Hierarchical linear and nonlinear adaptive learning model for system identification and prediction

• Published in: Applied intelligence (Dordrecht, Netherlands) Netherlands), 2020-06, Vol.50 (6), p.1699-1710
• Main Authors: Jami’in, Mohammad Abu, Anam, Khairul, Rulaningtyas, Riries, Mudjiono, Urip, Adianto, Adianto, Wee, Hui-Ming
• Format: Article
• Language: English
• Subjects: Adaptive learning; Adaptive systems; Algorithms; Artificial Intelligence; Autoregressive moving average; Autoregressive moving-average models; Computer Science; Control theory; Learning; Liapunov functions; Machines; Manufacturing; Mechanical Engineering; Model accuracy; Neural networks; Processes; System identification
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-019-01615-0

In this paper, we propose a method to increase the model accuracy with linear and nonlinear sub-models. The linear sub-model applies the least square error (LSE) algorithm and the nonlinear sub-model uses neural networks (NN). The two sub-models are updated hierarchically using the Lyapunov function. The proposed method has two advantages: 1) The neural networks is a multi-parametric model. Using the proposed model, the weights of NN model can be summarized into the coefficients or parameters of auto-regressive eXogenous/auto-regressive moving average (ARX/ARMA) model structure, making it easier to establish control laws, 2) learning rate is updated to ensure the convergence of errors at each training epoch. One can improve the accuracy of model and the whole control system. We have demonstrated by the experimental studies that the proposed technique gives better results when compared to the existing studies.