KPCS-kNN based fault detection for batch processes

Machine learning technology is already used in a wide range of industrial processes. For nonlinear, multimode and non-Gaussian batch process, a improved fault detection method based on kNN is proposed. Firstly KPCA is introduced to get kernel principal components (KPCs) and to form modeling data set. Secondly, k nearest neighbors of each KPC are found and the sum of k nearest neighbor squared distances is computed as monitoring statistic for each KPC. Then kernel density estimation method was used to set statistical threshold of the monitoring statistic. KPCs can extract nonlinear feature of batch process and its dimensions are lower than the one of raw data. So using KPCs to build model can reduce computation complexity and memory requirement. kNN can solve nonlinear and multi-modes questions of batch process. The use of kernel density estimation method can solve non-Gaussian characteristics of modeling data. In order to comparison, MPCA method, FD-kNN method, PC-kNN method and the proposed methods are applied in the semiconductor industry process. The experiment results show the good performance of the proposed method.