Adaptive resonance theory (ART) is widely used in real time monitoring and fault diagnosis of mechanical equipment. Traditional ART adopts hard competitive learning mechanism always leads to erroneous when only one winning node is allowed to be learned. In the light of this problem, the Yu’s norm similarity criterion, the lateral inhibition theory and fuzzy ART are combined to establish a soft-competitive learning ART (Soft-ART) mode, which allows multiple winning nodes to be learned. To improve the diagnosis success rate, an improved feature selection technique based on Yu’s similarity criterion is also proposed. Moreover, the feature selection technique and proposed Soft-ART mode are tested by bearing fault data. Compared with FCM, BP and Fuzzy ART, the Soft-ART has a higher diagnostic precision.