Robust Process Identification from Step Response Data and Parallel Implementation

Fast, robust and accurate system identification is of importance to the process industry, and identification from step response is a prevalent approach. Recently, a new method based on rank constraint using low-quality industrial data has been proposed. However, with mean square error (MSE) used as the loss function, this identification method is sensitive to outliers, which may occasionally lead to invalid models. In this paper, we propose an improved robust process identification approach from step response data based on the Huber loss, which is less sensitive to outliers than generic MSE, and leads to a higher successful rate. A tailored solution algorithm based on alternating direction method of multipliers is developed, which, however, requires heavy computational cost especially when there are massive control loops to be identified simultaneously. To address this issue, we leverage recent advances in parallel computing. We show that this solution procedure can be parallelized, which leads to significant computation savings with graphical processing units used, and thus better conforms to requirement in practical situations. Numerical studies demonstrate that our proposed method is more robust against outliers, and the parallel implementation gives a faster speed in the presence of massive data.