The use of the existing Loss Differentiation Algorithms (LDA) cannot get prior knowledge from actual networks, so the algorithms can hardly be generalized to various environments. To solve this problem, a new LDA is proposed by enhancing the Fuzzy One-class Support Vector Machine (FOCSVM). The new LDA converts the differentiation packet loss between congestion and bit error on wireless links into judging that whether the loss is because of bit error. The training set consists of only the in-sequence packets received by the receiver, so the classification model can be got and updated during the session, without the prior knowledge about the two types of loss, which makes the LDA have fine generalization. Simulation results show that the proposed LDA has good differentiation accuracy and can enhance the utilization of resources in wireless networks.