Abstract: In the high-end motor control systems such as robots, due to the difficulty in parameter identification and harmonic interference of the motor during the operation of the internal permanent magnet synchronous motor (IPMSM), it is difficult to accurately calculate the electromagnetic torque of the motor through the traditional mathematical model, which has strong nonlinearity. The increase of torque sensors will increase the system cost. In order to realize high-precision torque prediction of torque-sensorless condition, a torque prediction model of IPMSM based on model fusion method is proposed. A simple linear torque mathematical model is used to fuse data-driven neural network algorithm, which effectively reduces the complexity of neural network model and improves the accuracy of torque estimation. The modeling simulation and experimental verification are carried out by using two common networks: BP and RBF, which proves that the algorithm realizes real-time online torque prediction and has good dynamic stability.