Gearbox faults feature selection and severity classification using machine learning

The most widely used technique for gearbox fault diagnosis is still vibration analysis. The need for gearbox condition monitoring in an automated process is essential and there is still a problem with the selection of features that best describe a fault or its severity level. For this purpose, multiple-domain vibration signals statistic features are extracted through time and frequency domain by postprocessing of raw time signal, time-synchronous average signal, frequency spectra and cepstrum. Five different datasets are considered with different levels of fault analyzing gear chipped and a missing tooth, gear root crack, and gear tooth wear under stable running speed and load. A preliminary experimental study of a single stage test bench gearbox was performed in order to test feature sensitivity to type and level of fault in the process of clustering and classification. Selected features were finally processed using an artificial neural network classifier.