The emergence of mobile edge computing (MEC) supports large-scale computing demand in the Internet of Things era by moving computing pressure from cloud center to the edge. However, computing resources of edge servers are scattered and restricted. A carefully designed computation offloading mechanism is the key to MEC. Besides, it is necessary for cloud centers and edge servers to participate in computation together. Most of the existing researches focus only on a single computation architecture to offload computing tasks and neglect the impact of load balancing on offloading efficiency. In this paper, we consider a cloud-edge-device collaborative computing scenario, leveraging the advantages of different computing levels. In order to direct computing task to service node with sufficient computing resources and avoid edge servers overloading or being idle all the time, we introduce an occupancy table to dynamically track the edge load status. Then we propose a deep Q-network (DQN) based collaborative computation offloading (DQNCCO) method for edge load balancing, jointly minimizing delay, energy consumption and edge load balancing. Experimental results show that the proposed method can balance the use of the whole edge computing resources while realizing suitable fine-grained offloading decisions.
Photosensor which converts the optical signal into electrical signal is an essential element in optoelectronic integrated circuits. Lead halide perovskites have been widely investigated for photosensor applications due to their excellent photoelectronic properties and superior stability. Moreover, single crystalline perovskite films have fewer defects comparing with polycrystalline films, which can further improve the material properties. In this study, vapour phase epitaxy was used to grow large scale single crystalline CsPbBr3 thin films with lateral sizes up to 2.25 cm 2 , Microstructure characterization of the films reveals that they are of high quality, with smooth surfaces. Furthermore, photo sensors constructed on these single crystalline films exhibits excellent on/off current ratio of 2.4×10 5 . Our work not only presents a facile method for producing large area CsPbBr3 single crystalline films but also provide a platform for developing high-quality halide per-ovskite photo sensors and their large-scale integration.
Significant progress has been made in the supervised semantic segmentation of remote sensing images (RSI), mainly attributed to the design of models and the large amount of high-quality pixel-by-pixel annotated data. However, pixel-level annotated data is extremely costly and requires specific knowledge. In addition, different application scenarios usually require different target classes of segmentation, which makes the annotated data and the trained models lack generalization. In this paper, we devise a semantic-aware self-supervised learning method termed SASS which can learn potential patterns from massive unlabeled remote sensing images and obtain a dense feature representation with se-mantic information. Besides, an efficient contrastive-based loss function is designed to optimize the proposed SASS. After training, the features extracted by SASS can be borrowed to implement unsupervised semantic segmentation (USSS) of remote sensing images. Experiments show our proposed method surpasses recent works by a large margin, which achieves 55.7% mIoU and 71.89% overall accuracy in Potsdam-3 dataset with the improvement of 5.12% and 6%.
Space-Air-Ground Integrated Network (SAGIN) is a promising architecture for next-generation wireless networks. It combines satellite networks, aerial networks, and terrestrial networks, offering ubiquitous global network services to ground users and enhancing connectivity for a wide range of wireless communications. Additionally, free-space optical (FSO) communication has gained more attention recently because of its energy efficiency and exceptional high-speed data delivery capabilities. However, the data transmission efficiency in SAGIN remains limited due to the dynamic and time- varying network topology and long-distance transmission link. Moreover, efficiently utilizing the limited power resources in satellites and aircraft to achieve effective data routing is still a crucial challenge. To address these issues, we construct an FSO/radio frequency (RF) space-air-ground integrated network that aims to enable large-scale data trans-mission as well as degrading the burden on terrestrial networks. Furthermore, a decentralized deep-Q network-based reinforcement learning with two hops node information is proposed to execute dynamic energy-efficient hybrid routing while ensuring system load balancing. The experimental results demonstrate that the proposed method reaches significant network performance improvement compared with baseline methods.
In this paper we propose and implements a standing long jump movements analysis algorithm based on artificial intelligence to help students improve their performance. We select the BlazePose, a lightweight neural network under the Mediapipe framework for human pose estimation. We extracte the coordinate information of body joints with Mediapipe, and use the extracted coordinate information to design the algorithm of standing long jump action segmentation and key frame selection for quantitative analysis movements. To evaluate the effectiveness of the project, we recruit 24 students to attend the Experiments with two rounds of tests, Experiment results show that the proposed algorithm can automatically detect and identify problems and deficiencies in key movements, and give suggestions for improvement. By analyzing the data of jumpers' take-off skills, posture control and landing methods, we can provide suggestions to help them improve their skills and performance.
Real-World-Data always follows natural patterns with various undesirable situations. Chinese language is more prone to susceptibility and influence for its rich semantic information. When confronted with severely imbalanced distributions, the existing methods always tend to favor the majority class while neglecting the minority class. However, the performance of the minority class cannot be ignored, and it is particularly crucial not to overlook the performance of the minority class when dealing with situations involving offensive text filtering. In this work, we mainly focus on three aspects, (i): We developed a Chinese Offensive Message Text Dataset, which includes seven categories: Normal, Spam, Advertisement, Fraud, Violent, Pornography, Politics, exhibiting a significant imbalance between Normal and Abnormal. (ii): We proposed a feature space manipulation which combines samples optimization and spatial position optimization for the majority / minority and a hybrid feature prediction module which integrates multi-level feature fusion for comprehensive utilization of sample information. (iii): Experimental results show that our method achieves significant improvement on the offensive classes in our dataset.
The direct release of circularly polarized (CP) light from an emitting layer (EML) enhances the light extraction in typical organic light-emitting diode displays using a circular polarizer. High degree of CP light is generated from a twisted arrangement of mesogenic luminophores induced by adding chiral agents to mesogenic organic light-emitting molecules. Although quite a lot of studies have been carried out to enhance the degree of CP light emitted from this twisted structure, a comprehensive and theoretical analysis for elementary parameters, such as birefringence and twist angles of the mesogenic emitting materials, has not been completely conducted. In this investigation, we derive the degree of CP light with photoluminescence mechanisms by modeling the twisted structure of mesogenic organic light-emitting molecules and by optical calculating based on Stokes parameters and 4 x 4 Mueller matrix. Also, we analyze the simulation results of CP emission with various parameters investigated under various conditions. The results of our study are expected to help understand the emission mechanism of CP light and serve as the basis for analysis to improve the performance of displays using CP light and devices used in related areas.
This paper presents an adaptive dead-time control for the zero-voltage switching operation of a wireless power amplifier to improve efficiency. In this paper, a power amplifier is designed using a differential class-D structure for high efficiency and maximum output power. The proposed circuit enables zero-voltage switching operation of the power amplifier through a controller that detects various system parameter changes and adaptively adjusts the dead-time. The proposed circuit is fabricated in a CMOS 180-nm standard process. The proposed dead-time control system for zero-voltage switching is verified through the simulation result.
With the continuous development of mobile communication technology and the arrival of the 5G era, the demand for mobile data traffic is increasing. Accurate traffic prediction is of great significance in meeting users' data transmission needs and optimizing network resource allocation. However, due to the uneven distribution of cellular network traffic in large-scale urban area, the current cellular traffic prediction methods based on deep learning suffer from performance degradation as well as large storage overhead. Therefore, we design a Low Overhead wireless Traffic prediction Network (LOTNet) for performance enhancement. Firstly, to improve prediction performance under the large-scale city area, the Context Embedding and Multi-Scale Spatiotemporal Expression Long Short Term Memory (CMS-LSTM) structure and the dual attention mechanism are used in the prediction module for the extraction of spatiotemporal features. Secondly, to reduce the storage overhead caused by large-scale prediction, we use the autoencoder structure for hotspot cells extracting. The experiments on real cellular traffic data proves the effectiveness of our scheme.
Ensuring the quantity and quality safety of grain storage is very important for social stability and human health. However, fungal spore infection has become a challenge for grain storage, affecting grain quality and causing economic losses. The traditional fungal spore detection method is mainly operated manually by professionals, which is time-consuming and tedious. This study establishes an image dataset containing 36,643 fungal spores and proposes an automatic detection algorithm for fungal spores based on deep learning and image processing techniques. Due to the similar morphological characteristics of different fungal species, the insufficient visibility of color features, and the occurrence of overlapping and gathering of fungal spores, the detection of fungal spores in microscopic images faces challenges. Thus, an attention mechanism is incorporated into the YOLOv5 algorithm, along with optimizations in the loss function and post-processing of the prediction results. The proposed algorithm achieves the mAP of 0.899 in fungal spores detection, The experimental results demonstrate the efficacy of the proposed algorithm in detecting fungal spores within microscopic images and thereby offer substantial evidence supporting the early detection of fungal infection.
Sleep disorders seriously affect human health. Leveraging deep learning methods and Electroencephalography, automatic sleep staging can aid experts in accurately diagnosing patients' sleep disorders. However, the imbalance of the training data undermines the learning of minority class features. Besides, the performance of the automatic sleep staging model obtained on the training data tends to decrease on the practical data due to the difference in data distribution. As a result, an unsupervised domain adaptation algorithm combined with class rebalancing strategy and semi-supervised learning is proposed to solve the above problems. To alleviate data imbalance in sleep staging datasets, our paper devises a balanced sampler. Random logit interpolation and relative confidence threshold are introduced to improve the accuracy of pseudo-labels. Moreover, distribution alignment is introduced to mitigate the dissimilarity in data distribution between the source and target domains. Through experiments, the effectiveness of the proposed method is proved on the SHHS, Sleep-EDF and ISRUC-Sleep datasets. The average improvement in accuracy is around 5.27%. Both the F1 score and recall rate have also been significantly improved.
Tracing and restoring the bat trajectory from a table tennis match video plays a critical role in table tennis technique and tactics. However, directly detecting the bat location in each frame is challenging, because the bat shapes are changed be-cause of various movement speeds, and can sometimes be am-biguous. In addition, there are rare bat-related datasets. In this paper, we propose a novel method that decomposes the prob-lem into two stages. In the first stage, we employ YOLOv5 for bat detection in each frame, and filter out erroneous candidate boxes. In the second stage, we classify all frames into two cat-egories. For frames corresponding to forward swing motions, we utilize a temporal prediction model that integrates human body keypoints and confidence scores to fill missing values. For other frames, missing values are completed using inter-polation methods. By combining these two results, we obtain a complete tennis bat trajectory with reduced errors in trajec-tory reconstruction. To evaluate the effectiveness of the pro-posed method, we build a video dataset that with labeled bat locations in each frame. The evaluation results demonstrate that the proposed method outperforms traditional methods on precision performance metric.
Synthetic aperture radar (SAR) imagery is susceptible to the non-systematic motion error, which will defocus the data in the training set. Also, traditional convolutional neural networks (CNN) are random in the features utilized, many of which are useless for practical target recognition. To this end, quantizable feature oriented utilitarian CNN (QFO-UCNN) is proposed in this paper, where the defocusing can be alleviated via a quantizable entropy norm regularizer. To overcome the non-convexity of the optimization function, a tight surrogate function is designed to ensure the closed-form solution of the focusing feature. Besides, the dominant feature of sparsity is encoded by a l 1 norm, where the sparse feature is quantified via a closed-form soft-thresholding operator. In such cases, two useful features are quantified and constrained in the CNN, which improves the intentionality of the intended SAR target recognition. The canonical MSTAR dataset is employed to verify the effectiveness of the proposed algorithm, when the data encounters the non-systematic and multiplicative phase error.
Due to prolonged high-load operation and the impact of external environmental factors, insulators often undergo selfblasting, leading to significant disruptions in the functioning of the power system. This research introduces a novel approach for detecting insulator self-blasting incidents based on the YOLOv4 model. This method combines several advanced techniques, including feature fusion, attention mechanisms, and dilated convolution, to achieve highly accurate detection results. In the proposed method, shallow-level information is utilized to compensate for the loss of intricate details in deep-level features. Additionally, an attention mechanism is employed to enhance the precision of target positioning. Furthermore, dilated convolution is integrated to capture vital target context information and expand the network's receptive field. The experimental results demonstrate a remarkable improvement in the average detection precision of the enhanced model, showcasing an impressive increase of 6.86% compared to the baseline YOLOv4 model. This advancement promises to contribute significantly to the early identification and mitigation of insulator self-blasting issues, thereby enhancing the overall reliability and performance of the power system.
We propose a Multi-graph Attention spatial-temporal graph convolutional network (MGA-STGCN) for AHP risk forecasting. To describe the temporal and spatial features of the area, we use different kinds of dataset such as satellite images, crash records, taxi trip records, weather records and so on to construct feature matrix and adjacency matrix. We use the multi-graph fusion unit to construct the graph matrix between the blocks to reflect their spatial relationship, including the distance relationship between the two blocks in geographical location and the similarity of the two grids in satellite images. In addition, we improve STGCN by introducing an additional attention unit to help the model adaptively allocate the input data. After training the parameters of the model, the regional risk prediction is simulated, and the performance of the model is evaluated. It is found that compared with the baseline of traditional linear regression, CNN, LSTM and other prediction models, the MGA-STGCN has better performance and prediction accuracy.
The field of artificial intelligence is advancing rapidly, with reinforcement learning making significant strides in solving various sequential decision problems in machine learning. As research progresses, multi-agent reinforcement learning has emerged in the field of reinforcement learning and has been applied to numerous domains. Vehicle formation is an important means of transportation for reducing vehicle energy consumption and improving air quality. Controlling sparse vehicles on the road to form formations is a fascinating research topic. In this paper, we propose a planning framework for formation control based on federated learning and multi-agent reinforcement learning to address this problem. We model vehicle energy consumption to accurately assess energy usage during vehicle formation. Additionally, we incorporate reinforcement learning algorithms into the vehicle formation process to enable multi-vehicle asynchronous decision-making and save formation time. We also introduce federated learning into the training process to significantly reduce overall system communication.
In recent years, low Earth orbit (LEO) satellite networks have emerged as the primary service provider for future mobile communication systems. As the network scale increases, existing routing algorithms have become increasingly time-consuming, resulting in network load imbalance and potential communication delays that may not meet the needs of users. Therefore, this paper proposes a multi-path routing algorithm to jointly optimize the Quality-of-Service (QoS) and the load-balancing for cluster-based satellite networks. To account for the rapid changes in the satellite dynamic topology, an improved clustering structure is designed with two distinct phases: initialization and maintenance. Our strategy utilizes the concept of path deviation in the recursive method to quickly identify a multi-path routing that meets the QoS constraints and evenly distributes the traffic, thereby ensuring network stability and minimizing computational overhead. Simulation results indicate that the proposed algorithm significantly improves load-balancing in satellite communication traffic and enhances QoS for users. Overall, the algorithm represents a promising solution to the increasingly complex routing challenges posed by LEO satellite networks and demonstrates the potential to enhance the quality and efficiency of mobile communication systems in the future.
This paper introduces mmWave channel state information (CSI) acquisition testbed for the demonstration of several algorithms in mmWave communications research fields. The testbed comprises multiple field-programmable gate arrays (FPGA), a radio frequency (RF) converter, and a phased array antenna for mmWave signal generation and high-speed data processing. We also propose system design and implementation for the control of mmWave CSI acquisition testbed by NI LabView programming software. The experimental results show that received mmWave pilot signal data is adequately accomplished via the proposed testbed, which is an important prerequisite for channel estimation.
Most existing methods for question generation create questions purely depend on the source text. The problem is that these methods tend to directly copy the expressions from given passages and lack the sufficient capability to ask diverse and human-like questions. To alleviate this problem, we present a simple and effective Retrieve-Generate-Rerank (RGR) framework that guides the generation with references in training set. Specifically, for every inputted passage and answer, a variety of references are retrieved, with irrelevant information filtered to produce clues. Each clue guides the generation of a question with different patterns and expressions. These questions are then re-ranked to find the most human-like outcome. Experimental results show that our approach consistently improve the performance of existing models on two datasets, and achieves state-of-the-art result on NewsQA. In addition, further investigation and human evaluation demonstrate that our method can generate more diverse and consistent questions.
In this paper, we proposed charge pump system designed specifically for energy harvesting at low voltage and low frequency, focusing on the twistron, a carbon nanotube yarn-based harvester. The whole system is composed of a transformer, a voltage doubler, and the charge pump system. Through simulations, we demonstrate that the charge pump system is capable of generating output voltage from 3.97 V to 4.47 V for the load resistance with 10 kΩ 1 MΩ when the input voltage of the charge pump system is set to 1.03 V. The measured power conversion efficiency (PCE) reaches up to 79%.
