Machine Learning Techniques for Fault Diagnosis of Rotating Machines Using Spectrum Image of Vibration Orbits

Abstract

A comparative analysis of machine learning techniques for rotating machine faults diagnosis based on vibration spectra images is presented. The feature extraction of dierent types of faults, such as unbalance, misalignment, shaft crack, rotor-stator rub, and hydrodynamic instability, is performed by processing the spectral image of vibration orbits acquired during the rotating machine run-up. The classiers are trained with simulation data and tested with both simulation and experimental data. The experimental data are obtained from measurements performed on an rotor-disk system test rig supported on hydrodynamic bearings. To generate the simulated data, a numerical model of the rotating system is developed using the Finite Element Method (FEM). Deep learning, ensemble and traditional classication methods are evaluated. The ability of the methods to generalize the image classication is evaluated based on their performance in classifying experimental test patterns that were not used during training. The obtained results suggest that despite considerable computational cost, the method based on Convolutional Neural Network (CNN) presents the best performance for classication of faults based on spectral images.