Non-orthogonal multiple access (NOMA) has been considered as an essential multiple access technique for enhancing system capacity and spectral efficiency in future communication scenarios. However, the existing NOMA systems have a fundamental limit: high computational complexity and a sharply changing wireless channel make exploiting the characteristics of the channel and deriving the ideal allocation methods very difficult tasks. To break this fundamental limit, in this paper, we propose a novel and effective deep learning (DL)-aided NOMA system, in which several NOMA users with random deployment are served by one base station (BS). Since DL is advantageous in that it allows training the input signals and detecting sharply changing channel conditions, we exploit it to address wireless NOMA channels in an end-to-end manner. Specifically, it is employed in the proposed NOMA system to learn a completely unknown channel environment. A long shortterm memory (LSTM) network based on DL is incorporated into a typical NOMA system, enabling the proposed scheme to detect the channel characteristics automatically. In the proposed strategy, the LSTM is first trained by simulated data under different channel conditions via offline learning, and then the corresponding output data can be obtained based on the current input data used during the online learning process. In general, we build, train and test the proposed cooperative framework to realize automatic encoding, decoding and channel detection in an additive white Gaussian noise (AWGN) channel. Furthermore, we regard one conventional user activity and data detection scheme as an unknown nonlinear mapping operation and use LSTM to approximate it to evaluate the data detection capacity of DL based on NOMA. Simulation results demonstrate that the proposed scheme is robust and efficient compared with conventional approaches. In addition, the accuracy of the LSTMaided NOMA scheme is studied by introducing the well-known 10-fold cross-validation procedure.

Automatic modulation classification (AMC) is a promising technology for identifying modulation types, and deep learning (DL)-based AMC is one of its main research directions. Conventional DL-based AMC methods are centralized solutions (i.e., CentAMC), which are trained on abundant data collected from local clients and stored in the server and generally have advanced performance, but their major problem is the risk of data leakage. Besides, if DL-based AMC is only trained with the data from their corresponding clients, it may exhibit weak performance. Thus, a federated learning (FL)-based AMC (FedeAMC) is proposed under the condition of class imbalance and noise varying. Its advantage is low risk of data leakage without severe performance loss, because data and training are in each local client, while only knowledge (i.e., gradient or model weight), rather than data, is shared with the server. In addition, there is generally class imbalance problem in each local client, and balanced cross entropy is introduced as loss function for solving this problem. Simulation results demonstrated that average accuracy gap between FedeAMC and CentAMC is less than 2%.

The communication between vehicle-to-vehicle (V2V) with high frequency, group sending, group receiving and periodic lead to serious collision of wireless resources and limited system capacity, and the rapid channel changes in high mobility vehicular environments preclude the possibility of collecting accurate instantaneous channel state information at the base station for centralized resource management. In addition, road and traffic safety have very strict requirements for low-latency, high-reliability communication, and information transmission is only valuable within a certain range, and requires timeliness. For the Internet of Vehicles (IoV), it is a fundamental challenge to achieve low latency and high reliability communication for real-time data interaction over short distances in a complex wireless propagation environment, as well as to attenuate and avoid inter-vehicle interference through a reasonable spectrum allocation. To solve the above problems, this paper proposes a resource allocation (RA) method using dueling double deep-Q network reinforcement learning (RL) with low-latitude fingerprints and soft-update architecture (D3QN-LS) while constructing a multi-agent model based on a Manhattan grid layout urban virtual environment, with communication links between V2V links acting as agents to reuse vehicle-to-infrastructure (V2I) spectrum resources. These agents work in concert, interact with the environment, receive appropriate observations, rewards, and ultimately learn to improve power and spectrum allocation to provide a better entertainment experience and safer driving environment for users. We also address the shortcomings in existing literature studies. Firstly, the low upper limit of transmitted data volume; Secondly, the artificial assumption that the spectrum is consistent with the number of V2V links, which is insufficient for future applications in spectrum shortage sections. We accordingly extend the amount of transmitted data while adding a scenario where the spectrum resources are relatively short, i.e., the number of V2V links is significantly larger than the amount of spectrum. Experimental results show that with proper training mechanism and reward function construction, multiple intelligence can cooperate effectively. Therefore, the total link capacity of the V2I link and the success rate of periodic security information transmission are further improved.

Automatic modulation classification (AMC) is considered as one of the most critical technologies for the non-cooperative communication systems. Recently, deep learning (DL) based AMC (DL-AMC) methods have been attracted intensely attentions due to their preferable performance. However, most of DL-AMC methods are concentrated in the single-input and single-output (SISO) systems, while there are few works on DL-based AMC methods in multiple-input and multiple-output (MIMO) systems. Hence, we propose a convolutional neural network (CNN) and zero-forcing equalization-based AMC (ZF-AMC) method for MIMO systems. Simulation results demonstrate that the CNN-based ZF-AMC method achieves better performance than the artificial neural network (ANN) and high order cumulants (HOC)-based AMC method under the condition of the perfect channel statement information (CSI). Moreover, we also explore the impact of the imperfect CSI on the performance of the ZF-AMC method, due to its difficult acquisition of the perfect CSI. Simulation results demonstrated that the classification performance is not only influenced by the channel error, but also associated with the number of the transmitted and received antennas.