Serverless computing has emerged as a prominent research focus in cloud computing because it provides infrastructure-transparent development and elastic resource management. However, this computing paradigm still faces the inherent challenge of cold start. Existing approaches have two major limitations: insufficient workload prediction accuracy and inefficient allocation of reusable container replicas to incoming function requests. To address these challenges, we propose a container scheduling approach based on Workload Prediction and Particle Swarm Optimization (PSO), named WPPSO. WPPSO first leverages a code-pre-trained large language model (LLM) to extract intrinsic function features and then uses a spatio-temporal fusion-based temporal neural network (STF-TNN) to predict serverless workloads. It subsequently employs a greedy algorithm to construct a high-quality initial matching state and uses PSO to refine the container scheduling strategy. Finally, WPPSO introduces a hierarchical container recycling mechanism to reduce idle resource waste. Extensive experiments show that WPPSO reduces startup latency by up to 72.2% and memory footprint by 63.4% compared with the native Knative platform. Compared with RainbowCake, WPPSO achieves a 15.6% lower mean startup latency without statistical significance and a statistically significant 31% reduction in idle memory consumption.

In resource-constrained edge intelligence environments, Neural Processing Units (NPUs) have become the mainstream accelerators for Deep Learning Inference (DLI) tasks. However, the absence of fine-grained preemption in existing NPUs limits their ability to meet the strict real-time requirements of concurrent DLI workloads. To sustain performance under high concurrent DLI workloads, edge intelligence devices often overprovision NPUs, thereby incurring inefficient resource utilization and increased system cost. This paper presents a Multi-model Aware Real-time Scheduler designed for coordinated DLI execution on NPUs (MARS) without altering the model structure or parameters. MARS formulates the scheduling challenge as a blocking-aware job chain real-time optimization problem over a hyperperiod for single-NPU heterogeneous devices, and addresses it through a three-phase system. In the offline phase, a Utility-driven Partition Scheme Selection (UPSS) algorithm determines efficient model partitioning strategies to reduce the scheduling search space. During the online phase, the Simulated Annealing-based Earliest Deadline First for Heterogeneous Job Chain Real-Time Scheduling (SAEDF-HCRS) algorithm dynamically optimizes task scheduling decisions to maximize real-time task capacity and ensure deterministic execution. Finally, the pipeline execution phase applies multidimensional optimization to manage heterogeneous job chains, while maintaining responsiveness. Experiments show that MARS maintains millisecond-level response latency for concurrent tasks, while achieving up to 17

In edge computing (EC), services are deployed on edge nodes for lower latency and higher privacy. However, the resource-constrained characteristics of edge servers make services deployment hard to satisfy the quality of service (QoS) requirement. Recently, some research considers deep reinforcement learning (DRL) in service deployment especially for EC. However, these DRL-based service deployment methods ignore the optimization of initial state of DRL and have a low information utilization, which lead to a poor QoS of services. Therefore, we propose a service deployment method, named SDAD, in EC for that integrates association rules and DRL to optimize the initial state and fully utilize information of services and edge nodes. First, we mine association rules about services, edge nodes and QoS from history data. Then, we apply the mined rules to generate a helpful initial service deployment state and to form feature vectors to augment information for DRL. The service deployment is finally evolved through DRL. The experimental results show that our SDAD has a 6.4

This study proposes a multidimensional fault injection test methodology targeting ARINC429 avionics bus. ARINC429 avionics bus is currently a mature aviation bus standard used in many important onboard systems. Evaluating the reliability and safety of the ARINC429 avionics bus in critical system communication is particularly important. Completing fault injection test and evaluation of ARINC429 signals through injecting specific faults of ARINC429 avionics bus signals into the physical layer, electrical layer, and protocol layer, and completing fault injection test in different scenarios through signal channel routing. This article discusses the principle and architecture of fault injection test for the widely used ARINC429 signal in avionics systems in multi-level and multi scenarios. Based on this architecture, a test platform is built to complete the verification of ARINC429 avionics bus fault injection test. And taking the navigation system as an example, relying on this platform to complete fault injection test.

This paper introduces a novel method for matching visible light and infrared images, termed the Visible Light and Infrared Image Matching Network (VLIMNet). In the image encoding stage, we incorporate a generative architecture-based modality transformation network after the SuperPoint encoder, enabling the local geometric features extracted from infrared images to more closely resemble those of visible light images. This reduces the impact of modality differences. Simultaneously, we utilize the image encoder of the Segment Anything Model to obtain global semantic descriptors. During the decoding stage, we fuse the global semantic descriptors with the local geometric descriptors and perform joint decoding to obtain keypoints and new feature vectors. The matching process is computed by solving a differentiable optimal transport problem based on the LightGlue network, a graph neural network with an attention mechanism. Compared to other matching models, our approach demonstrates improvements across various metrics in the domain of infrared and visible light image matching, particularly excelling in matching images with significant pose differences. Specifically, our method achieves approximately a 5% improvement in matching accuracy compared to the highest accuracy matching methods.

Broken oracle bones record valuable information about ancient Chinese cultural heritage in the Shang Dynasty. However, many complete oracle bones have broken into pieces over the years, resulting in disjoining fragments scattered around the world, researchers have turned their attention to current technology to piece together oracle bone fragment images. A survey on rejoining object fragment algorithms and experiment datasets has been conducted, and it shows that the restoration work poses a serious challenge for current restoration models. To enhance the effectiveness of a deep rejoining model (DRM) with two stages, in the first stage, the longest similar edge segment (LSES) is proposed to match edge segments of two images, and a complete image rejoining (CIR) algorithm is given to rejoin the two images to be a complete image, in the second stage, we have embedded machine learning or deep learning methods into the DRM to determine whether the two images’ textures are continuity. In addition, a data set of oracle bone fragment image (OBFI) is given to evaluate current methods, it consists of three parts, and they are single high-resolution oracle bone fragment images, oracle bone fragment image pairs that can be rejoined together, and images that have rejoinable or unrejoinable textures. Extensive experiments on DRM and OBFI datasets demonstrate that DRM over previous state-of-the-art methods and the dataset can be used to evaluate the performance fragment rejoining algorithms.

Person re-identification (Re-ID) systems in edgecloud continuum face critical trade-offs between latency sensitivity and energy efficiency due to the resource-constrained edge environment. This paper proposes Re-IDFaaS, a serverless ReID system that dynamically optimizes energy consumption and computational performance across the edge-cloud continuum. Leveraging serverless architectures, our system implements three key improvements: (1) An event-driven workflow triggered by motion detection, eliminating idle GPU resource consumption during inactive periods. (2) A hardware-aware dynamic scheduler that allocates tasks based on real-time energy states and container availability, achieving balanced resource utilization across heterogeneous nodes. (3) An adaptive batching mechanism that reduces cold-start frequency through latency-constrained request grouping while maintaining the efficiency of GPU memory. Experiments demonstrate a 23.3% improvement in edge node availability and 55% reduction in memory usage compared to existing methods. The system design provides practical insights for building AI services in hybrid computing environments requiring cross-framework compatibility and adaptive resource orchestration, achieving 53% higher throughput than traditional architectures. These innovations address the challenges of dynamic workload scheduling and runtime optimization in hardware-diverse scenarios, ensuring sustainable operation under bursty surveillance workloads.

In recent years, gesture recognition based on data gloves has attracted increasing attention as a human-computer interaction (HCI) method that is natural, convenient, stable, robust, easy to recognize, and applicable to various usage environments. This research first proposes an advanced smart data glove that integrates cutting-edge flexible capacitive sensors on the fingertips and a 6-axis IMU on the back of the hand to recognize gestures. Secondly, this study proposes a personalized continuous gesture segmentation (PCGS) model that can adaptively calculate the most appropriate gesture segmenting threshold based on the current user and introduces the multi-sliding window theory and kinematic knowledge to perform personalized gesture segmentation. The accuracy of gesture segmentation can reach 94.3%. The result shows that our PCGS model achieves an average segmentation accuracy of 94.3% and outperforms the state-of-the-art pproaches by 11.2% to 18.5%.

Software-based distributed shared memory (DSM) allows multiple processes to access shared data without the need for specialized hardware. However, this flexibility comes at a significant cost due to the need for data synchronization. One approach to mitigate these costs is to relax the consistency model, which can lead to delayed updates to the shared data. This approach typically requires the use of explicit synchronization primitives to regulate access to the shared memory and determine the timing of data synchronization. To circumvent the need for explicit synchronization, an alternative approach is to manage shared memory transparently using the underlying system. While this can simplify programming, it often imposes a fixed granularity for data sharing, which can limit the expansion of the coherence domain and increase the synchronization requirements. To overcome this limitation, we propose an abstraction called the elastic coherence domain, which dynamically adjusts the scope of data synchronization and is supported by the underlying system for transparent management of shared memory. The experimental results show that this approach can improve the efficiency of memory sharing in distributed environments

KPIs (key performance indicators) reflect the operational state of Cloud services and infrastructure. With the rapid development of cloud services, many dimensional KPIs are constantly generated during service operations. Recently, deep-learning-based Multivariate KPI Anomaly Detection (MAD) methods have become the mainstream trend to meet the requirements of effective anomaly detection. As the number of KPI dimensions grows, the redundant dimensions in anomaly detection increase. Data with high dimensional complexity increases computational costs and storage requirements. This phenomenon is called the "Curse of dimensionality", which prevents common MAD strategies from being efficient. Therefore, dimensionality reduction is important in MAD. However, there are few researches on dimensionality reduction methods specifically designed for MAD. In this paper, we propose a feature selection KPI dimensionality reduction framework ITRMD, which reduces multivariate KPI series from two aspects: inter-metric dimensionality selection and temporal dimensionality selection. ITRMD preserves features of original data by selecting KPI dimensions with effective information and discards redundant KPI dimensions. Experiments on real KPI datasets show that ITRMD can greatly improve the detection accuracy of the state-of-the-art MAD and reduce the computation cost of model training.

Stream partitioning has a fundamental effect on the efficiency of data parallelism in distributed stream processing systems. The skewed and time-varying nature of streaming data makes it challenging to achieve load balancing while minimizing the cost incurred. The requirement of adaptivity further complicates the problem, that the partitioning mechanism should not only be able to capture the changes in workload and adjust itself but also be quite tolerant of the changes because of the lag in statistics. Existing approaches use one-choice or multiple-choice schemes to make tradeoffs between these factors, but they tend to treat them as opposites, which either fails to achieve good load balancing or incurs excessive cost. There is a lack of deeper insight into how partitioning behavior affects load balancing, cost, and adaptivity when the keys have a different number of candidate choices. Also, it requires a flexible partitioning scheme to allow different trade-offs among the three factors for various scenarios. To address the issues mentioned above, we propose a novel (1 +beta)-choice based stream partitioning scheme, which splits beta is an element of(0, 1) part of keys selectively to have multiple candidate choices. We demonstrate that just splitting beta part of the keys is sufficient to achieve optimal load balancing while minimizing cost and providing the required adaptivity to workload variance. In a new perspective, we analyze the relationship among load balancing, cost, and adaptivity, as the theoretical foundation of getting proper beta and the corresponding number of choices. Experiments on Apache Flink demonstrate that our approach outperforms state-of-the-art solutions, improving throughput by 7.3x and reducing latency by 85%.

The field of wearable sensing devices and technology has witnessed remarkable advancements, driven by the need for more integrated, efficient, and user-friendly systems [...]

As the editor of the Special Issue on “Exploring IoT Sensors and Their Applications: Advancements, Challenges, and Opportunities in Smart Environments”, I am delighted to present this collection of groundbreaking research that addresses the emerging needs and challenges in the field of IoT sensors and smart environments [...]

Summary Convolutional neural networks (CNNs) have shown remarkable advantages in a wide range of domains at the expense of huge parameters and computations. Modern CNNs still tend to be more complex and larger to achieve better inference accuracy. However, the complex and large structures of CNNs could slow down the inference speed. Recently, Compressing the convolutional weights to be sparse by pruning the unimportant parameters has been demonstrated as an efficient way to reduce the computations of CNNs. On the other hand, field‐programmable gate arrays (FPGAs) have been a popular hardware platform to accelerate CNN inference. In this paper, we propose an algorithm/hardware co‐optimized method for accelerating CNN inference on FPGAs. For the algorithm, we take advantage of unstructured and structured parameter sparsifying methods to achieve high sparsity and keep the regularity of convolutional weights. Correspondingly, hardware‐friendly index representations of sparse convolutional weights are proposed. For the hardware architecture, we propose row‐wise input‐stationary dataflow, which is tightly coupled with the algorithm. A row‐wise computing engine (RConv Engine) is proposed, which is based on the dataflow. Inside the RConv Engine, the scalar‐vector structure is applied to implement the basic processing elements (PEs). To flexibly calculate the feature map with various sizes, the PEs are organized in a 2D structure with two work modes. The experimental results demonstrate that our co‐optimized method implements high sparsity of convolutional weights, and the computing engine achieves high computation efficiency. Compared with other accelerators, our co‐optimized method implements a 10.9 speedup on FPS at most with the highest sparsity of convolutional weights and negligible accuracy loss.

Stream join is a fundamental operation in stream processing and has attracted extensive research due to its large resource consumption and serious impact on system performance. As the theoretical basis of stream join systems, the stream join model greatly affects system performance. State-of-the-art stream join models either consume too much computing resources or too much storage resources, thus resulting in lower throughput or higher latency. In this paper, we propose a new stream join model for processing arbitrary join predicates, called CoModel, which offers a flexible trade-off between memory and computing resource consumption. More importantly, CoModel can achieve the minimum sum of the number of store operations and join operations among all existing join models, and thus can achieve the lowest latency and highest throughput when the overheads associated with the local stream join for each input tuple are approximately constant. We give a trade-off strategy for CoModel and theoretically prove its performance advantages based on queuing theory. Furthermore, we design and implement an adaptive distributed stream join system, CoStream, based on CoModel. CoStream can adaptively adjust its structure according to resource constraints and statistics of input data. We conduct extensive experiments for CoStream to evaluate its performance and adaptivity, and the results show that CoStream has the lowest latency and highest throughput in various scenarios.

Loop-closure detection, also known as position recognition, is a function in the "Simultaneous Localization And Mapping（SLAM） system to judge whether the motion track passes through repeated places（i.e. there is loop-closure） according to the similarity between images, and plays a role in eliminating cumulative errors in stages. Focused on the research of loop-closure detection under the specific theme of visual SLAM system, we summarize the basic functions and basic components of SLAM system, analyze the principle and workflow of loop-closure detection, pre-problems and evaluation indicators. We analyze a series of methods generated in the development of loop-closure detection, classify two types of loop-closure detection algorithms in SLAM system, which are loop-closure detection algorithm based on bag-of-words model and loop-closure detection algorithm based on deep learning, and focus on the principle, advantages and disadvantages of these two types of algorithms. The analysis shows that the loop-closure detection algorithm based on bag-of-words model is still in the mainstream because of its real-time advantage. The loop-closure detection algorithm based on deep learning has excellent accuracy and robustness, but limited by the allocation of computing resources by devices, how to apply this kind of method to the visual SLAM system that pays attention to real-time is still an urgent problem to be solved. Finally, the challenges and problems of loop-closure detection are analyzed and prospected.

For a stream processing system that uses checkpoints as a fault-tolerant method, selecting the appropriate checkpoint period is the key to ensuring the efficient operation of streaming applications. State-of-art stream processing systems currently only support fixed-cycle checkpoints, which is difficult to make a good trade-off between fault-tolerant processing and the cost of failure recovery in dynamically changing streaming application scenarios. Moreover, in a complex distributed streaming application environment, the dynamic environmental indicators (e.g., the values of workloads and failure rates) are not in coincidence with the model assumptions, such as the dynamics of Twitter’s hot events data changing quickly. In this paper, we consider the dynamic changes of environmental indicators and adaptively optimize the processing delay and fault recovery time. Then, we propose a dynamic adjustment method for the checkpoint interval by reinforcement learning, which is named DACM. DACM adaptively optimizes the processing delay and fault recovery time, while avoiding the overall environment modeling of streaming applications. The experiments conducted on the Flink platform show that DACM reduces the processing delay by 10% and the failure recovery time by 37% compared with the existing checkpoint interval optimization models.

This study has designed and developed a smart data glove based on five-channel flexible capacitive stretch sensors and a six-axis inertial measurement unit (IMU) to recognize 25 static hand gestures and ten dynamic hand gestures for amphibious communication. The five-channel flexible capacitive sensors are fabricated on a glove to capture finger motion data in order to recognize static hand gestures and integrated with six-axis IMU data to recognize dynamic gestures. This study also proposes a novel amphibious hierarchical gesture recognition (AHGR) model. This model can adaptively switch between large complex and lightweight gesture recognition models based on environmental changes to ensure gesture recognition accuracy and effectiveness. The large complex model is based on the proposed SqueezeNet-BiLSTM algorithm, specially designed for the land environment, which will use all the sensory data captured from the smart data glove to recognize dynamic gestures, achieving a recognition accuracy of 98.21%. The lightweight stochastic singular value decomposition (SVD)-optimized spectral clustering gesture recognition algorithm for underwater environments that will perform direct inference on the glove-end side can reach an accuracy of 98.35%. This study also proposes a domain separation network (DSN)-based gesture recognition transfer model that ensures a 94% recognition accuracy for new users and new glove devices.

Quality of Service (QoS) prediction is a crucial aspect in service management. However, the existing QoS prediction methods face several limitations, such as loss of information during encoding, incomplete feature extraction and neglect of the interaction between features. To this end, this paper proposes a new QoS PRediction method based on a Multi-Scale convolutional neural Network, i.e., QPRMSN. For each service invocation, we build a feature matrix that encodes invocation context and QoS characteristics by using status codes with degrees of membership. Then, a multi-scale convolutional neural network is employed to extract features that keep detailed information during deep global features mining. Moreover, we introduce attention mechanism to learn the intrinsic relationships between features to strengthen key features. Finally, QPRMSN completes the QoS prediction based on a multi-level feature matrix. Extensive experiments are conducted on a real-world dataset to evaluate the performance of QPRMSN. The experimental results demonstrate that QPRMSN outperforms the state-of-the-art QoS prediction models and is better at QoS context encoding.

Convolution Neural Network (CNN) models have shown their dominance in computer vision tasks. Recently, a special convolution block, named MBConv block or inverted residual block, is proposed to construct CNNs to meet the real-time requirements on resource-constrained edge-computing platforms. The MBConv block is first proposed by MobileNetV2 and has been widely used to construct lightweight CNNs. However, the MBConv block brings new challenges to the structure of the computing engine, the bandwidth requirement of off-chip memory and the demand for on-chip memory when designing hardware accelerators. In this paper, a convolution Block Oriented Accelerator (BOA) is proposed for the inference of CNNs constructed on MBConv blocks. In BOA, the MBConv-based CNNs are performed block by block using a Block-Based Engine which consists of dedicated computing units for each layer of the MBConv block. To reduce both the bandwidth requirement of off-chip memory and the demand for on-chip memory, a two-level data flow optimization and an amortized weight loading method are proposed. Furthermore, a hierarchical scheduling scheme is proposed to improve the performance and flexibility so that BOA can guarantee all units running in parallel and support various MBConv-based CNNs. Finally, we deploy BOA on Xilinx VC709. We evaluate the accelerator on ImageNet for image classification. The results show that BOA can perform various MBConv-based CNNs and achieve 1.28x - 7.75x speedup on inference latency.