Wireless communications in aircraft cabin environments have drawn widespread attention with the increase of application requirements. To ensure reliable and stable in-cabin communications, the investigation of channel parameters such as path loss is necessary. In this paper, four machine learning methods, including back propagation neural network (BPNN), support vector regression (SVR), random forest, and AdaBoost, are used to build path loss prediction models for an MD-82 aircraft cabin. Firstly, machine-learning-based models are designed to predict the path loss values at different locations at a fixed frequency. It is shown that these models fit the measured data well, e.g., at 2.4 GHz central frequency the root mean square errors (RMSEs) of BPNN, SVR, random forest, and AdaBoost predictors are 1.90 dB, 2.20 dB, 1.76 dB, and 2.12 dB. Subsequent research is engaged to forecast path loss at a new frequency based on available information at known frequencies. Additionally, to solve the data limitation problem at the new frequency, we propose a path loss prediction scheme combining empirical models and machine-learning-based models. This scheme uses estimated values generated by the empirical model according to prior information to expand the training set. To verify the performance of this scheme, measured samples at 2.4 GHz and 3.52 GHz, as well as samples generated by the empirical model are employed as the training set for the path loss prediction at 5.8 GHz. The RMSEs of BPNN, SVR, random forest, and AdaBoost models are 2.49 dB, 2.78 dB, 2.54 dB, and 3.76 dB. In contrast, without samples generated by the empirical model, the RMSEs of those models are 3.84 dB, 4.94 dB, 6.57 dB, and 6.77 dB. Results show that the proposed data expansion scheme improves prediction performance when there are few measurement samples at the new frequency.

The dynamic position deployment and resource allocation of the unmanned aerial vehicle (UAV) communication networks has great significance in terms of interference management, coverage enhancement, and capacity improvement. Since the transmission power and energy resources of the UAVs are limited and the actual communication environment is complex and time-varying, it is challenging for the multiple UAVs to dynamically make decisions to ensure the communication performance of the system. Meanwhile, the centralized architecture may generate a certain degree of communication delay and affect communication efficiency. Facing this challenge, a resource allocation algorithm for the UAV networks based on multi-agent collaborative environment learning is proposed. This method is based on a distributed architecture. Each UAV is modeled as an independent agent, which improves the utility of the UAV networks through the dynamic selection decisions of its deployment position, transmission power, and occupied sub-channels. Each UAV learns the mapping of the network information to the position deployment and resource selection decisions based on the reinforcement learning algorithm according to partial of the state information it can observe. For the overall network, a multi-agent reinforcement learning method based on federated learning is designed on the purpose of realizing information interaction and combined dispatching of the UAVs. In the multi-agent system, the framework of federated learning is introduced to realize the sharing of non-privacy data among the UAVs. Simulation results indicate that the proposed method can effectively improve the network utility compared with the multi-agent deep reinforcement learning algorithm without information interaction.

Modulation recognition plays an indispensable role in the field of wireless communications. In this paper, a novel attention cooperative framework based on deep learning is proposed to improve the accuracy of the automatic modulation recognition (AMR). Within this framework, a convolutional neural network (CNN), a recurrent neural network (RNN), and a generative adversarial network (GAN) are constructed to cooperate in AMR. A cyclic connected CNN (CCNN) is designed to extract spatial features of the received signal, and a bidirectional RNN (BRNN) is constructed for obtaining temporal features. To take full advantage of the complementarity and relevance between the spatial and temporal features, a fusion strategy based on global average and max pooling (GAMP) is proposed. To deal with different influence levels of the signal feature maps, we present the attention mechanism in this framework to realize recalibration. Besides, modulation recognition based on deep learning requires numerous data for training purposes, which is difficult to achieve in practical AMR applications. Therefore, an auxiliary classification GAN (ACGAN) is developed as a generator to expand the training set, and we modify the loss function of ACGAN to accommodate the processing of the actual in-phase and quadrature (I/Q) signal data. Considering the difference in distribution between generated data and real data, we propose a novel auxiliary weighing loss function to achieve higher recognition accuracy. Experimental results on the dataset RML2016.10a show that the proposed framework outperforms existing deep learning-based approaches and achieves 94% accuracy at high signal to noise ratio (SNR).

In this paper, we propose a robust end-to-end classification model, Graph-in-Graph Neural Network (GIGNet), for automatic modulation recognition (AMR). In GIGNet, multi-level graph neural networks (GNNs) are utilized to extract internal graph-based features from signal samples and correlation information between different signals treated as nodes in a graph. Specifically, a graph-level GNN is utilized to extract local and global features of signal samples transformed into graphs. Next, a method for constructing a graph that corresponds signals to nodes is proposed to assess the degree of association between nodes and to find closer neighbors of nodes. These closer neighbors enable the subsequent node-level GNN to incorporate appropriate correlation information for the further classification task. Compared to classical deep learning models and existing GNN-based models, experimental results justify the advantages of the proposed GIGNet model on recognition accuracy and robustness at low signal-to-noise ratio (SNR).

In time-variant high-speed train (HST) radio channels, the scattering environment changes rapidly with the movement of terminals, leading to a serious deterioration in communication quality. In the system-and link-level simulation of HST channels, this non-stationarity should be characterized and modeled properly. In this paper, the sizes of the quasi-stationary regions are quantified to measure the significant changes in channel statistics, namely, the average power delay profile (APDP) and correlation matrix distance (CMD), based on a measurement campaign conducted at 2.4 GHz. Furthermore, parameters of the multi-path components (MPCs) are estimated and a novel clustering-tracking-identifying algorithm is designed to separate MPCs into line-of-sight (LOS), periodic reflecting clusters (PRCs) from power supply pillars along the railway, and random scattering clusters (RSCs). Then, a non-stationary geometry-cluster-based stochastic model is proposed for viaduct and hilly terrain scenarios. Furthermore, the proposed model is verified by measured channel statistics such as the Rician K factor and the root mean square delay spread. The temporal autocorrelation function and the spatial cross-correlation function are presented. Quasi-stationary regions of the model are analyzed and compared with the measured data, the standardized IMT-Advanced (IMT-A) channel model, and a published non-stationary IMT-A channel model. The good agreement between the proposed model and the measured data demonstrates the ability of the model to characterize the non-stationary features of propagation environments in HST scenarios.

High-altitude platforms (HAPs) are considered to be the most important equipment for next-generation wireless communication technologies. In this paper, we investigate the channel characteristics under the configurations of massive multiple-input multiple-output (MIMO) space and large bandwidth at millimeter-wave (mmWave) bands, along with the moving essence of the HAP and ground terminals. A non-stationary three-dimensional (3D) geometry-based stochastic model (GBSM) is proposed for a HAP communication system. We use a cylinder-based geometric modeling method to construct the channel and derive the channel impulse response (CIR). Additionally, the birth–death process of the scatterers is enclosed using the Markov process. Large-scale parameters such as free space loss and rainfall attenuation are also taken into consideration. Due to the relative motion between HAP and ground terminals, the massive MIMO space, and the wide bandwidth in the mmWave band, the channel characteristics of HAP exhibit non-stationarities in time, space, and frequency domains. By deriving the temporal auto-correlation function (ACF), we explore the non-stationarity in the time domain and the impact of various parameters on the correlations across the HAP-MIMO channels. The spatial cross-correlation function (CCF) for massive MIMO scenarios, and the frequency correlation function (FCF) in the mmWave bands are also considered. Moreover, we conduct simulation research using MATLAB. Simulation results show that the theoretical results align well with the simulation results, and this highlights the fact that the constructed 3D GBSM can characterize the non-stationary characteristics of HAP-MIMO channels across the time, space, and frequency domains.