The employment of non-orthogonal multiple access (NOMA) technology represents an innovative approach in enhancing spectrum efficiency and bolstering user access capabilities, which acts as a pivotal technology within 5G and future maritime intelligent transport applications. With numbers of users increasing significantly, complexity of implementation increases dramatically if traditional serial interference cancellation (SIC) technology is used for signal detection. Aiming at error propagation problem in the decoding process of the traditional SIC receiver, a novel downlink dual-user NOMA signal detection scheme based on convolutional neural network (CNN) is proposed in this paper. The CNN-based scheme can directly process the NOMA signals received at the receiver without relying on the traditional SIC receiver. Through the training and optimization of the neural network model, the accurate detection of NOMA signals has been achieved, and the technical advantages of this scheme in improving system transmission rate and reducing error propagation have been evaluated.