Modeling data-driven sensor with a novel deep echo state network

Abstract

Data-driven approach has been widely utilized in modeling soft sensor for predicting key quality variables in process engineering area. The soft sensor is generally a time dependent dynamical model between the input and the output. Echo state network (ESN) is a typical data-driven modeling tool, which has exhibited excellent performance in temporal data processing area. However, the memory mode in the traditional ESN lacks flexibility. It is sometimes hard to preserve sufficient input features in the states, especially for modeling long-term dependent soft sensors. To solve this problem, this paper proposes an asynchronously deep echo state network (ADESN), which is composed of a number of sub-reservoirs that are connected one by one in sequence. Additionally, time delay modules are inserted between every two adjacent layers. The ADESN scheme preserves more input history in the states. Moreover, it can realize a selective memory. The validity of the ADESN is demonstrated on modeling a number of numerical and real-life soft sensors.