Reinforcement Aided Latent Temporal Feature Transfer Learning: Time Series Prediction With Insufficient Labeled Data for Industrial Chemical Process

Time series data of industrial chemical process are typically collected in two ways: distributed control system (DCS) and laboratory test. With the popularity of DCS, a large amount of industrial chemical process monitoring data can be collected (label-sufficient), which is commonly feature rich but information poor. Due to the variant design of the processes, some target features are still monitored through laboratory tests. Data collected in this way cannot support the training of deep learning models as a result of low monitoring frequency and the lack of historical data (label-insufficient). In this study, a reinforcement aided latent temporal feature transfer learning (RALTFTL) method is proposed. It predicts a label-insufficient feature in target domain by learning knowledge from a similar label-sufficient feature in source domain. A transfer learning framework characterized by using latent temporal feature is constructed. Autoencoder is conducted to construct the latent spaces and unify the number of latent features. Reinforcement learning approach is introduced to the framework for feature selection. By taking monitoring features as candidates and taking the improvement of accuracy as reward, it explores a subset of monitoring features that maximizes the accuracy of the prediction model in the target domain. A gas emission (NOx concentration) prediction task is taken as experiment using the practical data from two industrial chemical devices. The performance of RALTFTL is assessed, and the effectiveness of introduced approaches and mechanisms is verified.