Machine Learning/Signal Processing

Project - 6 : Comparison of conventional method of fault determination with data-driven approach for ball bearing faults in a wind turbine gearbox

This project was conducted under the guidance of Dr. G R Sabareesh between August-December 2015. As part of this project, experimental acceleration data for healthy and faulty (inner and outer race faults) ball bearings was acquired using a Piezoelectric accelerometer in an in-house wind turbine gearbox test rig, shown below.

Experimental wind turbine gear box test rig at

BITS Pilani Hyderabad Campus

Following this, the experimental acceleration signals were de-noised using the wavelet de-noising techniques using the MATLAB Signal Processing toolbox. After de-noising, the Frequency spectrums were analyzed to identify the characteristic ball bearing frequencies using FFT plots and compared with the theoretical ball bearing defect frequencies. Further, statistical tools such as Principal Component Analysis (PCA), Zero-phase Component Analysis (ZCA) and wavelet transforms were used to pre-process the testing set to apply a Multiclass Support Vector Machine (MSVM) model for classification. This was done using the Statistics and Machine Learning MATLAB toolbox.

The procedure yielded significant enhancements in classification accuracy, because of which this work was submitted for review to International Journal of Performability Engineering and has been published by the Journal (Link)