This paper presents a data fusion approach for the diagnosis of bearing faults under different seeded fault scenarios. The approach is based on the extraction of three simple and intuitive features that fuse the information that comes from two accelerometers placed at two different sites of the test bed. The analysis shows that in the case of the occurrence of a fault even in an early stage the “footprint” left at the scatter plot of the measurements coming from the two accelerometers can effectively turned into features/descriptors by simple statistical measures such as the elements of the covariance matrix. Those features when fed to a k-nearest neighbor classifier or an ensemble of one class detectors can lead to a remarkably high detection/diagnostic performance.