A control-oriented operation mode recognizing method using fuzzy evaluation and attention LSTM networks

Abstract

Operation mode recognition is a prerequisite for precise regulation of key performance indicators (KPIs) in industrial processes. However, system uncertainties and the complexity of the process dynamics pose significant challenges in achieving accurate mode partitioning. This study proposes a control-oriented operation mode recognition method called attention-long short-term memory-Monte Carlo simulation (AT-LSTM-MC). First, a fuzzy inference-based ‘indicator regulation potential’ evaluation framework is established to quantitatively describe the maximum control potential of each control variable on KPIs. Subsequently, considering the temporal dependencies of industrial process data, a long short-term memory (LSTM) autoencoder network is employed as the core architecture for feature extraction, where the ‘indicator regulation potential’ guides the LSTM autoencoder through attention layers to extract control-oriented deep clustering features. Finally, the K-means clustering method is utilized to determine the system operation modes based on these deep clustering features. To address uncertainty-induced challenges, multiple Monte Carlo simulations are performed on the operation mode recognition for the same period, thereby obtaining a statistically convergent operation mode. The effectiveness of the proposed method is validated through a case study of an actual industrial process.