Transfer learning for process fault diagnosis: Knowledge transfer from simulation to physical processes

Abstract

Deep learning has shown great promise in process fault diagnosis. However, due to the lack of sufficient labelled fault data, its application has been limited. This limitation may be overcome by using the data generated from computer simulations. In this study, we consider using simulated data to train deep neural network models. As there inevitably is model-process mismatch, we further apply transfer learning approach to reduce the discrepancies between the simulation and physical domains. This approach will allow the diagnostic knowledge contained in the computer simulation being applied to the physical process. To this end, a deep transfer learning network is designed by integrating the convolutional neural network and advanced domain adaptation techniques. Two case studies are used to illustrate the effectiveness of the proposed method for fault diagnosis: a continuously stirred tank reactor and the pulp mill plant benchmark problem.