This paper presents a novel method to monitor crack length which based on binary tree support vector machines (BTSVM). In this method, matching pursuit method with Chirplet atom is applied to extract the matching parameters as feature vectors to train and test in the BT-SVM algorithm. Then one simulation of lug joint is carried out for studying the effect of crack extension on Lamb wave signals propagation. Fatigue loading experiments on lug joints are carried out at last. The results show that this new method can monitor the length of fatigue crack effectively.