Data-driven, mechanistic and hybrid modelling for statistical fault detection and diagnosis in chemical processes

摘要

Research and applications of multivariate statistical process monitoring and fault diagnostic techniques for performance monitoring of continuous and batch processes continue to be a very active area of research. Investigations into new statistical and mathematical methods and there applicability to chemical process modelling and performance monitoring is ongoing. Successive researchers have proposed new techniques and models to address the identified limitations and shortcomings of previously applied linear statistical methods such as principal component analysis and partial least squares. This thesis contributes to this volume of research and investigation into alternative approaches and their suitability for continuous and batch process applications. In particular, the thesis proposes a modified canonical variate analysis state space model based monitoring scheme and compares the proposed scheme with several existing statistical process monitoring approaches using a common benchmark simulator – Tennessee Eastman benchmark process. A hybrid data driven and mechanistic model based process monitoring approach is also investigated. The proposed hybrid scheme gives more specific considerations to the implementation and application of the technique for dynamic systems with existing control structures. A nonmechanistic hybrid approach involving the combination of nonlinear and linear data based statistical models to create a pseudo time-variant model for monitoring of large complex plants is also proposed. The hybrid schemes are shown to provide distinct advantages in terms of improved fault detection and reliability. The demonstration of the hybrid schemes were carried out on two separate simulated processes: a CSTR with recycle through a heat exchanger and a CHEMCAD simulated distillation column. Finally, a batch process monitoring schemed based on a proposed implementation of interval partial least squares (IPLS) technique is demonstrated using a benchmark simulated fed-batch penicillin production process. The IPLS strategy employs data unfolding methods and a proposed algorithm for segmentation of the batch duration into optimal intervals to give a unique implementation of a Multiway-IPLS model. Application results show that the proposed method gives better model prediction and monitoring performance than the conventional IPLS approach.