A Review of the Application of Transfer Learning in Fault Diagnosis and its Potential in Aerospace Condition Based Maintenance

Condition based maintenance (CBM), maintenance that is triggered by knowledge of component degradation, relies heavily on fault diagnosis to pinpoint the component or system that requires maintenance. While there have been many advances in applying machine learning in fault diagnosis in recent years, there remains a problem of insufficient data with which to train machine learning algorithms. One approach to this problem is to reuse lessons learnt on one system on another system by transfer learning (TL). Previous reviews about the application of TL in fault diagnosis have concluded that TL is effective to cross-domain fault diagnosis problems through leveraging data from other working conditions or similar machines, and they systematically covered how various types of TL methods apply to different fault diagnosis problems. However, they did not consider what TL algorithms have never been applied to fault diagnosis that can benefit fault diagnosis research. Therefore, there is the necessity to comprehensively study TL in general and identify, from the whole scope of TL, any novel methods that may further facilitate fault diagnosis and aerospace CBM. By investigating into the history of TL, one such novel TL method found is high-level TL methods that enables knowledge transfer between both dissimilar source and target domains. Developing high-level transfer learning solutions in fault diagnosis would improve the current lack of diversity in the specific applications and domains of transfer in this field. Regarding the potential in aerospace CBM, high-level transfer learning is expected to significantly improve the efficiency of data usage.