Kernel based manifold learning for complex industry fault detection

Abstract

Accurate and rapid fault detection based on the data from industry process is very important for the process control. This paper introduces a new multivariate statistical process control approach for fault detection using kernel method based manifold learning algorithm combining T2 statistic. The proposed approach is effective in the fault detection, which has two stages. Stage I: a kernel method based locally linear embedding is employed to extract the nonlinear features, preserve local structure and reduce dimensionality of the multivariate input data, and a new low-dimensional embedding method is developed to solve the "out-of-sample" problem. Stage II: the fault detection is performed by T2 statistic with control limits derived from the eigen-analysis of the kernel matrix in the Hilbert feature space. In this study, the method is applied for the fault detection of the benchmark Tennessee Eastman (TE) challenge process. The proposed method has been compared with conventional methods in terms of performances such as detection accuracy, detection delay and false alarm rate. It is demonstrated that the proposed method outperformed the others in fault detection on TE process. © 2013 Springer-Verlag.