Diagnostics of rolling element bearing using transfer learning

The Case Western Reserve University (CWRU) bearing diagnostics dataset is a benchmark dataset created to aid in the test and development of new classification methods for identifying bearing faults. In recent years, transfer learning has come to the forefront of deep learning endeavors as a viable avenue for research in the development of new algorithms. The goal of this thesis is to investigate avenues by which to enhance the methods of rotor diagnostics using artificial intelligence methods. As models have been created, one issue that has not been thoroughly investigated is the efficacy of transfer learning methods on bearing diagnostics data and the number of samples, needed from the CWRU dataset for proper classification. The primary research question to aid in the completion of our goal is can the sample size impact the classification accuracy of a transfer learning model trained on experimental bearing data? The specific aims of the study are to investigate the impact of load, sampling frequency, the location of the accelerometer on accuracy classification. It was found that all three vectors of the study affected the classification accuracy with 60 samples being sufficient for a 100% classification accuracy in all three studies.