In the motor fault diagnosis technique,vibration and stator current frequency components of detection are two main means.This article will discuss the signal detection method based on vibration fault.Because the motor vibration signal is a non-stationary random signal,fault signals often contain a lot of time-varying,burst properties of ingredients.The traditional Fourier signal analysis can not effectively extract the motor fault characteristics,but are also likely to be rich in failure information but a weak signal as noise.Therefore,we introduce wavelet packet transforms to extract the fault characteristics of the signal information.Obtained was the result as the neural network input signal,using the L-M neural network optimization method for training,and then used the BP network for fault recognition.This paper uses Matlab software to simulate and confirmed the method of motor fault diagnosis validity and accuracy. In the motor fault diagnosis technique, vibration and stator current frequency components of detection are two main means. This article will discuss the signal detection method based on vibration fault.Because the motor vibration signal is a non-stationary random signal, fault signals often contain. a lot of time-varying, burst properties of ingredients. The traditional Fourier signal analysis can not effectively extract the motor fault characteristics, but also also likely to be rich in failure information but a weak signal as noise. Herefore, we introduce wavelet packet transforms to extract the fault characteristics of the signal information. Obtained was the result as the neural network input signal, using the LM neural network optimization method for training, and then used the BP network for fault recognition. This paper uses Matlab software to simulate and confirmed the method of motor fault diagnosis validity and accuracy.