Abstract—  Non-orthogonal multiple access (NOMA),  is one of the significant candid approach for the Fifth-generation (5G)  has drawn a great deal of attention in the wireless communication. At receivers, previously Successive interference cancellation (SIC) technique was utilized most commonly for both uplink and downlink of  NOMA transmission, which complicates the receiver and raises issues with error propagation limit. In order, to avoid complications faced we investigate a tool with excellent performance and efficiency i.e, Deep Learning(DL). This approach analyses automatically the channel state information (CSI) and detects the original transmission sequence. The ideal channel gain in the present schemes, can be achieved by removing the signal which has a greater power allocation factor while detecting the signal with a lesser power allocation factor. In a Proposed model of  Deep Learning both channel estimation process recovers the  signal facing the channel distortion and multiuser signal superposition. Simulations of MIMO-NOMA-DL systems have been analyzed and compared to the previous SIC method. This approach  successfully construes channel impairment and obtains a  good detection performance, according to the findings of our simulations. Instead of employing clever detection methods, MIMO-NOMA-DL uses a neural network to find the best answer (NN). Deep learning is therefore a strong and useful tool for NOMA signal detection.