Closed-loop identification of LPV models using cubic splines with application to an arm-driven inverted pendulum

A method for the identification of MIMO input-output LPV models in closed-loop is proposed. The model is assumed to display both linear and non-linear behavior in which the latter is dependent on the scheduling parameters, and cubic splines are used to represent the non-linear dependence. For the estimation of both linear and non-linear parameters, the separable least square method is employed. The linear parameters are obtained by a least square identification algorithm, while the non-linear parameters are obtained using a recursive Levenberg-Marquardt algorithm. To identify such a model in closed-loop, we use a non-linear version of a two-step method. A neural network ARX model will be used in the first step for two purposes. Firstly, to generate noise-free input signal to get an unbiased model and secondly to generate noise-free scheduling signal for consistent identification. The proposed method is applied to an arm-driven inverted pendulum. The resulting model is compared with a linear time-invariant model, and with an LPV model that depends polynomially on the scheduling parameters. Experimental results indicate that the cubic spline model outperforms the other ones in terms of accuracy.