(187i) An Inverse-Model-Based Methodology for Real-Time Fault Diagnosis in Non-Square Multivariate Dynamic Systems

Early and accurate fault diagnosis for chemical plants could save millions of dollars of losses due to process safety hazard. Most current diagnostic methods are data-driven and usually provide only fault detection (fault existence) and isolation (fault locations). Unlike the data-driven methods, the model-based diagnostic methods rely on a detailed system model which requires extra effort to build, but they provide not only fault detection and isolation, but also fault recovery (fault sizes and types). Traditional model-based methods such as the Luenberger observers provide asymptotic convergence between real and estimated state values, in which close reference tracking are guaranteed only after a certain time period. However, the reference tracking might not perform well at the initial time period. We hereby present new stable inverse algorithms which perform close reference tracking throughout the time period based on an inverse model. The discussions are confined in the non-square (number of inputs not equal to outputs) linear time-invariant (LTI) systems of which the feedthrough matrices in the state-space model does not appear. In order to create such an inverse model, an optimal dummy feedthrough matrix is designed based on the Lyapunov’s test for system observability. The efficacy and merit of the proposed methods are also demonstrated by a high purity distillation tower case study.