Echo-state networks for soft sensor design in an SRU process

Abstract

The implementation of soft sensors for industrial processes is expanding in applications for recent machine learning techniques. In this work, strategies based on reservoir computing are applied to developing dynamical models of target variables in a sulfur recovery unit (SRU) of a refinery plant in Italy. In particular, a specific type of recurrent network, namely an echo-state network (ESN), is adopted to estimate key process variables on the SRU. Two process lines are considered to evaluate the proposed algorithm on different datasets in terms of estimation performance and computational effort of the learning process. The obtained results are evaluated in comparison with other recurrent networks, based on long short-term memory, and with other techniques reported in the literature, demonstrating the feasibility of the proposed approach. Furthermore, the introduction of intrinsic plasticity (IP) is also considered to adapt the reservoir parameters to the provided inputs, achieving a significant improvement in the statistical distribution of the results obtained for the pool of learned networks. The reported results show that ESN-IP represents a suitable solution for identifying dynamical models of the industrial processes, avoiding the time-consuming regressor selection procedure, which is needed when a static network is adopted to design a dynamical model.