Diagnosing localized and distributed faults of rolling bearing using kurstogram and machine learning algorithms using bearings audio signal in comparison with vibration signal

Abstract

Condition monitoring of rolling bearing using bearing’s audio signal and vibration signal via cost-effective accelerometer is experimented with and analyzed for both localized faults and distributed/ generalized roughness faults. Even though the Fast Fourier Transform (FFT) of the bearing’s audio signal is not appropriate to diagnose bearing faults under varying background noises it was possible to observe characteristic frequency hikes related to misalignment from the FFT of the vibration signal. Localized faults are processed using Kurstogram, and Hilbert transform. Misalignment experiments for different types of bearings at different speeds and related fault frequencies are identified through both bearing audio signal and vibration signal. Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) are trained using the Mel Frequency Cepstral Coefficient (MFCC) feature of bearing audio signals with various background noises and/or the FFT of the vibration signal to distinguish between healthy bearing and bearing having distributed /generalized roughness faults. The ANN and CNN models trained using fused MFCC of bearing audio signal and FFT of the vibration signal yield greater accuracy than that of models trained using MFCC of the audio signal or trained using FFT of the vibration signal