Data-Driven Design of Fault Diagnosis Systems

Due to the increasing demands on system performance, production quality as well as economic operation, modern technical systems become more complicated and the automation degrees are significantly growing. To ensure the safety and reliability of such complicated processes, an effective fault diagnosis system is of prime importance in process industry nowadays. Although the model-based fault diagnosis theory has been well established, it is still difficult to establish mathematical model by means of the first principles for large-scale process. On the other hand, a large amount of historical data from regular sensor measurements, event-logs and records are often available in such industrial processes. Motivated by this observation, it is of great interest to design fault diagnosis schemes only based on the available process data. Hence, development of efficient data-driven fault diagnosis schemes for different operating conditions is the primary objective of this thesis. This thesis is firstly dedicated to the modifications on the standard multivariate statistical process monitoring approaches. The modified approaches are considerably simple, and most importantly, avoid the drawbacks of the standard techniques. As a result, the proposed approaches are able to provide enhanced fault diagnosis performance on the applications under stationary operating conditions. The further study of this thesis focuses on developing reliable fault diagnosis schemes for dynamic processes under industrial operating conditions. Instead of identifying the entire process model, primary fault diagnosis can be efficiently realized by the identification of key components. Advanced design schemes like multiple residuals generator and state observer are also investigated to ensure high fault sensitivity performance. For the large-scale processes involving changes, e.g. in operating regimes or in the manipulated variables, the recursive and adaptive techniques are studied to cope with such uncertainty issues. A novel data-driven adaptive scheme is proposed, whose stability and convergence rate are analytically proven. Compared to the standard techniques, this approach does not involve complicated on-line computation and produces consistent estimate of the unknown parameters. To illustrate the effectiveness of the derived data-driven approaches, three industrial benchmark processes, i.e. the Tennessee Eastman chemical plant, the fed-batch fermen-