Multimode Process Monitoring and Fault Detection: A Sparse Modeling and Dictionary Learning Method

This study focuses on the performance monitoring of a non-Gaussian process with multiple operation conditions. By utilizing the Bayesian inference technique, the proposed method, locality preserving sparse modeling, can automatically identify the current operation condition. Then, the feature of the data structure is extracted by locality preserving projections (LPP) and modeled by the sparse modeling technique. This hybrid framework of sparse modeling and LPP provides a robust and accurate paradigm for process data clustering and monitoring. The validity and effectiveness of this approach are verified by applying it to both a synthetic numerical example and the Tennessee Eastman process benchmark process.