SVD and statistic theory based modified TPLS

Abstract

Modern industrial system is becoming more and more complex in order to produce the goods with high quality or achieve the functional requirements set by human beings. However, once the faults occur in the system, it's highly possible that the financial losses and even the operators' death may be caused. Therefore, it's necessary to improve the reliability of the system. The data-based fault diagnosis scheme is an important approach to realize the fault-tolerant control to further secure the system operation in the normal condition. This paper concentrates on the multivariate statistical analyses included in the framework of data-based scheme, more specifically, Total Projection to Latent Structures (TPLS). Although the traditional TPLS has achieved effective monitoring results in some practical applications, it should be noted that the decomposition principle of process variables is not appropriate. Furthermore, the test statistic it chooses can not reflect the subspaces they monitored. Both of the weaknesses make TPLS useless in some circumstances. To solve the problems, this paper proposes a Modified TPLS (MTPLS) based on TPLS, statistics theory and matrix analysis. Compared with TPLS, MTPLS has better fault diagnosis performance. A numerical example is used to validate the effectiveness of TPLS.