Adaptive Anti-Over-Fitting Soft Sensing Method Based on Local Learning

Local learning based soft sensing methods are effective in dealing with process nonlinearities as well as time varying characteristics. In this paper, an anti-over-fitting method is proposed for appropriate online local model adaptation. The proposed method is based on the weighted sum of the predicted errors for the newest few samples, the weights of which are determined adaptively. Moreover, to reduce the online computational load and memory cost, we propose two adaptive process states division schemes which consider the influence of both the variance and mean value of the predicted residual. Two case studies on continuous stirred tank reactor and debutanizer column demonstrate the effectiveness of the proposed soft sensing scheme.