Distant supervision relation extraction methods based on the multi-instance learning framework mostly rely on contaminated labels that are heuristically generated, and focus on predicting relations at bag-level. However, they show unsatisfactory performance on sentence-level prediction which is more friendly with comprehend sentence tasks, like question answering and knowledge graph completion. To solve the above problems, a novel distant supervision relation extraction method is proposed in this paper, in which we train the model at sentence-level via positive learning and negative learning to separate noisy data and enable faster convergence. Meanwhile, a constraint graph is constructed to encode the re-strictions between relations and entity types and is optimized by an auxiliary loss towards relation prototype, which allows information propagation among different relations that makes the model can learn essential and interpretable sentence representation. We not only identify noisy data but also revise the labels of them iteratively to refine the quality of distant data and further enhance model performance. This method performs well in the sentence-level relation extraction task of the NYT dataset, with an accuracy of 77.69%, which is 6.47% higher than the current optimal baseline model. The F1 score on the noisy annotated test set is as high as 85.88%, verifying its excellent denoising ability. The ablation experiment results show that the contribution of the constraint graph to the optimization of the relation prototype is 11.02%. The experimental results show that this method significantly outperforms the existing methods in the sentence-level relation extraction task, not only effectively reducing the impact of noise, but also significantly improving the model performance, providing an efficient solution for the remote supervision relation extraction task.