Wear State Estimation of Rolling Element Bearings using Support Vector Machines

Failure of mechanical bearings is extremely problematic with respect to the reliability of electric and rotating machines for various applications and technologies, in transport, energy systems and advanced robotic systems (industry 4.0). The occurrence of catastrophic failure modes gives rise to abrupt unscheduled down time and maintenance, machine overheating (burn-out), as well as significant reliability and health and safety considerations for various mission critical end-applications. This paper proposes a robust method of remaining useful life (RUL) estimation using Support Vector Machines (SVMs) to classify the wear states of the rolling element bearings. Different types of feature sets are derived from the Short-time Fourier Transform (STFT) as well as Envelope Analysis; which are compared for both linear and non-linear wear state models. The methods were optimised for different types of SVM kernelling functions and achieved classification accuracy of up to 67.6%. Medium and coarse Gaussian kernelling functions achieved the highest level of accuracy. This proposed method proves to be a valuable non-invasive predictive CbM approach for electric and rotating machines, using accelerometer based vibration signals mounted on the external races of rolling element bearings under test.