With the concept of ubiquitous power internet of things, the transient stability plays an increasingly important role in the operation and control of power system. Due to the extensive configuration of the phasor measurement unit (PMU), the real-time evaluation method of machine-based transient stability shows great potential for development. Aiming at the problem that data generation of the offline training is time-consuming and it is difficult to update the model quickly after the grid changes, this paper proposes a power system transient stability assessment method based on active learning. Firstly, implementing short-time simulations (simulation to fault clearing time) to generate unlabeled samples in different kind of operations and faults; and then, randomly select a part of samples for long-term simulation to mark these samples; Then, a part of samples are randomly selected for long-term simulation to be labeled with transient stability status, and support vector machine is further trained to build the transient stability assessment model; Finally, the data with high information entropy in the remaining unlabeled samples is re-trained until the model accuracy does not change. The simulation of the New England 10-machine 39-bus test power system shows that the proposed method can effectively reduce the time of offline simulation, which greatly improves the efficiency of model configuration, and it is robust to wide-area noise.