Intelligent Power Allocation Optimization Algorithm for Enhanced CR-NOMA Network via DenseNet-BiLSTM

• Published in: IEEE transactions on cognitive communications and networking Vol. 11; no. 4; pp. 2543 - 2553
• Main Authors: Fu, Yu, Chen, Bingcai, Liu, Chanjuan, Lin, Kai
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.08.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Artificial neural networks; artificial noise (AN); Autonomous aerial vehicles; Channel estimation; Codes; Cognitive radio; DenseNet-BiLSTM; Fading channels; intelligent power allocation; Machine learning; NOMA; non-orthogonal multiple access (NOMA); Nonorthogonal multiple access; Optimization; Optimization algorithms; Prediction algorithms; Relays; Resource management; Security; Underlay cognitive radio (CR); Unmanned aerial vehicles
• ISSN: 2332-7731; 2332-7731
• DOI: 10.1109/TCCN.2024.3516044

This paper presents a novel non-orthogonal multiple access (NOMA) network featuring an unmanned aerial vehicle (UAV) relay within underlay cognitive radio (CR) framework. Artificial noise (AN) is employed to enhance the system's security performance, under the realistic assumptions of imperfect channel state information (I-CSI) and imperfect successive interference cancellation (I-SIC). Accounting for Nakagami-m fading with path loss over all channels, as well as interference from the primary transmitter, the secrecy outage probability (SOP) expression for the whole system is derived to investigate the effect of power allocation on it. In response to this observation, we design a DenseNet-BiLSTM network and introduce a power allocation optimization algorithm based on secrecy considerations. The DenseNet-BiLSTM network consists of improved dense convolutional network (DenseNet) and bidirectional long short-term memory network (BiLSTM) architectures. Designed to address challenges encountered in training deep neural networks, DenseNet establishes dense inter-layer connections to mitigate vanishing gradients and promote feature reuse. Additionally, a self-attention layer is inserted into the DenseNet network to improve the efficiency of information transmission. BiLSTM, an extension of LSTM networks, improves modeling capabilities and predictive accuracy through processing sequential data bidirectionally and capturing both past and future context. DenseNet and BiLSTM are integrated by an attention mechanism to learn and represent complex relationships between data more accurately. Experimental results demonstrate that the proposed algorithm provides extremely satisfactory predictions and much more accurate coefficients in comparison to other prominent deep learning architectures.