For automatic detection/diagnosis of localized defects in bearings, a pattern recognition analysis scheme was developed for investigating vibration signals of bearings. Two normalized and dimensionless features are extracted by short-time signal processing techniques. Employing these two features, two linear discriminant functions have been established to detect defects on the outer race and rollers of bearings, respectively. Results of fault detection/diagnosis, based on the experimental data of imposed bearing defects, indicated the technique to be 14 percent better in the rate of success for the detection of defects than the best among the state-of-the-art. It takes 20 seconds for data processing and fault diagnosis on a PC-AT on-line implementation.