Aiming at the problems of large calculation of model predictive torque control (MPTC) for surface-mounted permanent magnet synchronous motor (PMSM), the convolutional neural network (CNN) is used to replace the MPTC for selecting voltage vector among candidate voltage vectors. Simulation results show that the steady-state control performances of the MPTC based on CNN are basically equivalent to those of the conventional MPTC. But the imbalance between the training data at steady and dynamic states deteriorates the control of the MPTC based on CNN at dynamic state and results in large stator flux ripple. Therefore, the adaptive switching strategy between the MPTC based on CNN and the direct torque control (DTC) is proposed, with the state of the system as the switching criterion. Simulation results show that the adaptive switching strategy can achieve high steady-state performance and suppress stator flux ripple at the dynamic state.