Extended Abstract#br# When monitoring sea targets, there are three-dimensional swing and translation components of ship targets in the navigation under the action of waves. The non-stationary motion component makes the synthetic aperture radar (Synthetic Aperture Radar, SAR) echo approximate to a multi-component polynomial phase signal, resulting in different degrees of defocusing of the three-dimensional rotating ship target, which makes the three-dimensional rotating ship target recognition accuracy rate low. At present, the research on the recognition of stationary ship targets has achieved good results, but there is still no good solution for the recognition of complex moving targets. Aiming at the complex information features of moving targets in SAR images, this paper applies the complex-valued convolutional neural network to SAR ship target recognition. In this paper, the deep mining of the target phase information is considered when designing the network architecture, combined with the advantages of the existing amplitude-phase-type and real-imaginary-type complex-valued convolutional neural networks, a mixed-type complex-valued convolutional neural network (Mix-CV-CNN) is proposed. Meanwhile, this paper derives the forward propagation and backpropagation algorithms of Mix-CV-CNN and proposes a recognition algorithm for the SAR ship target based on Mix-CV-CNN. The 3D rotating ship target has residual phase information after SAR imaging processing. Mix-CV-CNN could make full use of the amplitude and phase information of the complex SAR image and could better complete the recognition of the SAR complex moving ship without refocusing the target. The proposed method reduces the task difficulty of complex motion ship target identification and has certain innovation and practical value. Experiments show that the recognition performance of Mix-CV-CNN is improved compared with the real-valued convolutional neural network with the same degree of freedom, real-imaginary-type complex-valued convolutional neural network, VGG16, and ResNet18, and the average accuracy of measured data recognition is improved about 3.85% to 6.37%.#br#