The wavelet analysis and neural network were employed to study the dif ferent defects of wood materials medium density fiberboard(MDF) by the way of pa ttern recognition. The Daubechies wavelet packet was used to decompose vibration signals with different defects of the materials, and the energy rate in each fr equency segment from the calculation, as the training sample, was used to train the different topological structure neural networks. Then, the trained networks were used to recognize the defects. The results show that two kinds of poor adhe sive defects with similar properties are recognized as one pattern; the BP netwo rks with one layer or two layers could recognize bubbling, poor adhesive and no defects extreme exactly; The generalization of the BP network with two layers is better than that with one layer; and the fluctuation of the BP network with two layers is smaller than that with one layer. The optimal BP network to recognize the three pattern of MDF is the twolayer network with the hidden nodes of 20 and 6 respectively. And the accuracy of recognizing the MDF defects is 90%.