Data-Driven Design for static Model-Based fault diagnosis

This paper focuses on Data-Driven Design FOR Model-Based fault diagnosis, called D34MB for short. When the objective model is static, LVD (Latent Variable Detection) methods can be realized based on the LVE (Latent Variable Extraction) and LVR (Latent Variable Regression) techniques. A unified weight-framework for D34MB are proposed in this paper, which shows that all D34MB methods share the same procedures, i.e., LVE, LVR and LVD. The detection theorems shows that D34MB methods based on RRR (Rank Reduction Regression) and CCA (Canonical Correlation Analysis), compared with PCA (Principal Component Analysis) and PLS (Partial Least Square), tend to ensure higher calibration accuracy in terms of MSE as well as better detection performance in terms of FDR (Fault Detection Rate). In the case study, TEP (Tennessee Eastman Process) validates the correctness of our theoretical results.