Nowadays, vibration monitoring system (VMS) using machine learning has been increasingly used to predict rotor faults. However, a sufficient amount of fault data is harder to collect practically than the normal data, as a result, the imbalanced training data set can significantly affect the accuracy of the trained classifier. In this paper, we proposed a data augmentation approach that uses physics-based high-fidelity dynamics simulation as an alternative to acquiring practical fault data. The overall procedure includes: (1) The high-fidelity numerical simulation model reflecting the behavior of rotor to obtain the vibration signature of fault data; and, (2) data augmentation with the simulation fault data in conjunction with experimental normal and fault datasets. A fully connected Neural Network (FCNN) is applied to build the classification model that identifies rotor faults focusing on mass unbalance. The rotor system considered in this study consists of rigid discs, shaft, eccentric mass, and bearing housings. The numerical simulation model in this work considers high-fidelity physical behaviors such flexible multibody dynamics having centrifugal force and gyroscopic effects. Time domain data of the vertical and horizontal vibration responses of bearing housings are obtained from simulation and then FFT is applied to extract the main feature in frequency domain, which is the amplitude of the 1X harmonics of the vibration responses. The data augmentation is accomplished with frequency domain data both from simulation results and experimental acquisition. This approach can tackle data imbalance problem which is one of the most critical hurdles in neural net-based. fault diagnosis. From experimental verification, high accuracies more than 90 % of rotor fault diagnosis, which demonstrates the effectiveness of the proposed framework compared to the model with insufficient fault data.
|