A Study on the Use of Simulation Data for Data-Driven Fault Diagnosis of Various Rolling Bearings Using Transfer Learning

Rolling bearings are key components of numerous engineering systems and are subject to wear due to the mechanical contacts. Consequently, bearing fault diagnosis is imperative for the reliability and efficiency of these systems, such as rotating machinery. This paper explores the utilization of simulation data for training data-driven fault diagnosis methods. To this end, a self-developed bearing simulation and self-collected measurement data from test rigs are employed, considering varied operating conditions and bearing types. The study evaluates the effectiveness of simulation data in improving the diagnosis performance of real bearing faults. In particular, transfer learning methods are examined, encompassing both inductive and transductive transfer learning approaches, implemented with three types of neural networks. The findings demonstrate the effectiveness of the developed simulation model in generating data that is conducive to fault diagnosis. Already the training with simulation data alone indicates the potential benefits of incorporating simulation data. The study further demonstrates that inductive transfer learning exhibits superior performance in comparison to training with real measurement data alone. However, no improvements are achieved through transductive transfer learning.

Keywords: PHM, Prognostics and health management, Fault diagnosis, Data-driven methods, Simulation data, Measurement data, Transfer learning, Rolling bearing, Similar systems, Operating conditions.