Unmanned Aerial Vehicles (UAVs) have attacted much attention in the field of wireless communication due to its agility and altitude. UAVs can be used as low-altitude aerial base stations (UAV-BSs) to provide communication services for ground devices (GDs) in various scenarios, such as emergency communication and traffic offloading in hotspots. However, due to the limited communication ranges and high prices of commercial UAV-BSs, covering a target area all the time with sufficient UAVs is quite challenging, especially under dynamic environment. We need to design the trajectory of the UAV-BSs to optimize system performance. Most existing works focus on the energy-efficient coverage and throughput maximization but ignore the fairness of communication service, especially the fairness at user-level. Besides, reinforcement learning is suitable for solving decision problems in dynamic environments. However, most existing works use centralized deep reinforcement learning (DRL) approaches. Due to the scalability and low time complexity, a distributed DRL approach is more suitable for multiple UAV-BSs communication system in dynamic environment. Unlike previous works, we characterize the fairness at user-level based on proportional fairness scheduling and formulate a weighted-throughput maximization problem via designing UAV-BSs’ trajectory. Then we model the dynamic deploymentproblem of UAV-BSs as a Markov game and propose a multi-agent deep reinforcement learning-based distributed UAV-BSs control approach named MAUC. MAUC approach adopts the framework of centralized training with distributed execution. Simulation results show that the MAUC can improve fairness of communication service by sacrificing a small amount of throughput.

The marine Internet of things (MIoT) is the application of the Internet of things technology in the marine field. Nowadays, with the arrival of the era of big data, the MIoT architecture has been transformed from cloud computing architecture to edge computing architecture. However, due to the lack of trust among edge computing participants, new solutions with higher security need to be proposed. In the current solutions, some use blockchain technology to solve data security problems while some use federated learning technology to solve privacy problems, but these methods neither combine with the special environment of the ocean nor consider the security of task publishers. In this article, we propose a secure sharing method of MIoT data under an edge computing framework based on federated learning and blockchain technology. Combining its special distributed architecture with the MIoT edge computing architecture, federated learning ensures the privacy of nodes. The blockchain serves as a decentralized way, which stores federated learning workers to achieve nontampering and security. We propose a concept of quality and reputation as the metrics of selection for federated learning workers. Meanwhile, we design a quality proof mechanism [proof of quality (PoQ)] and apply it to the blockchain, making the edge nodes recorded in the blockchain more high-quality. In addition, a marine environment model is built in this article, and the analysis based on this model makes the method proposed in this article more applicable to the marine environment. The numerical results obtained from the simulation experiments clearly show that the proposed scheme can significantly improve the learning accuracy under the premise of ensuring the safety and reliability of the marine environment.

Network lifetime is one of the most critical issues for wireless sensor networks (WSNs) since most sensors are equipped with non-rechargeable batteries with limited energy. To prolong the lifetime of a WSN, one common approach is to dynamically schedule sensors' active/sleep cycles (i.e., duty cycles) with sleep scheduling algorithm. In this paper, we propose a new sleep scheduling algorithm, named EC-CKN (Energy Consumed uniformly-Connected K-Neighborhood) algorithm, to prolong the network lifetime. The algorithm EC-CKN, which takes the nodes' residual energy information as the parameter to decide whether a node to be active or sleep, not only can achieve the k-connected neighborhoods problem, but also can assure the k awake neighbor nodes have more residual energy than other neighbor nodes at the current epoch. Based on the algorithm EC-CKN, we can obtain the state transition probability at the n'th epoch, and upper bound and lower bound of the network lifetime by Markov chain and Markov decision chain.

Recently, using unmanned aerial vehicles (UAVs) to collect information from distributed sensors has become one of the hotspots in the Internet of Things (IoT) research. However, previous studies on the UAV-assisted data acquisition systems focused mainly on shortening the acquisition time, reducing the energy consumption, and increasing the amount of collected data, but it lacked the optimization of data freshness. Moreover, we hope that UAVs can perform long-term data collection tasks in dynamic scenarios within a constantly changing age of information (AoI) and within their own power levels. Therefore, we aim to maximize the quality of service (QoS) based on the freshness of data, while considering the endurance of the UAVs. Since our scenario is not an inertial order decision process with uniform time slots, we first transform the optimization problem into a semi-Markov decision process (SMDP) through modeling, and then we propose a hierarchical deep Q-network (DQN)-based path-planning algorithm to learn the optimal strategy. The simulation results show that the algorithm is better than the benchmark algorithm, and the tradeoff between the system QoS and the safe power state can be achieved by adjusting the parameter ${\beta }_e$.

By deploying edge servers on the network edge, mobile edge computing network strengthens the real‐time processing ability near the end devices and releases the huge load pressure of the core network. Considering the limited computing or storage resources on the edge server side, the workload allocation among edge servers for each Internet of Things (IoT) application affects the response time of the application’s requests. Hence, when the access devices of the edge server are deployed intensively, the workload allocation becomes a key factor affecting the quality of user experience (QoE). To solve this problem, this paper proposes an edge workload allocation scheme, which uses application prediction (AP) algorithm to minimize response delay. This problem has been proved to be a NP hard problem. First, in the application prediction model, long short‐term memory (LSTM) method is proposed to predict the tasks of future access devices. Second, based on the prediction results, the edge workload allocation is divided into two subproblems to solve, which are the task assignment subproblem and the resource allocation subproblem. Using historical execution data, we can solve the problem in linear time. The simulation results show that the proposed AP algorithm can effectively reduce the response delay of the device and the average completion time of the task sequence and approach the theoretical optimal allocation results.

WiFi-based human activity recognition (HAR) plays an essential role in various applications such as security surveillance, health monitoring, and smart home. Existing HAR methods, though yielding promising performance in indoor scenarios, highly depend on a massive labeled dataset for training which is extremely difficult to acquire in practical applications. In this paper, we present an automatic data labeling and HAR system, termed AutoDLAR. Taking a semi-supervised cross-modal learning framework with a hybrid loss function as the core, AutoDLAR transfers rich visual information to automatically label WiFi signals for WiFi-based HAR. Specifically, we devise a lightweight and multi-view WiFi sensing model with a parallel feature embedding method to accurately identify activities and accelerate recognition speed. Then, we exploit the video data to fine-tune a well-established visual HAR model, generating effective pseudo-labels for guiding the WiFi model’s training. We also build a synchronized Video-WiFi dataset with seven types of human activities under different scenarios to enable training and validating the semi-supervised HAR system. Extensive experiments on our collected activity dataset and the emotion recognition benchmark demonstrate that AutoDLAR attains an average accuracy of over 95.89% without manual labeling and only spends the inference time of 3.35 ms, outperforming the state-of-the-art (SOTA) methods.