Enhanced industrial process modeling with transfer-incremental-learning: A parallel SAE approach and its application to a sulfur recovery unit

Abstract

In industrial processes, quality variable prediction is important for process control and monitoring. Deep learning (DL) methods offer excellent prediction performance and potential paradigm shifts in quality variable modeling. However, in real-world production, the lack of offline labeled data and time-varying data distributions commonly exist, which seriously prohibits practical applications of DL-based predictive models. This paper introduces an enhanced quality variable prediction framework, Transfer-Incremental-Learning Parallel Stacked Autoencoders (TIL-PSAE), to address this challenge. TIL-PSAE integrates three key components: a parallel model structure, a transfer-learning (TL)-based offline training strategy that accumulates knowledge from multiple similar but different processes, and an incremental-learning (IL)-based online adaptation strategy. The model structure comprises two parallel SAEs for extracting process-invariant and target-process-specific features. Offline training involves sequential training using data from different processes, facilitating knowledge accumulation into different parts of model. During online adaptation, the accumulated knowledge remains unchanged while a new combination of knowledge is learned, thus improving online prediction accuracy and avoiding knowledge forgetting. The proposed model is applied to a sulfur recovery unit with four parallel sub-units. Experimental results demonstrate the effectiveness of the proposed model in both offline and online prediction performance.