An intelligent fault diagnosis method based on relevance vector machine (RVM) is proposed for incipient fault detection of gear. First, by combining wavelet packet transform with Fisher criterion, it is able to adaptively find the optimal decomposition level and select the global optimal features from all node energies of full wavelet packet tree. Then, the RVM is adopted to train the fault diagnosis model. Finally, to improve accuracy for online continuous fault diagnosis, moving average is applied to each feature before it is input into the fault diagnosis model. Experimental results demonstrate that the proposed method can achieve very high classification accuracy and indicate that RVM is more suitable than SVM for online fault diagnosis.