Context Aware Anomaly Detection for Condition Monitoring of Rolling Element Bearings

Anomaly detection based on machine learning methods has been an active area of research for condition monitoring of rotating machinery. However, the machine learning anomaly detection approaches so far have been mainly targeting a single operating condition of the machine. In this study the introduction of the rotational speed of the machine as a context into a Deep Support Vector Data Description (DeepSVDD) anomaly detector is examined. In the considered case, the rotational speed changes in discrete steps and is kept constant over the period of interest resulting in discrete contexts. Several different methods of introducing such a discrete context into the DeepSVDD model are examined. First of all, the use of a separate anomaly detector per context is considered. Next, the context information is introduced in the preprocessing stage by performing resampling. Finally, the context is introduced in the model structure by considering a multi sphere DeepSVDD where every sphere is related to a speed setpoint of the machine. To aid the model in the multi sphere DeepSVDD case different ways of encoding the context are evaluated. Every context aware model is compared to a baseline DeepSVDD model where the context is not taken into account. The methodologies are applied, validated and compared on a multiclass bearing fault dataset, where special consideration is taken that the training and testing data do not have any overlap and are coming from different assemblies of the setup such that the results are generally applicable.