2. Key Aspects for Practical Implementation of DL-Based NOMA
2.1. Resource Allocation
In NOMA, one resource block (RB) is shared by multiple users, and the SIC receiver is used to decode the user information at the receiver end based on the user’s channel gains. Interference between the users can be avoided by choosing proper power allocation algorithms. Otherwise, resource allocation issues such as user pairing and power allocation (PA) will arise. In user pairing, the users with less power are allocated with more channel gain, and users with more power are allocated with less channel gain to make channel fairness to all the users at the transmitter end. At the receiver end, the SIC receiver is used to decode the same. In this method, if the number of users increases, then the decoding complexity also increases at the receiver end. This is one of the major problems in user pairing. Along with this, another problem, i.e., if the users with high and low gain are transformed to mid-gain, then mid-gain users may be paired or may not, which leads to reduced channel capacity.
To overcome the user pairing issues, optimization techniques, game theory, machine learning and deep learning algorithms are proposed in the literature. The authors proposed an optimization method while pairing two users
[20][21][22]. To optimize the user pairing, the channel gain should not be less than the predefined threshold. A strong channel pairing algorithm can increase the system capacity and fairness in user pairing. In
[21], the authors used a new pairing concept, i.e., the highest channel gain users are paired with the next highest gain users. Different Game theory algorithms for multiple user pairing and machine learning algorithms for user pairing have been proposed in recent research. In
[23], the authors proposed an RL-enabled joint power allocation and user pairing scheme. Through Q-learning, they were able to successfully implement both power allocation and user pairing with reduced computational complexity. In
[22], the authors introduced an optimal power allocation technique with a given sub-channel assignment through a closed-form approach. Considering this, a traditional deep reinforcement learning (DRL) algorithm named Deep Q-Network (DQN) algorithm is used to investigate the optimal user pairing scheme. The DQN algorithm provides better performance of the feature extraction ability and higher learning efficiency than conventional reinforcement learning (RL) schemes.
2.2. Power Allocation
One crucial challenge is how to allocate power when there are limited resources to make the most of the benefits of the NOMA system. It has been established that this issue of optimum power allocation is NP-hard, indicating that it is impractical and expensive to study all possible channel assignments to find an ideal solution. As a result, several methods have been put forth by researchers to deal with this issue. Solutions include distributing power for a downlink single input and single output (SISO) NOMA system
[24], distributing power for the most equitable distribution of users
[25], and distributing power for the most energy-efficient use of resources
[26]. Deep learning techniques must be used because several solutions have been demonstrated to be less than ideal. A thorough literature assessment of deep learning-based approaches to the power allocation concern will be provided in sufficient depth in this section. Utilizing DL in NOMA, deep neural network generic architecture efforts are at the forefront of current technological advancements in power allocation. To distribute power to consumers in the best possible way,
[27] suggests a deep reinforcement learning (DRL) method; specifically, an artificial neural network (ANN) is employed to perform channel assignment. The system model is based on BS and several users in a downlink NOMA scenario. Users serve as the performance environment for the deep learning algorithm, which treats BS as an agent. To allocate resources and channels to users, BS first chooses a task (channel assignment) from a set. A feedback signal is then provided towards the BS to help assign users in the following transmission based on the users’ responses. The three crucial parts of this process are the status space, action space, and reward function. The channel information is responsible for the state space. The agent (BS) chooses a single channel for data transmission for a single user in the action space. The collection of actions is constrained to meet the requirements of user channel allocation, so each user is associated with a unique action. After the user acts, the allocation procedure is complete. The signal returned to the BS as a result of a failed or successful transmission at the conclusion of each time slot is the reward function, as shown in
Figure 3. The data rates each user experiences and is sent by the BS to make up the signal. The goal of
[28] is to maximize this incentive signal and, in turn, optimize each user’s data rate. The acquired findings provide a sum-rate comparison of Joint Resource Allocation (JRA) without downlink versus JRA with DL, where the non-DL variant is significantly outperformed by the DL counterpart.
[29] suggests a power allocation plan that uses DL approaches to optimize the system sum rate in a downlink NOMA environment with an incomplete SIC. The algorithm for finding the best power allocation is exhaustive. In a recent study
[30], a power allocation approach for imperfect SIC to enhance the experienced system sum rate is suggested.
Figure 3. Power allocation and channel assignment of DL-based NOMA system.
2.3. Channel State Information
Practically speaking, channel state information (CSI) significantly influences the NOMA system’s performance, and several efforts have been made to implement channel estimation using NOMA scenarios
[31]. In
[32], a new linear estimator was developed to maximize the average effective signal-to-interference noise ratio (SINR) of the strong user, with a finite SINR required for the weak user to identify the CSI. Meanwhile, several researchers are looking at NOMA-based solutions in various CSI circumstances because the CSI is difficult to collect using conventional approaches. Two power allocation techniques and the performance of NOMA in an incomplete CSI environment were reported
[33]. Furthermore, using uplink NOMA systems, the researchers demonstrated that insufficient CSI causes improper decoding and additional interference with the intended signal
[34].
Consequently, how to efficiently collect flawless CSI is a crucial challenge in NOMA-aided approaches, and new techniques must be used to address this issue. Although numerous recent research contributions have developed various reliability and sum data rate optimization methods, these techniques demand high computational complexity because of the nonlinear optimization. They are unable to produce an associated power allocation mechanism against a given CSI. In particular, virtually all of the key benefits of NOMA techniques rely mainly on CSI; as a result, several strategies have been presented in previous studies to further enhance the effectiveness of channel estimates
[35]. Conventional approaches, however, are unable to trace the alteration in the channel state in real time due to the complexity of the channel conditions in multiple-user systems
[36]. Usually, the drastically fluctuating channel characteristics cause CSI acquisition to be disrupted and the NOMA system efficiency to suffer.
Nevertheless, nonlinear reconstruction techniques are inevitable since the channel sparsity trends have been frequently taken for granted in previous studies. Therefore, super-resolution direct arrival (DOA) estimates, and signal identification cannot be accomplished using standard approaches since they are inefficient and unreliable. The NOMA system has recently been enhanced with a promising machine learning (ML) approach to enable the auto-detection of the CSI. The DL
[37] idea, introduced in 2006 and a typical branch of machine learning, is a particularly effective technique for managing large amounts of data and resolving challenging nonlinear issues. A few earlier papers
[38][39][40][41] included DL in communication in relation to the physical layer, channel coding, and MIMO. The intriguing system that incorporates the DL into the OFDM context has been demonstrated in
[42], and its outstanding performance in the context of signal recognition and channel estimation has been confirmed. DL has also been used in traffic monitoring systems, which work admirably
[43][44][45]. Additionally, DL-based communication systems have shown certain benefits in terms of security, BER, and throughput performance.
2.4. Successive Interference Cancellation
The drawbacks of SIC might also be addressed using the DL technique. Due to the SIC receiver’s poor cancellation, overall capacity decreases
[46]. As a result of different hardware limitations, decoding and canceling may be faulty in real-world systems, making SIC possible. The performance gain of NOMA may be enhanced by using SIC at cell-edge users, as demonstrated by the authors in
[47]. The creation of an easy-to-use, effective SIC receiver is essential to NOMA. Multi-stage SIC lowers multi-path fading and BER. The performance of the system is impacted by the signal’s decoding sequence. High signal-to-noise ratio (SNR) signals are initially deciphered. The performance of the SIC receiver is enhanced by a low complexity, highly effective power allocation algorithm
[48]. A real implementation’s non-idealities and flaws cause error propagation in SIC, which is utilized to decode and identify desirable signals. Due to the signal processing required for SIC, receiver complexity increases as user equipment (UE) numbers rise. A deep neural network (DNN) is used in
[49] to approximate the SIC receiver. In the MIMO-NOMA system, the combined optimization of precoding and SIC decoding minimizes the total mean square error between the user’s intended signal and their decoded signal. Users and their sub-bands are grouped according to the status of the channel, ascending. The binary dislocation principle pairs them (BDP). As a result, users with excellent and bad channel conditions will be paired.
A sub-band that satisfies user demands is selected. EP at the receiver can be removed if the signal-to-interference-plus-noise ratio (SINR) difference among a pair of users in the sub-band is sufficiently great. Users that share a band are given authority by BS. The minimum mean square error (MMSE)-SIC method with Interference Rejection Combining (IRC), which analyzes noise and interference independently and enhances average channel capacity and, therefore, system performance, can be used to attain the best performance at the receiver
[50]. Some recent works are based on the theory that most EP-related problems in SIC may be handled by appropriately grouping or clustering users. Additionally, by concurrently optimizing precoding and SIC decoding using DNN and domain-specific information, the mean square error (MSE) value between the intended and decoded signals would be reduced to the absolute minimum.
2.5. User Fairness
Several recent studies
[51][52] have discussed the benefits of employing NOMA. This increases system throughput, spectrum efficiency and user fairness. Researchers can also obtain an extremely highly reliable connection. Time and frequency resources are distributed to users in the spatial domain through the power domain or code domain NOMA
[53]. A recent work
[51] uses a deep learning algorithm to ensure user fairness by dividing users into low-rate and high-rate requirement users considering their mobile phone usage habits. The authors consider a NOMA system with DL-based coordinated multi-point (CoMP), used in 5G cellular networks to guarantee the rate requirements from the different edge users.
Figure 4 evaluates the performance of user sum rates in dynamic point selection CoMP (DPS-CoMP) subchannels and the number of cells subchannels.
Figure 4 shows that the user sum rate in the DPS-CoMP subchannel of each cell in the DPA algorithm, the NOMA-CoMP algorithm, and the maximum throughput (MT) algorithm all increase as the number of subchannels increases. In the NOMA power domain, power distribution among users varies depending on the channel characteristics and user-specific channel quality. As a result, consumers located far from the BS will receive more power, and vice versa. According to the works cited in
[54], fairness for NOMA in 5G is highlighted by the fact that, in downlink mm-wave NOMA, various data from all users is thus overlaid in the power domain at the transmitter, and the SIC is performed at the receiving side. By integrating SIC and superposition coding (SPC) at the transmitter end and receiver side, respectively, researchers may use NOMA to increase spectral efficiency. The max-min fairness of both the average CSI and instantaneous CSI is also discussed in this literature. Considering the interaction between NOMA and cooperative transmission, the integration of NOMA with several emerging 5G technologies, the correlation with other NOMA variants, and the resource control of NOMA, the authors in
[55] focus on state of the art in power-domain multiplexing-based NOMA. Considering the hybrid beamforming system described in the research
[55], which employs phase shifters and sets of switches, down-converter, LNA, ADC, and DAC are components of the radio frequency (RF) chain. The price of the system rises along with the number of RF chains. To decrease the number of RF chains, the hybrid beamforming approach is applied, and the system’s price will be reduced as a result. 5G deep learning systems have been researched in the literature. Power allocation, DoA estimation
[56], physical layer security
[57], channel estimation
[58], energy optimization, etc., are all included in the program, which significantly addresses user fairness issues in NOMA.
Figure 4. Sum rate of DPS-CoMP users (Mbps) versus the number of subchannels per cell.
2.6. Impulse Noise
2.6.1. Impact of IN in NOMA
Numerous obstacles to the adoption of NOMA systems have been raised by the broader literature on NOMA, which primarily occurs with respect to next-generation networks like the Internet of Things (IoTs) and smart grids. Analysis has also been done on the impact of IN on the NOMA downlink
[59][60] and uplink
[61] systems. The NOMA uplink systems’ outage performance in the occupancy of IN is given in the research work
[61]. Analytical findings and comprehensive Monte-Carlo simulations were used to verify the NOMA system’s sensitivity to IN. The effect of IN on the cumulative rate capacity of NOMA downlink systems was given by the authors in
[60]. The real loss from IN was calculated using a particle IN scenario. The authors also examined how well NOMA performed across Rayleigh-fading channels with composite noise (impulse with AWGN). A union constraint on the bit error rate (BER) was developed using the pairwise loss of bits (LoB) formula. The research study quantified the variance in channel conditions that NOMA users encounter when there is composite noise. The operational reduction of NOMA-aided IoT networks caused by IN was examined, and a mitigation approach was suggested in
[62]. For acquired OFDM symbols generated from the power domain multiple-NOMA (PDM-NOMA) strategy, a multistage nonlinear solution based on deep learning was presented.
2.6.2. IN Mitigation Techniques
The threshold-aided IN technique is defined as a memoryless nonlinear mitigation strategy that comprises blanking
[63], clipping
[64], and clipping/blanking
[65]. In this method of mitigation, the high amplitude and short duration of IN are studied by employing a threshold whose adaptation seems difficult. The authors of
[66] describe a threshold optimization strategy considering the Neyman-Pearson criteria. In
[65], the authors presented a mathematical solution for IN mitigation utilizing clipping and blanking. In
[67], a comparison of numerous analog domain processing strategies for IN mitigation demonstrates that threshold value selection is the most important aspect for enhancing the efficacy of threshold-assisted nonlinear techniques. Once the threshold varies due to channel circumstances, the model gets mismatched. As a result, extremely impulsive environments have a negative impact on the effectiveness of all conventional threshold-based approaches. In
[59], the authors have successfully used DL approaches for IN mitigation.
Figure 5 presents the DNN performance for IN mitigation in User 1 and compares it with User 2. User 1 uses SIC to reduce inter-user-interference. Thus, it will suffer from IN only. While User 2 is affected by both IN and inter-user interference. Therefore, it has variant BER performance according to SNR values. The results show that the DNN approach can be effectively utilized to overcome IN.
Figure 5. Performance of NOMA user pair through DNN-based technique.
2.7. Transceiver Design
A recent study
[68] focuses on explaining how deep learning aids in overcoming the NOMA as mentioned above difficulties. The multiuser receiver (Rx) is initially improved using deep learning from a model-driven and data-driven perspective. The authors briefly explain how deep learning may enable the optimization of end-to-end NOMA transceivers with practical transmitter (Tx) restrictions or domain knowledge. End-to-end learning is used in NOMA to integrate computation and communication. The authors investigate how deep learning can extract and use upper-layer data for transceiver design. They conclude by outlining some exciting new avenues for deep-learning-enhanced NOMA in mMTC.
Multiuser Detection Design [69][70][71][72][73][74][75][76][77]
Different users’ signals are sent in a non-orthogonal fashion in NOMA. Generally, multiuser detectors (MUDs) are used at the receiving end to differentiate between the overlapping signal streams, thereby minimising inter-user interference (IUI). For several NOMA systems, state-of-the-art MUDs have been created, including parallel interference cancellation (PIC), sequential interference cancellation (SIC), and message-passing algorithms (MPA). Unfortunately, multi-user detection still lacks a unified signal processing framework. By using DNN to improve MUD, researchers may get a more unified architecture, higher detection accuracy, and shorter processing times. DNN-based concepts may be roughly divided into two distinct camps: data-driven and model-driven. Vanilla DNNs are used in a data-driven strategy, which reduces the time spent on design but increases the amount of data needed for training. Alternatively, a model-driven method uses domain-specific knowledge from NOMA to reduce the need for data and increase learning efficiency. In a recent study
[74], the authors propose a DL method that automatically analyzes the CSI of the communication system and detects the original transmit sequences.
Figure 6 shows the symbol error rate (SER) and SNR curve of the numerical simulation. The proposed MIMO-NOMA-DL reached 12.6 dB, whereas the traditional scheme reached 16.2 dB—a difference of approximately 3.6 dB. The authors used powerful DL tools to perform accurate signal detection rather than traditional complex signal processing for channel estimation and demodulation.
Figure 6. Performance comparison of MIMO-NOMA-DL and MIMO-NOMA-SIC.
2.8. DL for Channel Estimation
In a MIMO-NOMA system, an accurate channel estimate is crucial since it influences the system’s performance. Appropriate CSI is necessary for interference cancellation at the receiving end. DNN may be a good option for calculating precise CSI and channel estimates. The researchers of
[74] created an algorithm that automatically assesses and seeks the best logical plan for AWGN channels and MIMO Rayleigh fading channel-state information to recover the signal. For every conceivable scenario, including power allocation parameters, it was demonstrated that DL-based approaches might outperform SIC receivers in terms of symbol error rate (SER) performance. Channel estimation and detection were carried out in batches throughout the training phase. In the testing phase, channel error was included, and the authors investigated how the DL method behaved when the estimated CSI and the actual channel state were different. Throughput is decreased because channel estimate errors cause residual and SIC decoding errors. Impact channel estimate error and reference signaling are reduced by the transmission rate back-off method (in which the transmission rate is regulated). Random beamforming is a useful technique for lowering CSI feedback
[78]. Since flawless interference cancellation depends on the correct CSI estimate, NOMA system performance is impacted. Practically speaking, it is challenging because of the complicated fluctuations in channel conditions brought on by high mobility. Utilizing the spatial diversity of massive MIMO, DL methods may be utilized to evaluate the DOA and real-time channel estimates. The sparse features may be fully extracted and efficiently used in the DL technique to learn the entire system. DL also performs better than traditional approaches when SNR is high
[27].
The articles make it clear that estimate accuracy is a performance parameter and that DL is highly preferred to estimate CSI in real-time with less complexity and pilot overhead than the conventional alternatives. Large datasets with different channel conditions are still difficult for supervised learning, and offline network training takes time. Its efficacy in cases with high mobility is constrained since it is challenging to predict the channel.
2.9. DL for Beamforming and Selection
The performance of 5G technology is also determined by the beamforming process. A quick unsupervised learning-based beamforming design methodology has been put out by authors in
[79]. In this approach, DNN is trained offline and provides real-time assistance for simple neural network tasks once it is online. DNN in the downlink records the channel’s characteristics, takes the channel coefficients as input, and produces a beamformer. Pruned DNN is used because it decreases the parameters and, as a result, the computational complexity and time required by DNN. The simulations showed that, although degrading with increasing SNR and transmit antenna count, deep neural network performance is comparable to that of WMMSE. In
[80], DL-aided hybrid beamforming (HB) is suggested, where supervised learning and an autoencoder build the HB. Compared to other traditional beamforming methods, this approach performed better in the context of bit error rate. A recent study
[81] introduces a novel MIMO-NOMA system that addresses partial CSI feedback. Channel quality information (CQI), the best beam, and beam correlation are used to cluster users. The user pair chosen for clustering had the greatest CQI differential and the highest beam correlation. HB is created through clustering. As an analog beamforming vector, the best beam from a high data rate user is selected. Thus, inter-cluster interference is decreased. For weak users, digital beamforming is used to reduce intra-cluster interference. Furthermore, the authors developed a system with efficient power allocation by optimizing the power differential between UEs in a cluster while subject to rate constraints. The system that was presented had a greater sum rate. With DL, choosing a beam is simpler. Using two optimal beam indices as inputs and an estimated power delay profile (PDP) as a label, the DL model is trained using supervised learning. Adam is used for optimization, cross-entropy is utilized as the cost function, and softmax activation is employed at the output layer, where the number of beams equals the number of neurons
[82]. DNN may be used to perform beam selection and hybrid beamforming with little latency. Additionally, it produces better results from the perspective of summation rate and BER.
2.10. DL for Modulation and Signal Processing
At high SNR, long-short-term memory (LSTM) and the deep residual network (ResNet) may achieve high classification accuracy. Still, the convolutional long-short-term deep neural network (CLDNN) and ResNet performed well at low SNR. Furthermore, principal component analysis and subsampling were used to minimize training time
[83]. In the presence of faulty CSI, CNN for feature extraction and DNN for joint channel equalization and decoding have high accuracy. In terms of BER and decoding rate, DNN outperforms CNN
[84]. A system that combines CNN and LSTM is thought to perform well in automated modulation classification (AMC) at varied SNRs
[78]. For signal demodulation using Rayleigh and AWGN channels, CNN and a bidirectional gated recurrent unit layer known as a mixed neural network model are utilized. CNN is utilized to extract features, whereas RNN is used for time-series analysis
[85]. In
[76], the authors propose a deep residual network-based blind modulation detection technique that uses a noisy joint constellation as input. Wavelet denoising is used to increase constellation quality. To demodulate the signal, the SIC receiver at the distant UE requires information on the modulation mode. This technique significantly reduces signaling overhead while improving service quality in NOMA systems. However, for higher-order modulation, the constellation becomes more difficult. A CNN-based AMC with an extended symbol rate sequence and an estimated SNR is a near approximation to a maximum likelihood-based AMC (ML-AMC), learning from raw data and processing in parallel, making it quicker and better than feature-based approaches and ML-AMC
[86]. A survey on DL in signal recognition reported in
[87] highlighted the difficulty in developing an accurate and effective DL signal recognition system in coexistence. For modulation recognition under various channel impairments and datasets, a modified deep residual network (RN) has been deployed. This outperforms CNN in terms of efficiency. Transfer learning is employed to accelerate the suggested model. In addition, the authors compare the baseline approach and strongly boosted gradient tree classification for radio signal classification utilizing over-the-air observations
[88]. In
[89], the authors propose employing a single DNN for joint optimum MIMO signal detection and channel decoding. The suggested DNN model has the limitation of requiring training for different channel matrices as well as having a high decoding latency. A DNN that can handle multiple channel matrices with a single training is offered as a research path. For modulation categorization, signal identification, and decoding, prominent models include CNN, RN, LSTM, and customized DNN. In a recent work
[90], the authors investigated a deep learning-based SIC scheme for NOMA communication systems and compared its performance with
[19] and
[91] as shown in
Figure 7. The authors propose a convolutional neural network (CNN)-based SIC scheme to enhance the single BS and multiuser NOMA scheme. The proposed CNN-based SIC scheme can effectively mitigate losses resulting from imperfections of the SIC. The findings also indicate that the CNN-based SIC method can achieve good detection performance and relieve conventional SIC impairments.
Figure 7. Sum rate versus SNR for the conventional and proposed SIC schemes with varying power allocations.