Rolling bearing elements is a key component that plays an important role in maintaining a rotating machine overall health condition. Failures in these components lead to huge financial losses and unplanned maintenance disrupting the continuous production. These failures are a result of a sequential degradation and, the condition monitoring of these elements helps to detect their anomalous behavior. The vibrations of the bearings change due to the degradation and this characteristic feature of bearings can be leveraged to identify their behavior. In this thesis, a condition monitoring system is developed based on autoencoders which is a type of artificial neural network that learns the representation by reconstructing the input data. By training the autoencoders with data from normal working conditions of the component, it learns the representation in the data. And, for a data with deviated behavior, it gives a high reconstruction error indicating the fault in the component.
However, several systems have been developed previously to monitor the components. But, autoencoders are more promising for high dimensional and correlated data and thus it is further investigated. In this thesis, the features are extracted in the frequency domain using Fast Fourier Transform (FFT) for an unlabeled dataset and then a condition monitoring system is developed based on autoencoders. The developed system is later compared with a traditional Principal Component Analysis (PCA) system.