Fault Detection Method based on Incrementable Laplacian Eigenmaps and Normal Space

Abstract

Aiming at the difficulty of embedding new samples added and the loss question of information when using Laplacian Eigenmaps (LE) for process monitoring, a fault detection method based on Incremental Laplacian Eigenmaps and Normal Space (ILENS) is developed. After feature extraction of the training data using Laplacian Eigenmaps, the local projection matrix is constructed from the local information for embedding new samples. It ensures that normal samples can be embedded in the manifold of the training data in the feature space and that most fault samples can be separated from normal samples. Then Normal Space of the manifold is constructed based on the local information of the samples. In the Normal Space, a small number of fault samples falling into the low-dimensional manifold can be separated from normal samples. The ILENS method was compared with PCA, KPCA, FD-KNN, and RP-KNN through a numerical simulation process and a turbocharged spark-ignited engine system simulation process. The results show that ILENS possesses a higher fault detection rate compared to other classical methods.

DOI: 10.61416/ceai.v26i1.8932