Lifetime Based Health Indicator for Bearings using Convolitional Neural Networks
MetadataVis full innførsel
Out of all the components in rotating electrical machinery, bearings have the highest failure rate. Bearingdegradation is a seemingly random process which is hard to both model and predict. Countless of con-dition based methods and algorithms have been proposed in order to accurately diagnose incipient faultsand estimate the remaining useful lifetime of bearings. These methods are often complex and hard to im-plement. In this thesis, a data-driven method of estimating a linear lifetime based health indicator (HI)using convolutional neural networks (CNNs) is proposed. The idea behind the method is to train a CNNmodel to recognize the shapes and distributions of vibration data in order to predict a HI with minimalpre-processing. Two models are presented: A CNN that takes time-series vibration data as input and aCNN that takes vibration frequency spectrum data as input. Finally, HIs are predicted on unique datasetsand their respective remaining useful lifetimes (RULs) are estimated as part of the model validation process.The results show that the models are able to recognize relevant fault features to a certain degree. However, accurate predictions have proven difficult in many cases.
Master's thesis Renewable Energy ENE500 - University of Agder 2019