A data-driven approach for on-line auto-tuning of minimum variance PID controller

A data-driven approach for on-line tuning of minimum variance (MV) PID controller is proposed in this paper for a linear system subject to stochastic disturbances, in which none of a prior knowledge or/and external excitation signals is required. The main procedure is that two different rough tuning controllers are employed and switched from one to another such that two sets of output data are collected under routine operating conditions. Subsequently, based on FCOR (Filtering and CORrelation analysis) algorithm, the corresponding linear MV controller is estimated on-line for the linear system. The parameters of MV-PID controller is tuned to approximate the estimated MV controller by means of solving an optimization problem subject to a constraint of the closed-loop stability, where the weighted penalty function is composed of the inverse of controller parameters and the difference between the proposed controller and the minimum variance controller. By using a different selection of the weighting coefficients in the penalty function, the final tuning parameters of MV-PID controllers are determined by the practical consideration of step disturbance attenuation or sometimes the trade-off between stochastic and step signal disturbance attenuation.