In float glass fabrication, once defects are detected, recognition of the defect type is important for adjustment of the process conditions. Based on analysis of the float process conditions and float glass defects, this paper outlines an adaptive neutral network-based classification. 12 features of the defect are picked out as the inputs for defect classification. The ReliefF method is used to evaluate the features, and the sequence of defect features is decided based on the results of the evaluation. The momentum term and adaptive learning rate is used to address the classification's disadvantage of slow learning speed and easy falling into a local minimum. Experiments and real application proves that the float glass defect classification presented in this paper correctly recognizes defects in real time, and satisfies the requirements of float glass fabrication.