Neural network based classification of bearing faults in rotating machines: Augmentation of vibration measurements with power measurements

This thesis investigated the integration of power consumption measurements with condition monitoring data to enhance predictive diagnostics on rotating machines. More specifically, it focused on rotating machines operating at constant speeds, and the diagnosis of bearing faults on such machines based on vibration measurements, with or without augmentation by power consumption measurements. Power consumption monitoring is becoming more prevalent due to rising energy costs. Vibration measurements and power consumption measurements were collected for a range of known bearing fault conditions, under controlled operating conditions, through laboratory experimentation. A large number of neural network based diagnostic classifiers were implemented. The performance of these classifiers, using feature vectors derived from vibration measurements alone, was investigated with respect to recognition, interpolation, and extrapolation of bearing fault. The same set of classifiers was then tested with augmented feature vectors, consisting of the concatenation of the vibration based feature vector and the magnitude of the measured power consumption. Comparison of the two sets of results showed that, for the case of multiple accelerometer signals, i.e. sensor fusion, measured power consumption was a useful feature with which to augment the neural networks for achieving detailed diagnostic classification of bearing faults: it enhanced extrapolation capability at the cost of a slight degradation in interpolation capability, while preserving recognition capability.