Soft sensor development and applications based on LSTM in deep neural networks

Abstract

In chemical industrial processes, some quality variables are difficult to measure, and thus soft sensors have been proposed as an effective solution. Deep learning has been introduced in soft sensors to deal with the complex nonlinearity of the process, yet lacking the ability for dynamics. This paper introduces Long Short-Term Memory (LSTM) and develops a deep neural network structure based on LSTM as a soft sensor method to deal with strong nonlinearity and dynamics of the process. The effectiveness of the improved modeling method is validated by a sulfur recovery unit benchmark. Then it is applied in a real case of coal gasification, which shows that it is especially suitable for dynamic soft sensor modeling.