ABSTRACT In multistage manufacturing processes, soft sensors have become an important tool for predicting quality. However, existing quality soft sensing methods focus on the correlation between process parameters and quality characteristics rather than causality, which increases the difficulty in determining the root cause that affects product quality characteristics. At the same time, the sampling rates between quality variables are also not the same, and multirate sampling introduces data with unequal‐length‐parameter, which leads to the waste of sampling time information. To address this issue, a causality embedded quality soft sensing framework is developed for identifying key process parameters and predicting quality characteristics. Meanwhile, considering the inconsistency in data length caused by sampling, the data fusion method including weighted and Kalman filtering is developed for preprocessing. Subsequently, a hierarchical causal discovery algorithm of polynomial complexity is proposed to extract their causal features and enhance the interpretability of the framework. Several popular prediction models are employed separately within the soft sensing framework to analyze the performance of different interpretability models. Experimental results from simulation and an industrial tobacco‐production case demonstrate the accuracy of the proposed causal discovery algorithm in finding key parameters and the effectiveness of this framework in quality soft sensing.
Cui et al. (Fri,) studied this question.