A novel method is proposed for identifying vehicle loads on bridges and selecting the error pattern based on Bayesian inference. Initially， the expression of the relationship between the vehicle loads and the measured responses is constructed using the static influence line. The modified coefficient surface is then established to eliminate the identification error caused by the dynamic effects. Afterwards， five error modes related to the structural responses and vehicle speed are introduced. According to the assumed prior distribution， the posterior distribution of vehicle axle loads is derived. On that basis， the optimal estimated value and confidence interval of vehicle loads can be obtained. Furthermore， the posterior probability of each error mode is calculated. Finally， the identification accuracy and reliability of the proposed method for scenarios under different vehicle speeds are validated by the numerical example of simply supported beam and the dynamic load test of a continuous beam bridge. The results show that the modified coefficient surface can effectively eliminate the vehicle dynamic impact and improve the accuracy of vehicle load identification. The result of vehicle load identification is presented in the form of confidence interval to quantify the uncertainty. Bayesian method can identify the optimal error pattern， which further improves the robustness of load identification.