In this paper we proposed a knowledge tracing method based on Transformer structure, improved the embedded representation of interactive records, designed a gate unit suitable for this model, and optimized the input processing of self-attention sublayer to improve the predictive performance of deep knowledge tracing model. The experimental results on four commonly used public data sets show that compared with previous methods, the model proposed in this paper can better reflect learners’ mastery of knowledge points, and has better performance on data sets with large sample sizes.