Multimodal Data Fusion Using Multivariate Empirical Mode Decomposition for Automatic Process Monitoring

Abstract

The recent leap forward in wireless sensing and communication technologies has brought forward the proliferation of multimodal sensing data and offered a unique viewpoint to study numerous complex systems and processes. Indeed, the mining of such sensing data is poised to promote the paradigm-shift from experience-based to evidence-driven process monitoring, control, prognostics, and optimization. Notably, an integration and effective fusion of such multimodal data has been an impending obstacle to utterly marshal their prowess. Despite the extensive attempt in the literature, multimodal data fusion is still in the embryonic stage, particularly considering the inherent nonlinear, nonstationary, and heterogeneous nature of the sensing data. In this article, we report a novel approach to fuse multimodal data from different channels via nonparametric multivariate empirical mode decomposition (EMD). The EMD decomposes each signal into a series of oscillatory modes called the intrinsic mode functions (IMFs) with a progressively lower frequency. The multivariate EMD guarantees that signals from different channels have the same number of components and are mode aligned. Capitalizing on the recurrence quantification analysis of the intrinsic state space reconstructed from the IMFs, we build a control chart to monitor process dynamics. Applications in ultraprecision machining processes suggest the effectiveness of the proposed algorithm.