In order to reduce the problem of traditional evaluation methods being difficult to provide accurate feedback due to the diversity of students' self-directed learning behaviors and differences in technology platforms in online teaching, the LMBP （Levenberg Marquardt Back Propagation） algorithm is introduced to construct an automatic evaluation model that can quantitatively analyze students' learning performance using weighted evaluation indicators. Determine the weight of evaluation indicators for online teaching and learning, screen out key evaluation indicators, and allocate weight values reasonably to reduce data disorder. Based on the LMBP algorithm, an automatic evaluation model is constructed to automatically calculate the online learning evaluation score of each student through the operation of the model, reducing the lag of evaluation and achieving objective and accurate evaluation. The experimental results show that the weight values of various indicators calculated by the model are above 0.96, the fitting degree is higher than 0.98, and the evaluation score is higher than 97 points, which can achieve effective evaluation of online teaching.