Machine-Learning Based Fault Diagnosis for a Rotordynamic System Using Multibody Simulations

Abstract

Machine-learning based fault diagnosis plays an important role in condition monitoring for rotating machinery to prevent systems from catastrophic faults. It is important to note that the performance of data-driven methods relies highly on a large quantity of training fault data. Since rotating machinery operates under normal condition most of the time, collecting sufficient fault data from experiments takes a huge amount of time and expense, and under various operating conditions. To overcome the fault data insufficiency, building a virtual testbed for generating fault data is a promising way in bridging the gap between data requirement and prediction accuracy. Many simplified dynamic models have been proposed to generate a single fault on some rotordynamic systems. These methods, however, cannot reflect complex operation conditions such as variant rotation speed or multi-faults. To better reveal vibration responses of local defects, this research aims to establish a multibody dynamics (MBD) model that can simultaneously analyze complete dynamic behavior and simulate a wider range of fault scenarios. In this research, a simulation-driven fault diagnosis method is proposed to generate the simulation fault data. Firstly, a rigid-flexible hybrid model of a single-rotor-bearing system is established using MSC ADAMS, which is based on MBD and finite element analysis. Different fault conditions are simulated including outer race bearing faults, inner race bearing faults, and rolling element faults. After generating fault data, a timefrequency feature extraction method is developed based on Hilbert envelope and wavelet packet decomposition, extracting a large amount of features from the original signals. In addition, an autoencoder model is built to highlight the critical features, enhancing the performance of the classifier. This feature extraction is made to obtain fault-related features, which train the machine learning classifiers for discriminating the fault categories. To validate the simulation results, the Case Western Reserve University (CWRU) bearing dataset that has been widely accepted as a standard reference is introduced. A comparison of bearing fault frequencies between simulations and the CWRU dataset is then conducted. Meanwhile, a transfer learning method is applied using the CWRU dataset to fine tune the fault diagnosis classifier. This research lays a solid foundation for future development of a digital twin and simulation-driven transfer learning for fault diagnosis of rotating machines.