Traditional mechanical manufacturing experimental teaching is limited to one teacher demonstrating operations to several students at the same time, which is difficult to take into account and evaluate the differences in knowledge mastery of different students. In order to improve the above teaching defects, firstly, the teaching evaluation of students’ experimental level is carried out based on their experimental operation behaviors through K-means clustering. On this basis, a deep learning-based knowledge tracking SAFFKT model is designed to empower and update students’ knowledge status. A personalized teaching recommendation method for virtual simulation is proposed based on students’ knowledge state, and the hidden semantic matrix decomposition recommendation algorithm for teaching recommendation is improved and implemented. The AUC and ACC of SAFFKT model are significantly higher than that of the comparison model (p<0.01), and it is robust. The F1 value of the recommended experiments was 0.775, indicating a better recommendation effect. The teaching evaluation model achieves accurate classification of students' experimental behavior and yields different learning characteristics of three types of students. Therefore, the innovative work of virtual simulation teaching strategy in this paper is of practical significance.