Intelligent fault-tolerant (FT) computing has recently demonstrated significant advantages in predicting and diagnosing faults proactively, thereby ensuring reliable service delivery. However, due to the heterogeneity of fault knowledge, dynamic workloads, and limited data support, existing deep learning-based FT algorithms face challenges in fault detection quality and training efficiency. This is primarily because their homogenization of fault knowledge perception difficuties to fully capture diverse and complex fault patterns. To address these challenges, we propose FT-MoE, a sustainable-learning fault-tolerant computing framework based on a dual-path architecture for high-accuracy fault detection and classification. This model employs a mixture-of-experts (MoE) architecture, enabling different parameters to learn distinct fault knowledge. Additionally, we adopt a two-stage learning scheme that combines comprehensive offline training with continual online tuning, allowing the model to adaptively optimize its parameters in response to evolving real-time workloads. To facilitate realistic evaluation, we construct a new fault detection and classification dataset for edge networks, comprising 10,000 intervals with fine-grained resource features, surpassing existing datasets in both scale and granularity. Finally, we conduct extensive experiments on the FT benchmark to verify the effectiveness of FT-MoE. Results demonstrate that our model outperforms state-of-the-art methods.

With the popularity of electric vehicles (EVs), wireless power transmission (WPT) technology has become a hot research topic for next-generation battery charging technology. However, the vulnerability of wireless networks to malicious interference attacks is inherited by WPT. To alleviate the privacy and security issues of WPT, we propose a novel FedDQ, a federated deep reinforcement learning with Q-ensemble, to cope with interference attacks in EV wireless charging network environments. Federated learning protects the security privacy of EVs by training a global model that exploits the property that data and models will not be transmitted. In order to trade-off the training cost and efficiency, we introduce offline-to-online training models by pre-training the offline Q-network with pre-collected data, and the trained model serves as an initialization of the online model. Then, the online Q-network is obtained by weakening or removing the original pessimistic constraints to enhance the training speed. Secondly, we introduce the intelligent reflective surface (IRS) to enhance the security performance of WPT by modifying the IRS phase shift and amplitude to cancel the malicious interference signal. Experimental results show that our proposed FedDQ algorithm has superior performance and outperforms existing baseline methods in terms of anti-jamming metrics.

Transformer-based models have recently adopted increasingly complex structure (e.g., deeper or wider stacked network) to promote the representation learning capabilities of vision recognition. However, progressively deeper or wider stacked network cause the expensive computation cost, which hinders their effective deployment in resource-constrained edge clouds or end devices. In this paper, we propose DTSNet, a dynamic transformer slimming model, which scales vision transformers (ViTs) down across layers from both of the model depth and input width. This is the first time to explore the joint reduction of input tokens and model parameters for ViTs under maintaining performance. Specifically, DTSNet adopts a diversity-enhanced weight sharing module to reduce network parameters, where the weight knowledge of multiple adjacent blocks is effectively integrated into one block. Furthermore, DTSNet designs a unified and massively scalable token pruning mechanism that dynamically discarding less important tokens with a model-driven manner, by introducing a series of discriminant parameters, which is a simple change to the common architecture of vision transformers. Extensive experiments are conducted to verify that DTSNet is able to yield high efficacy in compressing parameter space and accelerating model inference. DTSNet-T/-S/-B on ImageNet achieves 3.0M/11.1M/42.9M parameters and 0.8/2.9/13.7 GFLOPs, where number of parameters are reduced by 48%$\sim$51% and inference speed are improved by 1.3$\times \sim 1.5\times$. Experiments results on semantic segmentation and object detection dataset further demonstrate the potential of DTSNet on complex dense prediction tasks. Code will be available upon publication.

Intelligent adaptive bitrate (ABR) schemes have been widely recognized for their excellent learning strategies. However, existing intelligent ABR methods have limitations, i.e., the lack of logical reasoning capability for video-aware symbolic representations leads to low sampling efficiency and fails to achieve the optimal performance of Bitrate Adaptation. We introduce NeuroBA, a learning-based approach to realize ABR using neuro-symbolic deep reinforcement learning. NeuroBA trains a neuro-symbolic deep network model without making any assumptions about the edge video scene and without relying on a predefined model. Instead, it enables bitrate decision-making under uncertainty and partial observability by knowledge-driven video quality perception in symbolic first-order logic. To enhance wireless signals, we have introduced Intelligent Reflecting Surface (IRS) technology to address this issue. By dynamically adjusting the phase shift of IRS, the throughput performance of wireless networks is significantly improved. Based on trace-driven and real-world experiments covering a variety of edge video scenarios, and network performance metrics, NeuroBA is compared with state-of-the-art ABR schemes, and NeuroBA exhibits superior performance, with an average QoE improvement of 16.58% (BOLA)-25.34% (Fugu). In particular, it outperforms existing baseline approaches even without pre-programmed models and network scenarios assumed for the edge network.

Video streaming has emerged as a widely used Internet service, in which adaptive bitrate (ABR) algorithms play a critical role in delivering high quality of experience (QoE). However, existing learning-based ABR methods often suffer from limited generalization in unseen and dynamically changing network conditions. Although some meta-reinforcement learning techniques have been proposed to mitigate this issue, they generally depend on additional online training or fine-tuning. To overcome these limitations, this paper introduces EAStream, an environment-aware ABR algorithm based on meta-reinforcement learning for reliable video streaming services. The method employs a variational autoencoder to extract a latent representation of the current network environment from historical interaction data. This latent variable, along with the current system state, is fed into a policy network that perceives network conditions in real time and adapts bitrate decisions accordingly, without requiring further online training. A comprehensive evaluation is conducted using diverse real-world network traces. Experimental results show that EAStream not only achieves leading performance on in-distribution test sets compared to state-of-the-art ABR algorithms, but also demonstrates superior generalization capability on out-of-distribution test scenarios.

Approximate nearest neighbor search (ANNS) is crucial in many applications to find semantically similar matches for user queries. Especially with the development of large language models (LLMs), ANNS is becoming increasingly important in retrieval-augmented generation (RAG). An in-depth analysis of ANNS reveals that its diverse operations, from extensive memory access to intensive sorting, are key performance bottlenecks, imposing significant strain on both the memory system and computing resources. Based on these observations, we present SeIM, a hierarchical in-memory architecture to accelerate ANNS. SeIM is designed to accommodate the diverse operational characteristics of ANNS. Specifically, SeIM offloads highly parallel memorybound operations to the memory bank level and introduces a unified execution model to reuse hardware units, requiring only lightweight modifications to standard DRAM architecture. Additionally, SeIM places compute-bound sorting operations, which require cross-unit data access, at the memory controller level and employs an adaptive transmission filtering technique to reduce unnecessary data transfers and processing during sorting. Our evaluation shows that SeIM achieves $268 \times 22 \times$, and $5 \times$ higher throughput, $306 \times 59 \times$, and $4 \times$ lower latency, and $3081 \times$, $287 \times$, and $2 \times$ higher power efficiency than state-of-the-art CPU-, GPU-, and ASIC-based ANNS solutions.

Hypergraph Neural Networks (HGNNs) are increasingly utilized to analyze complex inter-entity relationships. Traditional HGNN systems, based on a hyperedge-centric dataflow model, independently process aggregation tasks for hyperedges and vertices, leading to significant computational redundancy. This redundancy arises from recalculating shared information across different tasks. For the first time, we identify and harness implicit dataflows (i.e., dependencies) within HGNNs, introducing the microedge concept to effectively capture and reuse intricate shared information among aggregation tasks, thereby minimizing redundant computations. We have developed a new microedge-centric dataflow model that processes shared information as fine-grained microedge aggregation tasks. This dataflow model is supported by the Read-Process-Activate-Generate execution model, which aims to optimize parallelism among these tasks. Furthermore, our newly developed MeHyper, a microedge-centric HGNN accelerator, incorporates a decoupled pipeline for improved computational parallelism and a hierarchical feature management strategy to reduce off-chip memory accesses for large volumes of intermediate feature vectors generated. Our evaluation demonstrates that MeHyper substantially outperforms the leading CPUbased system PyG-CPU and the GPU-based system HyperGef, delivering performance improvements of $1,032.23 \times$ and $10.51 \times$, and energy efficiencies of $1,169.03 \times$ and $9.96 \times$, respectively.

By integrating external knowledge bases, Retrieval-augmented Generation (RAG) enhances natural language generation for knowledgeintensive scenarios and specialized domains, producing content that is both more informative and personalized.RAG systems typically consist of two fundamental stages: retrieval and generation.The retrieval stage experiences low bandwidth utilization due to its random and irregular memory access patterns.Meanwhile, the generation stage is also constrained by memory bandwidth limitations, which arise from involving a significant number of General Matrix-Vector Multiplications (GEMV) operations.These two stages collectively lead to memory bottlenecks within RAG systems.Recent efforts leverage HBM-based Processing-in-Memory (PIM) to accelerate conventional Large Language Models (LLMs).However, the retrieval stage incurs substantial storage overhead due to the need to maintain large-scale knowledge bases, resulting in a capacity bottleneck.Solely relying on HBM-based PIM in RAG is both costly and insufficient to meet the capacity demands.Fortunately, DIMM-based PIM provides a low-cost, high-capacity alternative that complements HBM.In this work, we propose HeterRAG, a novel heterogeneous PIM acceleration system for RAG.It combines

Vehicular edge computing (VEC) is an emerging computing paradigm that is rapidly advancing the development of the Internet of Vehicles (IoV). However, edge server has limited data storage capacity and computing resource, making it difficult to handle the massive offloading requests from IoV applications. Moreover, the mobility of vehicles and dynamic data traffic make it highly challenging to design optimal offloading and resource allocation strategies. To address the challenges mentioned above, we design a cloud-edge–vehicle hierarchical architecture for IoV task offloading, introducing a cloud server to assist in computation and alleviate the overload pressure on edge server. Considering the impact of vehicle mobility on task offloading, we propose a mobility detection method to predict which vehicles might leave the communication range of the base station, thereby preventing task offloading failures. Additionally, to achieve efficient task offloading and resource allocation in this complex IoV system, we propose a multiagent-reinforcement-learning-based vehicle proposal offloading algorithm (MVPOA). This algorithm enables vehicles to autonomously decide whether to process tasks locally or propose offloading to edge server. The edge server then decides whether to accept offloading requests based on task priority and sends rejected tasks to cloud server for processing, thereby maximizing the utilization of resources at each layer of the system. Simulation results demonstrate that MVPOA outperforms other baseline approaches in optimizing system delay and energy consumption.

Heterogeneous graph neural networks (HGNNs) have recently demonstrated significant advantages of capturing powerful structural and semantic information in heterogeneous graphs. Different from homogeneous graph neural networks directly aggregating information based on neighbors, HGNNs aggregate information based on complex metapaths. ReRAM-based processing-in-memory (PIM) architecture can reduce data movement and compute matrix-vector multiplication (MVM) in analog. It can be well used to accelerate HGNNs. However, the complex metapath-based aggregation of HGNNs makes it challenging to efficiently utilize the parallelism of ReRAM and vertices data reuse. To this end, we propose ChainPIM, the first ReRAM-based processing-in-memory accelerator for HGNNs featuring high-computing parallelism and vertices data reuse. Specifically, we introduce R-chain, which is based on a chain structure to build related metapath instances together. We can efficiently reuse vertices through R-chain and process different R-chains in parallel. Then, we further design an efficient storage format for storing R-chains, which reduces a lot of repeated vertices storage. Finally, a specialized ReRAM-based architecture is developed to pipeline different types of aggregations in HGNNs, fully exploiting the huge potential of multilevel parallelism in HGNNs. Our experiments show that ChainPIM achieves an average memory space reduction of 47.86% and performance improvement by$128.29\times $compared to NVIDIA Tesla V100 GPU.

Edge computing, as an emerging paradigm in distributed computing, introduces a novel data access framework for latency-sensitive applications, enabling data retrieval from edge servers situated closer to users, rather than the remote cloud. This significantly reduces data retrieval latency, thereby enhancing the quality of experience. However, the resources of edge servers are highly constrained. Recent studies in erasure coding-based edge data storage have effectively sealed the gap between storage cost and data retrieval latency. Despite these advantages, the highly dynamic edge computing environments and frequent data updates introduce significant challenges in cache replacement for erasure coding-based edge storage systems. Specifically, direct data replacement, similar to traditional replica-based storage methods, may result in insufficient encoded blocks for reconstructing the original data during retrieval, which can increase data retrieval latency and even compromise data availability. In this paper, we identify and address, for the first time, the cache replacement problem in erasure coding-based edge storage systems. We propose a novel cache replacement algorithm, named ECCR, based on Lyapunov optimization, which effectively solves the cache replacement problem in dynamic edge computing environments. Theoretical analysis and extensive experiments on real-world datasets demonstrate the effectiveness and efficiency of the proposed method, which outperforms two state-of-the-art approaches and achieves an average system cost reduction of 49.78%.

As the mainstream solution for semi-supervised learning (SSL), pseudo-labeling-based approaches have achieved re-markable success. However, an obvious drawback of existing methods is that the valuable semantic relationships among categories are often ignored, thus leading to suboptimal encoded embeddings. To address this, we present a novel Semantics-Enhanced Pseudo Labeling Network, called SEPL-Net, for image analysis in a semi-supervised manner. SEPL-Net explores the prior knowledge of visual similarity between different classes to improve the quality of pseudo label decision making. Particularly, we encode semantic labels combined with the one-hot label to jointly train our network by exploiting their disagreement. To alleviate the difficulty of labeling unlabeled images due to the introduction of semantic labels, we further design different classifiers with differentiated strong augmentation modes to enable cooperative pseudo labeling. Extensive experimental results show that our SEPL-Net outperforms existing SSL methods with the averaged 1.84% accuracy improvement on image classification task. Code is available at https://github.com/sweetvicky/SEPLNet.git.

Multi-subject video analysis is one of the most important problems in the field of visual perception for human activity recognition on multiple subjects nowadays. However, multi-subject video analysis is difficult to achieve real-time performance at the edge due to the limited resources of edge devices and the high complexity of the Convolutional Neural Networks (CNN) model used in this task. The common processing method is to upload the video data to the cloud. However, due to the influence of network bandwidth, the transmission time is not fixed, and the latency cannot be guaranteed. Thus, statically deployed model configurations cannot meet some dynamically changing scenarios. To address these challenges, in this paper, we propose an edge-cloud collaboration processing system for multi-subject video stream analysis, which can dynamically configure and optimize the related configurations according to specific scenarios. Specifically, we provide an adaptive configuration optimization solution based on context awareness for edge devices with limited resources such that multi-subject video stream analysis can be processed completely at the edge. For other complex scenarios, we propose an edge-cloud collaboration method to achieve task segmentation and collaboration to meet the performance requirements of the complex scenarios. Experimental results show that our method can achieve an average accuracy of 91.3% and the latency of less than 78ms with arbitrary runtime state.

With the widespread application of artificial intelligence and the Internet of Things, the intellectualization of the industrial Internet of Things (IIoT) has received more and more attention. However, in the application scenario with numerous sensors, the contradiction between massive requests of computing tasks and high requirements of inference quality affects the operation efficiency and service reliability. Moreover, due to the heterogeneity of computing resources and the randomness of communication environments of the cloud-edge system, how to compute and deploy deep learning models in a cloud-edge collaborative environment has also become a challenging problem. Therefore, this article presents a collaborative cloud-edge service cognitive framework for deep neural network (DNN) model service configuration to provide dynamic and flexible computing services. In order to adapt to different service requirements, we explored the tradeoffs between accuracy, latency, and energy consumption indicators, and a revenue target is established, which considers the quality of service experience and the system energy consumption to improve resource utilization efficiency. By transforming the optimization of the revenue target into a partially observable DNN configuration reinforcement learning problem, a dueling deep Q-learning network-based self-adaptive DNN configuration algorithm is proposed. Experimental results show that the proposed mechanism can effectively learn from external experience, adapt to the dynamic network environment, and reduce delay and energy consumption while meeting the service requirements.

With the rapid economic development, the prominent social competition has led to increasing psychological pressure of people felt from each aspect of life. Driven by the Internet of Things and artificial intelligence, intelligent psychological pressure detection systems based on deep learning and wearable devices have acquired some good results in practical application. However, existing studies argue that the psychological stress state is influenced by the current environment. They put much attention on the momentary features but ignore the dynamic change process of mental status in the time dimension. Besides, the lack of research in the general laws of psychological stress makes it difficult to quantitatively evaluate the stress status, resulting in the inability to perceive the stress state of users effectively. Thus, this article proposes an evaluation mechanism of psychological stress for adjusting the mental status of users. Specifically, we design a multi-dimensional feature space and a time-aware feature encoder, which integrate various stress features and capture time characteristics of stress state change. Moreover, a novel mental state model is proposed, which uses the pressure features with time characteristics to evaluate the pressure stress level. This model also quantifies the internal relationship between pressure features. Last, we establish a practicable testbed to demonstrate how to evaluate and adjust mental state of users by the proposed evaluation mechanism of psychological stress.

The traditional natural gas Internet-of-Things (IoT) system has many problems, such as centralized management of resources, noncirculation of data between stations, insecurity of transaction information or account books, and lack of contract consensus. In order to ensure data security and reliable transaction, this article introduces artificial intelligence (AI) and blockchain technology and constructs an AI-enabled and blockchain-powered natural gas IoT system in a smart city. In this article, the natural gas output prediction model based on temporal pattern attention-based LSTMs (TPA-LSTMs) is used to enable the system to sense the change of natural gas deliverability. In addition, we establish a blockchain-based secure natural gas transaction scheme, which dynamically matches the purchase contract and sale contract to maximize the interests of the buyer and the seller and obtain a transaction contract. The experimental results show that our model can predict the output value of natural gas in real time and select the appropriate transaction matching scheme according to the dynamic demand for sales.

Autism spectrum disorder (ASD) is a serious hazard to the physical and mental health of children, which limits the social activities of patients throughout their lives and places a heavy burden on families and society. The developments of communication techniques and artificial intelligence (AI) have provided new potential methods for the treatment of autism. The existing treatment systems based on AI for children with ASD focus on detecting health status and developing social skills. However, the contradiction between the terminal interaction capability and availability cannot meet the needs for real application scenarios. At the same time, the lack of diverse data cannot provide individualized care for autistic children. To explore this robot-based approach, a novel AI-based first-view-robot architecture is proposed in this article. By providing care from the first-person perspective, the proposed wearable robot overcomes the difficulty of the absence of cognitive ability in the third-view of traditional robotics and improves the social interaction ability of children with ASD. The first-view-robot architecture meets the requirements of dynamic, individualized, and highly immersed interaction services for autistic children. First, the multi-modal and multi-scene data collection processes of standard, static, and dynamic datasets are introduced in detail. Then, to comprehensively evaluate the learning ability of children with ASD through mental states and external performances, a learning assessment model with emotion correction is proposed. Besides, a wearable robot-assisted environment perception and expression enhancement mechanism for children with ASD is realized by reinforcement learning, which can be adapted to interactive environments with optimal action policies. An interactive testbed for children with ASD treatments is demonstrated and experimental cases for test subjects are presented. Last, three open issues are discussed from data processing, robot designing, and service responding perspectives.

With the advent of the Internet era, the trend of highly informed society has been becoming more and more obvious, and the requirement of society on communication is also increasing. flexible satellite communication mode has many advantages such as large communication load and no geographic restriction, which cannot be replaced by other communication modes. In the satellite communication system, the most important is the satellite earth station (SES). When receiving signals from the target satellite, the SES terminal must accurately point to the satellite and track it to obtain the maximum receiving signal and reduce the interference with other signals simultaneously. However, the motion of either satellite or terminal can cause a change in signal intensity, so it is necessary to adjust the pointing of the SES antenna in time to maintain optimal signal receiving conditions. In order to satisfy different satellite communication scenarios, in this paper, Artificial intelligent (AI) technology is applied to the satellite communication process, mainly to optimize the optimal antenna angle and time consumption reduction. Firstly, the process of antenna pointing is introduced, and the traditional antenna search algorithm Auto-Acqire algorithm (AA algorithm) is analyzed in detail. Considering that the satellite system needs to adapt to the communication requirements of different terminals, based on AI antenna pointing algorithms are proposed. In order to verify this research, we build an experimental platform and compare the traditional AA algorithm as a benchmark algorithm with FI-GRU and II-DRL algorithms. According to the experimental results, the two algorithms proposed in this paper can improve the efficiency of satellite pointing and tracking tasks.

As power level grows, the power converter of wind turbines tends to use parallel-connected smaller converters. The main reason for parallel is that modular design of those smaller converters is relatively mature and cost-effective. However, the parallel of GSCs (Grid-Side Converters) is notorious for its difficulty in control. Bulky and costly filtering capacitor bank is common and LCL resonance problem often occurs in wind farm. Due to communication speed, synchronization and engineering limitation, conventional methods cannot obtain a good balance between the control performance and the system reliability. This paper first proposes to utilize the real-time Ethernet technology for the control of parallel-connected GSCs of a wind turbine. With the proposed technology, the conventional control deficiencies of parallel-connected GSCs can be satisfactorily overcome. The reliable real-time Ethernet single line connection greatly simplifies the system interconnection and provides a unified information interface for all control devices. The high-speed communication and real-time response capability make it possible to perform fast closed-loop control in one master controller. Moreover, as all control information is collected in the master controller, the best coordination strategy can be applied to improve the system performance. The high accuracy of synchronization enables slave controllers to synchronize at carrier frequency level so that they can use phase-shifted carrier to cancel the switching ripple of branch currents. The HIL (Hardware-in-The-Loop) experimental results verify the feasibility and high performance of the proposed control system. The application of real-time Ethernet is promising to become a new trend in wind power converter design.