Model predictive torque control (MPTC) selects the optimal voltage vector by traversing all candidate voltage vectors, which results in large calculation and poor real-time performance. In order to solve these problems, the data of MPTC are collected to train a convolutional neural network (CNN) and the CNN is then used to replace MPTC for the optimal voltage vector selection. In order to solve the out-of-control problem in using CNN, the MPTC based on direct torque control (DTC) and CNN is proposed. Voltage vectors selected by CNN and DTC are used as candidate voltage vectors of MPTC. Simulation results show that the proposed strategy can effectively solve the out-of-control problem of CNN, its control effect is basically equivalent to MPTC, while the torque and stator flux ripples of the proposed strategy are significantly lower than those of DTC.