On-line detection of ball bearings can improve product quality and enhance productivity. Three features, including peak amplitude of the frequency domain, percent power, and peak RMS, have been extracted from the radial acceleration of ball bearings. The Sequential Forward Search (SFS) algorithm has been applied to select the best vibration features. Adaptive Neuro Fuzzy Inference Systems (ANFIS) have been used. A 2 x 2 ANFIS using the pi-shaped built-in membership function can distinguish normal bearings from defective bearings with 100% reliability. Furthermore, a 3 x 5 ANFIS can classify ball bearings into six different conditions with a success rate of over 95%. In simple words, on-line detection of ball bearings can be performed successfully using SFS and ANFIS.