This paper proposes a new bearing fault detection framework that is based on multivariate statistical process control methods. In this framework, historical offline normal data are used to train the models and calculate the control limits of the monitored metrics. Then, bearings’ new online data are the input to the trained models to obtain their monitoring metrics, which are compared with the control limits to determine the healthy status of bearings. Unlike most conventional methods, the proposed framework does not need any faulty data at the training stage. Therefore, the proposed framework is flexible and applicable to most practical cases in which few or even no faulty data are available at the training stage. Two bearings’ life data sets are used to validate the proposed fault detection approach. Results show that the higher order cumulants analysis-based approach exhibits better performance in bearing fault detection when compared with principal component analysis-based and independent component analysis-based approach.