It is difficult to identify the motor parameters of permanent magnet synchronous motor, and the electromagnetic torque is also difficult to accurately estimate by mathematical model, which leads to the decrease of the motor control precision and the overall performance of the drive system. A motor electromagnetic torque network topology based on back-propagation (BP) neural network is designed. The network is packaged into a torque observer by MATLAB/Simulink for accurate calculation of motor torque. Experimental verification and comparison with the traditional calculation method are carried out by the experimental platform. Experimental results show that the torque observer has high-precision torque output performance and the control precision is higher than that of the traditional torque estimation mathematical model.