[Objective]By constructing an effective expression data set of engagement degree and designing a joint evaluation model of students' engagement degree in class, the problems of lack of relevant expression data set and low accuracy of the model existing in the evaluation methods of students' engagement degree are solved. [Methods]Data were collected based on real online classroom scenes, and an expression dataset suitable for engagement recognition was constructed. An improved VGG model was designed to evaluate the dataset and recognize engagement expression. The expression score and face score were combined to construct a joint evaluation model of students' engagement, and the actual class engagement scores of tested students were calculated.[Results]For engagement expression recognition, the network structure was adjusted and verified through parameter tuning optimization. The results showed that the improved model VGG16+Dense+Dropout(lr = 1e-5) had the highest accuracy. For engagement evaluation, the joint engagement score is more accurate than the single expression engagement score. [Limitations]No more ablation studies were designed in the process of training the model, and deeper neural networks were not explored. [Conclusions]The face dataset of W-AttLe is suitable for verify students' engagement in class. The proposed joint engagement evaluation model makes up for the deficiency of single index model. The proposed weighted test scheme combining knowledge point test and self-test of comprehension degree validates the joint engagement degree model.