A dynamical convolutional neural network–based adaptive observer for information-poor systems

Abstract

The objective of this paper is to propose a novel self-learning observer for effectively estimating unmeasured states of a dynamical information-poor system. By inspiring the success of convolutional neural network (CNN) in image and voice recognition applications, the dynamical convolutional neural network (DCNN) architecture (which is nonlinear-in-parameter unlike traditional basis functions–based neural network or fuzzy logic architectures) is proposed for approximating the unknown system dynamics, uncertainties, and disturbances. Instead of using an output observation error filtering method, complex weight-updating mechanism, strictly positive real (SPR) assumption, offline-tuning mechanisms, static approximation of gradients in back-propagation algorithm, or any other strong or strict requirement, the simple yet effective and novel rules for tuning DCNN weights and kernels are derived. Furthermore, the problems such as parameters over-tuning and unnecessary consumption of large amounts of computational power commonly found in traditional neuro-adaptive and fuzzy-adaptive controllers/observers are addressed and their solution is incorporated in the design of the observer. Thus, the proposed DCNN-based adaptive observer is energy-efficient. A quadratic function based on the convolution operator is also proposed which is utilized to provide the closed-loop stability analysis using Lyapunov’s approach. An example is simulated to demonstrate the performance of the proposed observer.