Abstract The paper presents an improved binary tree-based SVM multi-class classification algorithm. Guided by the principle for those class samples most easily to be separated should be divided firstly, the method, after defining class recognition rates of feature parameters, calculates level-by-level the class recognition rates of each feature against each class training samples of the corresponding level SVM; Then, according to the synthetic sorting result of the class recognition rates, chooses the class placed first as the positive class of the corresponding level SVM. From the cylinder head vibration signal envelopes, nine sensitive temporal domain parameters and wavelet package frequency band power percentage parameters are extracted and formed the diagnosis feature vector. The proposed binary tree SVM, as well as the commonly used 1-a-r (one-against-rest), 1-a-1 (one-against-one) and DDAG (Decision Directed Acyclic Graph) multi-class SVM classification method, are respectively used to recognize six simulative diesel valve gap states. Result shows, among the four multi-class classification methods, the proposed strategy has the best classification precision.