Bearing fault detection with vibration and acoustic signals: Comparison among different machine leaning classification methods

Abstract

Despite the recent advances in supervised ML-based methods for fault bearing detection is that most published work uses only vibration data for damage detection. However, depending on the type of bearing failure and the frequencies of the signals, the combination of non-invasive techniques such as vibration signals and acoustic signals can be an alternative to increase the precision when detecting a failure. In this research, simulated faults were created in a healthy bearing. Most of the values from defect areas are between 0.007 and 0.014 in. of diameter. A database was generated from these bearing faults, so that the acoustic and vibratory signals of the same fault condition could be compared, and a better diagnosis could be made. For this, an experimental system was designed. The present work, therefore, describes and compares three different supervised Machine Learning-based methods based on acoustic and vibration data. So far, there have been two approaches to obtain data. For this work, acoustic and vibration data were obtained from an own experimental system, integrated by transducers (microphone and a tri-axial accelerometer), data system acquisition, and a signal recording system. The three different methods proposed in the present work achieved a fault classification above 96%, thus presenting a new versatile approach to enhance predictive maintenance methodology in a rotary machine with vibration and acoustic signals.