Contribution to Fault Detection and Diagnosis: Mixture-based Modelling
by Tatiana V.Guy, Dani Juricic, Miroslav Kárny and Andrej Rakar
A multi-model approach complemented by probabilistic mixture modelling has been applied to solve fault detection and isolation problem. The resulting algorithm provides efficient tool for process condition monitoring.
Complexity of modern manufacturing and process industries emphasises the importance of the fault detection and isolation for the purpose of process condition monitoring. Unlike occasional inspections, periodic checks, etc, on-line condition monitoring seems to be the most promising way to decrease the number of unanticipated system shutdowns while optimising the production costs. Extension of recently developed probabilistic mixture-based approach to the domain of systems' fault detection and isolation provides an adequate support to solution of process monitoring problem. Performance of the approach was verified on a real process of gas conditioning.
Within an EC IST-99-12058 project ProDaCTool, a powerful generic tool for describing and optimising complex uncertain multi-dimensional systems by multiple models was developed and successfully applied. Hypothesis on existence of several operational regimes - normal as well as faulty - makes multi-model approach a natural framework for solving the fault detection and isolation (FDI) problems. This has motivated us to test its applicability to FDI domain.
The whole framework relies on a probabilistic mixture modelling that provides suitable description of non-linear dynamic systems. Each of the operation modes as well as each of the faulty states is represented via probabilistic model. A convex probability density functions &Mac246; called components - describing respective normal dynamic regressions, each associated within some operation mode, models the process data. Quasi-Bayes estimation provides the model of the supervised process. Provided rich process data exist, the model learned represents all the operation states occurred in the past, each associated with the corresponding component. The probability of occurrence of a particular operational state is expressed by the weight of the respective component. During on-line operation, the learned components' weights are updated using available process measurements and exploited for effective estimation of the actual operation state. A proper recommendation is then generated in a timely fashion. Effectiveness of the outlined approach has been verified on a gas-conditioning unit, a part of an industrial-scale pilot installation in Josef Stefan Institute for treatment of technological wastewater. Real-time diagnostic results show that each component correctly represents faulty states of the system. Apart from short transient periods, during changes of process states, all faults were successfully detected and uniquely isolated. It is important to note that the presented diagnostic approach relies on availability of rich past data. These should also include all faulty states that have to be diagnosed. Unfortunately this is not always the case in real applications, especially if a diagnosis system is designed from scratch. The on-line applicability of the quasi-Bayes estimation allows to design adaptive FDI version. The available self-learning capability will enable to consider newly met faulty states.
Tatiana V. Guy, Institute Information Theory and Automation, Academy of Sciences of the Czech Republic /CRCIM
Tel: +420 2 6605 2254