A Machine Learning approach to multicomponent fault diagnosis of rotating Machines using sound and vibration signals

Abstract

The objective of the study is to identify the feature-classifier pair for automated multi component fault diagnosis of the rotating machine by applying machine learning method to sound and vibration signals. Centrifugal pump, turbine, compressor, electric motor, generator etc., belong to the category of rotating machines. Maintenance of this critical machine is important in the industry to reduce the unexpected machine breakdown or downtime, productivity loss, economic loss and also to enhance the human safety. The predictive maintenance is much practised in many of the industries for maintenance decision making. The predictive maintenance or condition based maintenance determines the need for maintenance by monitoring the condition of machine continuously or periodically. The information about the machine is obtained by monitoring suitable variables and fault can be diagnosed. The fault diagnosis methods are vibration monitoring, wear debris analysis, acoustic emission, ultrasonic detection, sound monitoring etc. Each method has its own advantages in specific applications. Among these, vibration monitoring is the widely accepted technique for rotating machines. Sound signal can also be used like vibration in a cost effective way. Spectral or frequency domain signal analysis is the basic approach for vibration signal monitoring. Cepstrum analysis, envelope analysis, octave analysis, demodulation analysis, order analysis etc., are a few of the advanced vibration analysis techniques practiced in the industry with the help of highly skilled and experienced technician. The fault diagnosis process should be automated using computational intelligence for easy identification of the faults. The machine learning techniques are the best choice for automated fault diagnosis. The machine learning algorithm first learn and store the signal pattern of various faults and then it monitors the machine by mapping the acquired signal pattern with the various fault pattern in its memory.