Filter-based Online Neuro-Fuzzy Model Learning using Noisy Measurements

Neuro-Fuzzy (NF) model is capable of learning the nonlinear mapping between inputs and outputs accurately from training data and is thus a powerful tool for identification of nonlinear dynamic systems. However, when deploying the trained model, the noisy measurement leads to bias model predictions. Besides, training data is insufficient to cover the whole operating space for nonlinear systems. To well capture the system response, this paper proposes a recursive least squares algorithm to enable the NF model self-adaptive to different operating conditions whilst being robustness against measurement noise. Building on the data filtering technique and the auxiliary model theory, the proposed algorithm achieves high model prediction accuracy for online implementations. Efficacy of the algorithm is demonstrated by two simulation cases.