Study of a machine learning based methodology applied to fault detection and identification in an electromechanical system
Document typeBachelor thesis
Rights accessOpen Access
This work addresses the application of two di erent methods in order to detect faults in bearings operating within an electromechanical system, based on the measurement of vibrations and stator currents. The electromechanical system considered is a shaft connected to an electric induction motor. Two bearings are mounted on the shaft; these bearings can be metallic or ceramic. The bearings can be found in three di erent conditions: healthy or with a inner race hole of 1 mm or 2 mm. First of all the analisys of theorical fault frequencies was explored. The goal of this method is to identify theorical fault frequencies, depending on features of the bearing, in order to verify the presence of peaks in the frequency signals obtained from laboratory measurements. The accuracy of the theoretical frequency calculation was demonstrated by the actual presence of these peaks in the frequency signals, however it was expected to be found a proportion between the peak heights, and the severity of the fault, but this didn't happened. That led to the development of the second method, based on the building of a neural network able to classify the bearings with respect to their conditions, starting from 15 di erent statistical time domain features as input. Two reduction technicques, LDA and PCA, were implemented in order to reduce the number of input to the two most signi cant features; after that the neural network was built. The results obtained with this second method are very satisfactory as they allow to classify with a good performance both considering the di erent scenarios of bearing material and measured signal taken individually, but also considering all four di erent scenarios at the same time.
SubjectsElectromechanical devices, Bearings (Machinery), Neural networks (Computer science), Dispositius electromecànics, Coixinets (Maquinària), Xarxes neuronals (Informàtica) -- Aplicacions