Soft sensor modeling based on auxiliary error probability density function shape control

Abstract

Soft sensor has been widely used for estimating product quality or other important process variables when online analyzers are not available. In order to cope with estimation performance deterioration when process variables abruptly change, a new soft sensor modeling method based on auxiliary error neuro-fuzzy model is proposed. The model mean square error (MSE) is used as an evaluating index in traditional data-driven modeling, while only seeks the minimum error function from the vision of a single sample without considering the spatial state of model error data points. To overcome this shortcoming, the auxiliary error model and probability density function (PDF) are combined to adjust the model parameter by controlling auxiliary error PDF shape to track a given target PDF. Furthermore, soft sensor model parameters are determined by means of gradient descent method. The actual operation process data of coal-fired power plant are selected as the modeling data to justify the effectiveness of the proposed method, experimental results show that the prediction accuracy and generalization ability of combining auxiliary error neuro-fuzzy model (AENFM) and PDF-based soft sensor modeling method are superior to other three data-driven methods using MSE criterion. The results indicate that the proposed method can be applied to soft sensor modeling of complex nonlinear system.