The wear of abrasive belt is monitored based on the grinding sound and current. First, the feature of these signalsare extracted by the time domain analysis method and the wavelet packet analysis method, and merged by Naive Bayes classifier to identify the abrasive belt. Furthermore, to increase the accuracy of identification, the Naive Bayes classifieris improved by introducing a signal feature selection method based on Fisher discriminant rate and mutual information. The experimental results show that the signal features with good separability and weak correlation can be selected by the improved feature selection method, and the wear state recognition method based on Naive Bayes classifier can identify the abrasive wear state accurately.