Achieving a reliable fault diagnosis for gears under variable operating conditions is a pressing need of industries to ensure productivity by averting unwanted breakdowns. In the present work, a hybrid approach is proposed by integrating Hu invariant moments and an artificial neural network for explicit extraction and classification of gear faults using time-frequency transforms. The Zhao-Atlas-Marks transform is used to convert the raw vibrations signals from the gears into time-frequency distributions. The proposed method is applied to a single-stage spur gearbox with faults created using electric discharge machining in laboratory conditions. The results show the effectiveness of the proposed methodology in classifying the faults in gears with high accuracy.