In transient stability assessment using the classical support vector machine（SVM）， all the training samples are required at the training stage at the same time. If new samples are added， the SVM need to be retrained with all the samples. Considering this limitation， a transient stability assessment method is proposed based on SVM incremental learning. A fast SVM incremental learning method is adopted. It constructs a recursive solution and adds new data to the solution. Karush-Kuhn-Tucker conditions are maintained for all the previous used training data. Consequently， assessment models are updated through incremental learning once a new sample is added. The simulations on the New England 39-bus test system demonstrate the proposed method can effectively update models and reduce training time dramatically. The method provides a new thought for online learning in power system transient stability assessment based on machine learning method.