A Novel Fault Detection Scheme Based on Difference in Independent Component for Reliable Process Monitoring: Application on the Semiconductor Manufacturing Processes

Abstract

Compared with principal component analysis (PCA)-based methodologies, independent component analysis (ICA) is more suitable for extracting the variation information of latent variables. The advanced fault detection approach based on ICA utilizes two statistics, I²_d and square prediction error (SPE), to monitor the health status of a single mode process. However, when these statistics are used for monitoring a multimode process, their fault detection performances are disappointed. Aiming to detect some faults caused by abnormal changes of latent variables in a multimode process, a novel health status monitoring scheme based on difference in independent component (Diff-ICA) is proposed in this paper. In Diff-ICA, the conventional ICA is firstly implemented to extract the latent information of a process. Next, the estimated independent components of each sample are calculated using k nearest neighbors (kNN) rule. After that, a new statistic I²_diff, which is built based on the difference between real independent components and estimated independent components, is used for monitoring the variation of latent variables in this process. Meanwhile, another statistic q_diff, which is built in the residual subspace that is reconstructed using these estimated independent components, is applied to monitor the health status of this process. Diff-ICA is capable of both reducing the impact of multimode characteristics on process monitoring and improving fault detection rate of a multimode process. The efficiency of Diff-ICA is demonstrated by a simulated case as well as an industrial case. The experimental results indicate that the proposed method outperforms PCA, kernel PCA (KPCA), ICA and FD-kNN.