The interior permanent magnet synchronous machines have advantages of high power density, high reliability, field weakening performance etc. However, subject to the nonlinear characteristics of motor parameters, accurate estimation of the electromagnetic motor torque is very difficult. In this paper, a convolutional neural network based electromagnetic torque estimation method, i.e., a torque observer is investigated. Training data of the convolutional neural network are collected from the simulations of a high fidelity nonlinear interior permanent magnet synchronous machine by the means of finite element analysis. Then, a control scheme is adopted to control the interior permanent magnet synchronous machines with the proposed torque observer. In order to reduce the torque estimation error, different parameters and structures of the convolutional neural network are compared. Experimental results show that the proposed convolutional neural network based torque observer can estimate the electromagnetic torque accurately.