In this paper, a novel similarity classifier which synthesizes the adaptive resonance theory (ART) and the similarity classifier based on the Yu’s norm is proposed. The proposed ART-similarity classifier can not only carry out training without forgetting previously trained patterns but also be adaptive to changes in the environment. In order to test the proposed classifier, it is applied to the fault diagnosis of rolling element bearings. Before application to the fault diagnosis of bearings, considering computation burden principal component analysis (PCA) is proposed to reduce the number of features. The PCs are input the proposed classifier to diagnose the faulty bearings. The experiment results testify that the proposed classifier can identify the faults accurately. Furthermore, in order to validate the effectiveness of the proposed classifier further, it compares with other neural networks, such as the fuzzy ART, self-organising feature maps (SOFMs) and radial basis function (RBF) neural network through diagnosing the bearings under the same conditions. The comparison results confirm the superiority of the proposed method.