A Dynamic Doft Sensor Based on Hybrid Neural Networks to Improve Early Off-spec Detection

Soft sensors are widely used to predict hard-to-measure quality variables in industrial processes. For efficient quality control, prediction of quality dynamics is essential to prevent off-specification production in a process. Recently, dynamic soft sensors have been developed using machine learning techniques. Time-sequential information of quality variables is important to develop a robust dynamic model, but it is rarely considered in soft sensor modeling because there are insufficient data available to construct a time series of quality variables. Hence, we propose a hybrid sequence-to-sequence recurrent neural network-deep neural network (Seq2Seq RNN-DNN) to predict the quality dynamics for an early off-spec detection system. In the RNN unit, the encoder extracts the dynamic states of the process variables, and the decoder generates a time-relevant sequence to improve the long-term time-series prediction of sensor variables. Quality dynamics are then predicted using sensor variables in the DNN unit, trained using combined dataset consisting of offline analysis and simulation data to solve the problem of insufficient data. Finally, the effectiveness of the proposed networks is demonstrated using a 2,3-butanediol distillation process.