An online fault detection algorithm for automobile rear drive asse m bly (RDA) using wavelet decomposition was proposed. The fast Fourier transform ( FFT) and the wavelet decomposition were used to decompose and find the frequency band which mostly represents the RDA faults, and then the most indicative featu res from the reconstructed signal of the frequency band, including the variance and kurtosis of the signal, were extracted. To find an optimal fault detection i ndex, a sample of RDA workpiece, including normal and abnormal ones, was randoml y selected from the vibration signal database which was gathered from the assemb ly line of RDA. The features of every signal were extracted and used to train th e support vector machine (SVM) neural network to find the optimal classification hyperplane. The algorithm is currently used on the assembly line. It is found t hat the algorithm is robust to different working conditions, and is able to dete ct the faults in RDA online and effectively.