Fault diagnosis based on the wavelet packet decomposition, one-against-one support vector machine (SVM) and genetic algorithm (GA) is proposed in order to realize the real-time sensor fault diagnosis accurately. The input feature vectors of one-against-one SVM are produced by wavelet packet decomposition of the sensor output signal. GA is used to obtain optimal parameters of one-against-one SVM network model automatically, which can enhance the training speed and performance. The experiments of photoelectric encoder fault diagnosis show that the combination of these methods makes SVM own a better recognition rate and overall performance, which can improve the accuracy and time efficiency of fault diagnosis.