A new 3D model retrieval method was proposed to improve feature extraction accuracy, which could further satisfy the needs of agricultural intelligence retrieval. Firstly, the CAD models in agricultural database were normalized. Then CAD models were transformed into eight subband partitions by 3D discrete wavelet. The two-level wavelet transform was made for LxLyLz, which were the lowest frequency parts. After that a single CAD model could be represented by 15 frequency functions. The AHP method was used to calculate the weight of different functions, with which the similarity value was defined for similarity comparison. Finally, the effectiveness of the proposed method was proved with ESB and agricultural model database by using VS2010, Matlab 2016b and Open Cascade. The retrieval results in ESB database showed that the ranking deviation of wavelet algorithm was lower than those of others under the same retrieval condition. The harvester reel retrieval results presented that the similar evaluation considered the geometry morphology as well as functional request. The applied instance demonstrated that once the similar models were selected, designer only needed to make a small re-edit and modification to finish the total design work. The precision of wavelet transform retrieval algorithm varied with different wavelet functions and coif wavelet showed high applicability. APH method could sufficiently distinguish the weight of various decomposed parts. At present, two-level wavelet transform was quite enough to meet model retrieval requirement. The wavelet transform method aiming at design retrieval could satisfy design needs and broaden design ideas, which can promote intelligent design.