Rolling element bearing is one of the most critical components in rotating machinery, whose failures can result in costly downtime, equipment damage and even catastrophic accidents sometimes. In the last decades, signal based approaches have become the mainstream in fault detection and diagnosis of bearings. They are normally model-free, but many of them require a good deal of expertise and human intervention to apply them successfully. Thus alternative methods should be developed for bearing fault detection and diagnosis. This work is devoted to studying the bearing fault detection and estimation issues in the framework of the model-based FDI techniques. To this end, dynamic modeling of rolling element bearing systems is presented first in this work. Then, a model-based fault estimation scheme for rolling element bearings is proposed, where an H2 optimal observer is designed to estimate the impulse fault force in bearings and the energy level of fault force. Besides, an adaptive least square observer-based residual generator scheme is proposed for rolling element bearing systems, which can be further used for bearing fault detection. Finally, to tackle the problem of lacking access to collect real fault data in large equipment, a blind identification approach is proposed for a three-channel system with partial common unknown inputs for purpose of data fusion to generate synthetic fault data. The proposed approach allows to fill in the gaps of fault states database for large equipment in the case of lacking access to collect real fault data.