Memory-Adaptive Supervised LSTM Networks for Deep Soft Sensor Development of Industrial Processes

In the recent years, deep long short-term memory (LSTM) networks have achieved tremendous successes for soft sensor developments of dynamical industrial processes. Furthermore, the supervised LSTM models have also been proposed to enhance the prediction accuracy, which, however, requires real-time measurements of quality variables. In this article, we propose a novel memory-adaptive supervised LSTM (MA-SLSTM) network, where the historical states in the deep LSTM models, including the long and short memories over the recent time window, are adaptively updated to track the quality variables. To perform such supervisions, a new optimization strategy is designed, where the objective function is defined as the adaptation cost, while the prediction error is adopted as the constraint. Efficient solution algorithms solving the proposed optimization problem are derived accordingly. The MA-SLSTM is adaptive to varying industrial conditions; meanwhile, it does not impose strict assumptions on the measurements of quality variables. The advantages of the MA-SLSTM are illustrated through three industrial benchmark cases.