Neural-Network-Based Fuzzy Observer With Data-Driven Uncertainty Identification for Vehicle Dynamics Estimation Under Extreme Driving Conditions: Theory and Experimental Results

We present a neural network based Takagi-Sugeno (TS) fuzzy observer to estimate the lateral speed (or sideslip angle) of nonlinear vehicle dynamics subject to modeling uncertainties and unknown inputs. To this end, we first propose a TS fuzzy reduced-order observer design, which can be implemented with low computation effort, for nonlinear systems. The stability and robustness of the observer scheme against the modeling uncertainty is guaranteed by the
$\mathscr {H}_{\infty }$
filtering method. A data-driven approach is proposed to construct feedforward neural networks (NNs) for uncertainty approximation. This data-driven approach exploits a specific sliding mode observer (SMO) to identify the model uncertainty data from the collected training data. The NN-based uncertainty approximation is incorporated into the TS fuzzy observer structure to mitigate the effect of uncertainty and improve the estimation quality. Via Lyapunov's stability theory, design conditions of both the TS fuzzy reduced-order observer for dynamics estimation and the SMO for uncertainty identification are derived in terms of linear matrix inequalities. Experimental results obtained with the INSA autonomous vehicle on a real test track demonstrate the effectiveness of the proposed TS fuzzy observer under various driving scenarios. Performance comparisons are also performed to illustrate the interest of using NN-based uncertainty approximation for the new reduced-order observer scheme, especially under extreme driving conditions.