The artificial neural network (ANN) is a very popular fault classification tool in the condition monitoring of rotating machines. It uses statistical methods to preprocess vibration signals to obtain input features. ANNs offer the advantages of automatic detection and diagnosis of machines since they do not require an in-depth knowledge of the behavior of their rotating/reciprocating elements or of their internal vibration generating mechanisms. However they require a large amount of training data patterns. This paper addresses the development of a condition monitoring procedure for a gear transmission system. A two stage helical gearbox operating at a constant speed and load has been considered for the experiments. For controlled power transmitted by the gearbox, vibration signals were picked up by accelerometers mounted on the bearing and were analyzed to obtain statistical features for different conditions of faults induced in a gear tooth. Seven conditions of the gear were investigated: healthy gear and gear with six stages of depthwise wear induced on the gear tooth by grinding operation. The features extracted from the measured vibration signal were mean, root mean square (rms), variance, skewness and kurtosis, which are known to be sensitive to different degrees of faults. These characteristics were used as input features to an ANN with five input nodes, fifteen hidden nodes and seven output nodes to classify faults in the gearbox. The ANN was trained using 20 sets of data, each data set consisting of 30 time domain signals (each time domain signal consisting of 16 averages over 5 revolutions of the gear tooth). The analysis bandwidth was chosen up to 5th gear mesh frequency with a resolution of 20 Hz. The ANN was optimized for best accuracy by arriving at the best possible combination of number of epochs, hidden layers and learning rate at constant momentum rate. Results show that the classification performance of the ANN yielded a maximum accuracy of 93% over the chosen range of defect stages.