Causality enhanced deep learning framework for quality characteristic prediction via long sequence multivariate time-series data

Abstract Prediction of product quality characteristics plays an important role in the timely identification of quality conditions and in triggering an alarm for abnormal products. In modern manufacturing, the large number of parameters collected by sensors and complex interactions between operational parameters have led to complexity and difficulty in quality monitoring during the production process. To minimize noise interference, reduce modeling complexity, and improve prediction accuracy and interpretability, this study proposes a time-series causal discovery and quality prediction framework for multistage manufacturing processes. Initially, a hierarchical Peter–Clark momentary conditional independence algorithm with multiple time-lag detection accuracy algorithm was proposed. It is designed to identify optimal time lags, establish causal relationships between process parameters and quality characteristics, and efficiently extract the root process parameters. Furthermore, a temporal pattern attention–long short-term memory model is employed to predict the quality characteristics for multivariate time series data, and is aided by the obtained causal structure. Finally, data obtained through a simulation and a case study involving a multistage continuous production chemical process are utilized to verify the performance and superiority of the proposed method.