Abstract
Fault diagnosis plays an important role in ensuring safe and efficient chemical production. This
study explores methodologies that integrate simulation and data analytics for fault diagnosis.
Such integration is motivated by the advances in modelling and computing capabilities and
strength of data-driven methods in utilising simulated and historical data. The project explores
three research avenues along this direction.
The first avenue considers to train deep learning model for classifying fault sample and use
simulated data to overcome the fault data scarcity. Its practical implementation faces two
challenges: 1) the discrepancy between simulation and physical domains; 2) the problem of
non-identical fault label sets across the two domains. We use advanced transfer learning
techniques to address them. Case studies show the feasibility of integrating computer
simulation and transfer learning for fault diagnosis, necessity of using special transfer learning
models in presence of non-identical fault label sets, and advantages of our models over
compared traditional alternatives.
The second avenue focuses on diagnosing both single and multiple faults and fault of different
magnitude, motivated by: 1) the occurred fault can have a different magnitude from simulated
or historical faults; 2) diagnosing multiple faults is an important but challenging problem. For
this purpose, we propose an approach that integrates pattern matching and online simulation.
Case studies show its effectiveness in fault diagnosis and estimating the magnitude and
occurrence time of fault.
The third avenue focuses on Active Fault Diagnosis (AFD) for enhancing diagnosability. In AFD,
system inputs are deliberately designed to increase the discriminative information in the data,
resulting in improved diagnosis performance. We discuss a new AFD method which considers
fault magnitude uncertainty, constrained state estimation, and decentralized implementation
for reducing computational complexity. Experiments demonstrate its performance over
passive diagnosis and traditional AFD method. Its decentralized implementation presents
slightly degraded but acceptable performance.