In order to overcome this defect of traditional magnetic target motion estimation′s dependence on the initial state information of the target, a three-axis projection model of magnetic moving ship targets was established, and 10 kinds of target training datasets, validation data sets and test data sets of magnetic ship moving targets with variable parameters were generated. Multi-channel convolutional neural network was designed to estimate the distance abeam and velocity of the target, and the effects of different learning methods and activation functions on the performance of the network were compared and analyzed. The results show that the performance of Adam+tanh method is better than other methods, and the estimation effect of motion parameters is accurate. Compared with Kalman filter and particle filter, this estimation algorithm is calculated with preferable efficiency and independent of initialization for estimation.