Gear pitting diagnosis has always been an important research topic and different diagnostic methods are used. This paper uses artificial neural network (ANN) to diagnose early gear pitting. There are several issues that are inevitable with the application of ANN for diagnosis of early gear pitting: feature extraction, neural network structure optimization, and training stability. This paper proposes a new feature extraction method that uses fast Fourier transform (FFT) to select a number of frequencies in the frequency spectrum and use the frequency amplitudes as the inputs of ANN. The particle swarm optimization (PSO) is used to optimize the initial value of the network to make the training more stable. Furthermore, the performance of the ANN diagnosis under different working conditions are also compared and analyzed in the paper. The proposed method has been validated through the data collected from the gear pitting test experiment. The validation results have shown that the faults diagnosis accuracy could reach 100%, which proves that the proposed method is reasonable.