Early fault detection methodology in gear box diagnosis has been proposed to find the status of the gear based on vibration signals obtained from the experimental test rig. Signal processing categorized to time-frequency domain such as continues wavelet transform is used in the proposed work for statistical feature extraction. Feature selection method is used for selecting the extensive useful features among the extracted features to reduce the processing time. A famous optimization technique, Genetic algorithm (GA) and rough set based approach is used to select the best input features to reduce the computation burden. The efficiency of this feature selection method is evaluated based on the classification accuracy obtained from the proposed algorithms: back propagation neural network (BPNN) a famous artificial neural network algorithm and C4.5. Performance of classifiers are evaluated with the different signals acquired from the experimental test rig for different states of gears. © 2014 Published by Elsevier Ltd.