A Fused Deep Fuzzy Neural Network Controller and Its Application to Pneumatic Flexible Joint
Cheng Chen, Yu Cao, Xinxing Chen, Dongrui Wu, Caihua Xiong, Jian Huang
Abstract
With multilayered deep neural networks (DNNs) distinguished from the traditional single-layer neural networks (SNNs) by the depth, deep learning is a powerful paradigm possessing the capability of feature learning from big data. However, as a completely deterministic model, deep learning does not perform well in dealing with uncertain information and must be pretrained before being applied, which hinders the broad application of deep learning in real-time control. In this article, we introduce the knowledge-based fuzzy logic system (FLS) into deep learning and propose a fused deep fuzzy neural network (FDFNN) controller. The FDFNN controller derives the control signal from both deep fuzzy neural networks and the knowledge-based FLS. At the beginning of control, the FDFNN controller performs like a traditional fuzzy logic controller (FLC), which eliminates the pretraining procedure of deep learning. With another FLS providing supervised signal, the network weights of DFNN's hidden layers can be updated during the control process, which endows the FDFNN controller with online learning capability. A Lyapunov-based convergence analysis is conducted to guarantee the stability of the learning procedure. Both numerical simulations and real-world experiments validate that the proposed method effectively improves the real-time tracking control performance of a pneumatic flexible joint system.