Fault sound simulations from normal sounds for data-driven prognosis based on human expert and vibration knowledge

This paper discusses how to simulate plausible machine fault sound data from normal machine sounds using a combination of human empirical knowledge and past case studies of vibration monitoring. We focus on the fact that human experts in many maintenance areas generally listen to sounds when monitoring machine conditions. They can determine machine conditions by listening. Therefore, it is possible to assess machine fault sound simulators based on whether humans find the sound plausible. We successfully generated abnormal sounds that were synthesized from normal machine sounds using several parameters. The simulated sounds were assessed in two steps. One used theoretically synthesized sounds at various loudness levels and a vibration knowledge database. The other involved assessment by human experts by ear. We successfully generated plausible sounds for three rotor rotation machine failures and four roller bearing faults. The sounds were assessed by three human experts who had more than 10 years of experience in machine maintenance in two different areas. Their assessment results were consistent for our parameters.