Bearing fault simulation and transfer learning
Aduwenye, Presley (2023)
Diplomityö
2023
School of Energy Systems, Konetekniikka
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2023062057315
https://urn.fi/URN:NBN:fi-fe2023062057315
Tiivistelmä
Significant number of breakdowns in rotating equipment can be attributed to defective bearings. Thus, the close monitoring of bearing condition has seen increased attention in recent times. The difficulty in sample acquisition associated with defective bearings is one of the bottlenecks stifling the advancement of effective and reliable diagnosis model for real-time bearing condition monitoring.
This thesis aims to simulate the behavior of common bearing defects using a dynamic model that was previously proposed. It also demonstrates how bearing diagnosis knowledge from simulation can be applied in real-world settings, bridging the gap between the two domains and ultimately resolving the pervasive sample acquisition problem.
A simplified CNN architecture was used in developing a pretrained model using simulation data, thereafter, several transfer learning techniques were tested and evaluated on the basis of parameter modification. The results showed remarkable performance in three out of four techniques investigated.
Since the performance and generalizability of data-driven models are tied to the quality and scope of the data used for their development, further studies were suggested in that regard. Overall, a diagnosis framework has been studied and presented which would serve as a base for future research.
This thesis aims to simulate the behavior of common bearing defects using a dynamic model that was previously proposed. It also demonstrates how bearing diagnosis knowledge from simulation can be applied in real-world settings, bridging the gap between the two domains and ultimately resolving the pervasive sample acquisition problem.
A simplified CNN architecture was used in developing a pretrained model using simulation data, thereafter, several transfer learning techniques were tested and evaluated on the basis of parameter modification. The results showed remarkable performance in three out of four techniques investigated.
Since the performance and generalizability of data-driven models are tied to the quality and scope of the data used for their development, further studies were suggested in that regard. Overall, a diagnosis framework has been studied and presented which would serve as a base for future research.