<p>Failure of the gear drive provides a large adverse influence on the machine systems. Detection of early failure on the gear tooth is crucial to prevent the serious damage in the machine systems. In this study, we diagnosed the tooth surface of normal gears without a failure and spot damaged gears. Diagnosis was accomplished with the vibration acceleration on a bearing stand. A synchronous averaging processing for the measured vibration acceleration waveform in one rotation period of the gear was employed to exclude the noise of measured signal. Then, the synchronous averaged data was divided into each one pitch. Tooth surface damage was diagnosed with a nearest neighbor method for the pattern recognition. The nearest neighbor method identified the class through the distance of the feature vector between input factors and a prototype. We demonstrated that the presence of damage can be detected with the nearest neighbor method.</p>