For the permanent magnet synchronous motors, the controlling algorithms are usually complex and the motor parameters identification are difficult. Since the electromagnetic torques are difficult to estimate through mathematical models, which leads to a decline in motor control accuracy and overall performance of the drive system. In this paper, a topological model of the electromagnetic torque network of the motor was investigated based on the dynamic recursive feedback (ELMAN) neural network. At the same time, the neural network is built as a torque observer by the MATLAB / Simulink for accurate estimation of the motor torque. In the experiments, traditional torque calculation method and the back propagation neural network are compared with the proposed approach. In comparison, the proposed torque observer has better performance in both torque estimation accuracy and control precision.