<p>Signal detection is essential in wireless communication systems because it enables the simultaneous decoding of multiple overlapping signals from different users. The non-orthogonal multiple access (NOMA) technique has been paid much attention due to its high bandwidth, low latency, and massive connectivity. It applies successive interference cancellation (SIC) techniques to detect the transmitted signals in the receiver. The decoding performance is depreciated due to the effect of error propagation. This article suggests a self-attention-enabled deep learning (DL)–based hybrid model to improve the signal detection performance of a multi-user downlink NOMA system. The model is based on the temporal convolutional network bidirectional long short-term memory (TBLSTM) and a self-attention mechanism. The prediction errors are illustrated with various evaluation standards such as mean absolute error (MAE), overall dynamic time wrapping (DTW) score, mean square error (MSE), and root-mean-square error (RMSE). The bit error rate (BER) performance of the proposed DL-based hybrid framework for the different users is discussed with the standard SIC-based least squares (LS) technique and different DL-based bidirectional long short-term memory (BiLSTM), temporal convolutional network (TCN), and hybrid TBLSTM approaches. The simulation results illustrate that the proposed model can attain an effective signal detection approach for multi-user downlink NOMA systems.</p>