In this study, Hilbert-Huang Transform (HHT) is utilized for fault diagnosis under fixed rotating speed. The time-frequency analysis is to identify the severity of the gear faults. The experimental cases include the common faults of the gearbox, such as broken teeth, gear wearing and gear unbalance. The complicated vibration signals due to faults are first decomposed into a number of Intrinsic Mode Functions (IMFs), and then the envelope analysis is employed to extract the fault characteristics. Specific features of time-domain signals as well as the results of HHT analysis are extracted for Principal Component Analysis (PCA) to achieve the characteristic dimension reduction. The composite indicators obtained from PCA are used as the inputs of Neural Network to classify the different gear faults. The analysis results show that through PCA, the characteristic dimension can be reduced and the classifying accuracy of neural network can be also improved.