A Feature Adaptive Asynchronous Deep Reservoir Computing for Modeling soft sensors

Soft sensor has been widely studied and applied in engineering area, which aims at building a model to estimate quality variables that are difficult to measure by hardware sensors. Asynchronous deep reservoir computing (ADRC) is a promising tool to build data-driven soft sensors due to its training simplicity and high performance in dealing with time-dependent tasks. Nevertheless, how to build a good reservoir for the ADRC has been an open issue owing to many random factors in its structure. This paper proposes a feature adaptive (FA) way to optimize the ADRC's topology. In the FA scheme, a projector-like matrix is introduced into the reservoir matrix, and the singular value distribution of the reservoir matrix can be adjusted by learning the projector-like matrix adaptively. After the projector-like matrix is learnt, the optimal reservoir matrix can be acquired through singular value decomposition (SVD). The FA approach simplifies the reservoir structure. Moreover, it is helpful to improve the generalization capability and suppress the reservoir noise. The FA - RC is applied on modeling two soft sensors in chemical production process. The experimental results shows that the FA scheme can significantly improve the modeling precision and the performance stability of the ADRC soft sensors.