University of Science and Technology Beijing
The National Natural Science Foundation of China (61873024)
The characteristics of industrial process such as multivariate, high-dimensional data and nonlinearity complicate the quality monitoring and quality-related fault diagnosis. In this paper, we present a novel quality-related fault detection method for industrial process by combining kernel entropy composition analysis (KECA) and canonical correlation analysis (CCA) algorithm for feature extraction, which reduces the number of input space dimension and ensures the maximum correlation between the extracted features and quality variables simultaneously. Firstly, KECA algorithm is used to extract the features of the standardized data, in which phase, motivated by idea of CCA algorithm, canonical correlation is utilized to maximize the correlation between the extracted features and quality variables. Secondly, the monitoring statistics are constructed for process failure detection and the control limits are estimated via invoking a Parzen window density estimator. The proposed method was applied to the actual data of hot strip mill process (HSMP). Comparing with the performance of other four classical algorithms which are also suitable for nonlinear data, the results verify the accuracy, efficiency and advance of the method proposed.