In the multi-sensor fusion tasks of automatic drive, the strategy and the results of the data fusion are greatly influenced by the uncertainty of each subtask. To keep the whole system run steadily in multiple circumstances, the calculation model must operate with low uncertainty. The existing methods can only obtain uncertainty in the neural network prediction process, and few methods can reduce the uncertainty of the model in a self-learning method. To address the above problems, the concepts of uncertainty learning layer and uncertainty loss term are proposed, and a neural network architecture (ULNN) which can reduce uncertainty by self-learning method is designed to enhance the robustness of neural network model prediction. Experiments on CIFAR-10 and CIFAR-100 datasets show that ULNN can effectively reduce the model uncertainty and obtain 26 and 12 times lower uncertainty on the two data sets respectively. The universality of ULNN is proved by the experimental results of semantic segmentation on CamVid dataset.