristics such as strong nonlinearity, time-varying behavior, and low sampling rates of quality variables. Conventional offline-trained deep learning models often fail to provide accurate predictions for such semi-supervised data streams. Therefore, a quality-relevant deep rule-based system with complementary lifelong learning (QDRSCLL) is proposed to enable adaptive prediction of critical quality variables in streaming data environments. QDRSCLL comprises a deep backbone network and a shallow predictor. The former utilizes a semi-supervised quality-relevant stacked autoencoder (SQSAE) for feature extraction, while the latter employs a hierarchical fuzzy rule system (HFRS) to perform fuzzy inference on hierarchical hidden features. Furthermore, a novel complementary lifelong learning mechanism is proposed to enable QDRSCLL with online incremental learning capabilities. Additionally, semi-supervised learning is integrated into the online learning process to further enhance its deep feature extraction capabilities and the prediction performance. The feasibility and superiority of the proposed method are demonstrated through two real-world processes and four synthetic datasets. Compared to the traditional evolving fuzzy system (EFS), the RMSE of QDRSCLL is reduced by more than 25% in all application scenarios.