In order to solve incomplete prior information of the low probability of intercept(LPI) radar in non-cooperative electronic countermeasure environment, the recognition algorithm based on multi-fractal and semi-supervised expectation-maximization(EM) is proposed.Firstly, the multi-fractal spectrum of signal is calculated, and multi-fractal spectrum characteristic parameters are extracted.To solve the problem that the slower convergence speed and lower accuracy rate are caused when all unlabeled samples are joined in samples sets, this paper proposes a semi-supervised EM algorithm based on self-training ideology.By selecting one or some unlabeled samples to update sample sets, the unlabeled sampled sets are reduced, which can speed up the training speed of the classifier, avoid error accumulation, and improve the classification accuracy.The simulated results show that, the proposed algorithm has higher recognition rate and better real-time performance in LPI recognition.