Novel dynamic data-driven modeling based on feature enhancement with derivative memory LSTM for complex industrial process

Abstract

Effective feature extraction is important for accurate data-driven soft sensor modeling in large-scale dynamic industrial processes. However, due to the temporal, nonlinear, and high-dimensional nature of the data collected from large-scale industrial processes, traditional data-driven modeling methods often suffer from imperfect feature extraction. To overcome this challenge, this paper proposes a novel feature enhancement framework based on derivative memory long short-term memory (FEDM-LSTM) algorithm for soft sensor modeling. First, inspired by the proportional–integral–derivative control theory, the LSTM is equipped with a derivative gate that dynamically predicts future information. Combined with inherent gates that resemble the proportional and integral parts in traditional LSTM, the derivative memory LSTM (DM-LSTM) captures the dynamic information of the past, the present, and the future. Then, to adapt to multiple phases in complex industrial systems, a feature enhancement framework is designed for DM-LSTM in which features representing important dynamic information from previous phases are fed to the DM-LSTM as additional input in the current phase. Finally, the effectiveness of the proposed method is evaluated through two real industrial datasets, showcasing its ability to achieve high prediction accuracy.