Efficient Deployment Research Articles

Purpose This study aims to address the key challenges in multi-unmanned aerial vehicle (UAV) data sensing systems, where energy-constrained UAVs require real-time trajectory optimization to simultaneously maximize coverage efficiency and minimize energy consumption. Traditional centralized optimization approaches struggle with scalability and computational complexity due to the non-convexity and high dimensionality of the joint optimization problem. To overcome these limitations, the authors propose a distributed multi-agent deep reinforcement learning (MADRL) framework that leverages the autonomous decision-making capability of deep reinforcement learning to achieve distributed continuous action control of UAVs, thereby enabling efficient autonomous deployment. Design/methodology/approach This study proposes a multi-UAV energy-efficient data collection scheme (multi-UAV E2DC) based on an MADRL algorithm. The proposed approach enables UAVs to dynamically collect data from multiple ground sensors while accounting for practical constraints such as communication range, motion limits and energy consumption. To achieve this, the authors first construct a multi-objective optimization model by integrating an air-to-ground communication model with a UAV energy consumption model. Building on this foundation, the authors further develop an enhanced MADRL algorithm within a centralized training and decentralized execution (CTDE) actor-critic framework, which supports efficient continuous trajectory control and deployment of multi-UAV. Findings Extensive simulations demonstrate that the proposed approach achieves superior performance in multi-objective optimization compared to benchmark methods, including random policy, K-means clustering and multi-agent deep deterministic policy gradient. Specifically, the proposed method outperforms in terms of average coverage density, data volume and coverage energy efficiency index. Originality/value This study proposes an MADRL-based energy-efficient data collection framework for UAVs, which integrates air-to-ground communication and UAV energy consumption models to formulate a multi-objective optimization problem. By adopting a CTDE framework for continuous trajectory control, it effectively overcomes the computational challenges of traditional non-convex optimization methods in complex environments. The proposed approach offers a theoretically sound and practically applicable solution for distributed UAV sensing in extreme disaster scenarios.

Acute atherosclerotic tandem occlusions of the internal carotid artery (ICA) present significant technical challenges for endovascular intervention. Optimal strategies for the efficient and safe deployment of embolic protection devices (EPDs) in this context remain to be established, as conventional approaches often necessitate multiple device exchanges, prolonging procedural time and potentially increasing risk. In this retrospective series, 16 consecutive patients with acute ischemic stroke due to ICA occlusion underwent endovascular therapy utilizing a microcatheter-facilitated EPD (MICRO-EPD) delivery technique. This novel approach, leveraging a 0.027-inch microcatheter for direct EPD deployment, eliminated the need for standard exchange maneuvers and reduced procedural complexity. Technical success was achieved in all cases, with a median puncture-to-recanalization time of 55 minutes. At 90 days, 56.3% of patients achieved favorable functional outcomes (modified Rankin Scale score ≤ 2). Symptomatic intracranial hemorrhage and mortality each occurred in 1 patient (6.3%). No device-related technical failures or unanticipated complications were observed. The MICRO-EPD technique demonstrates a promising advance in the management of acute atherosclerotic tandem occlusions, offering a simplified, safe, and effective method for EPD deployment. This strategy has the potential to optimize procedural efficiency and improve patient outcomes, meriting further prospective evaluation in broader clinical settings. https://thejns.org/doi/10.3171/CASE25622.

The increasing demand for liquefied hydrogen (LHS) as a clean energy carrier necessitates the development of structural materials capable of withstanding extreme cryogenic conditions. This paper reviews three primary candidate materials austenitic stainless steels, high nickel steels, and high-manganese steels focusing on their mechanical properties, fracture toughness, and susceptibility to hydrogen embrittlement. Austenitic stainless steels provide excellent toughness and hydrogen resistance but are limited by low strength and high cost. Conventional 9%Ni steels have demonstrated reliable performance in LNG service, while newly developed 13-15%Ni steels exhibit stable fracture toughness down to lower temperature, extending their applicability to LH2 environments. High-manganese steels offer an attractive balance of high toughness, strength, and cost effectiveness; however, further validation of hydrogen embrittlement resistance and international standardization are required before large-scale implementation. The comparative evaluation highlights that each alloy system presents distinct advantages and limitations, and that future efforts should focus on comprehensive cryogenic testing and the establishment of consistent qualification standards to ensure safe and efficient deployment of LH2 storage and transportation systems.

This research focuses on the development of an interactive web-based dashboard to support a precision agriculture system for chili plants. The primary focus of this research is on the back-end development of the system. The system integrates several internal and external APIs, including the Flask API (internal) for plant disease classification and growth prediction, and the Google Gemini API for the AI-powered chatbot that provides consultation to farmers (external). These features allow farmers to receive automatic disease diagnosis and growth predictions, improving decision-making and crop management. The dashboard also presents weather information, environmental data, and nanobubble data, along with Echarts gauge charts for seven essential metrics: Electrical Conductivity (EC), temperature, humidity, pH, nitrogen, phosphorus, and potassium. Data for the environmental and nanobubble data is retrieved from the ThingSpeak API (external), while weather information is fetched from the OpenWeatherMap API (external). The system was thoroughly tested using Postman to ensure all API endpoints function correctly. The results confirmed that all endpoints responded with status code 200 OK, indicating stable back-end performance. Performance testing showed response times stabilizing at 2000 ms after initial 4500 ms peaks, confirming efficient handling, reliable endpoints, and deployment readiness.

Synthetic Aperture Radar (SAR) onboard satellites provides high-resolution Earth imaging independent of weather conditions. SAR data are acquired by an aircraft or satellite and sent to a ground station to be processed. However, for novel applications requiring real-time analysis and decisions, onboard processing is necessary to escape the limited downlink bandwidth and latency. One such application is real-time target recognition, which has emerged as a decisive operation in areas such as defense and surveillance. In recent years, deep learning models have improved the accuracy of target recognition algorithms. However, these are based on optical image processing and are computation and memory expensive, which requires not only processing the SAR pulse data but also optimized models and architectures for efficient deployment in onboard computers. This paper presents a fast and accurate target recognition system directly on raw SAR data using a neural network model. This network receives and processes SAR echo data for fast processing, alleviating the computationally expensive DSP image generation algorithms such as Backprojection and RangeDoppler. Thus, this allows the use of simpler and faster models, while maintaining accuracy. The system was designed, optimized, and tested on low-cost embedded devices with low size, weight, and energy requirements (Khadas VIM3 and Raspberry Pi 5). Results demonstrate that the proposed solution achieves a target classification accuracy for the MSTAR dataset close to 100% in less than 1.5 ms and 5.5 W of power.

Understanding the brain requires modeling large-scale neural dynamics, where coarse-grained modeling of macroscopic brain behaviors is a powerful paradigm for linking brain structure to function with empirical data. However, the model inversion process remains computationally intensive and time-consuming, limiting research efficiency and medical deployment. In this work, we present a pipeline bridging coarse-grained brain modeling and advanced computing architectures. We introduce a dynamics-aware quantization framework that enables accurate low-precision simulation with maintained dynamical characteristics, thereby addressing the precision challenges inherent in the brain-inspired computing architecture. Furthermore, to exploit hardware capabilities, we develop hierarchical parallelism mapping strategies tailored for brain-inspired computing chips and GPUs. Experimental results demonstrate that the deployed low-precision models maintain high functional fidelity while achieving tens to hundreds-fold acceleration over commonly used CPUs. This work provides essential computational infrastructures for modeling macroscopic brain dynamics and extends the application of brain-inspired computing to scientific computing in neuroscience and medicine.

Abstract. Airborne wind energy systems (AWESs) offer a promising route to high-altitude wind harvesting, but their commercialization remains limited by the challenge of converting highly dynamic tethered flight power into stable electrical energy. While most research has focused on flight trajectories and control, the mechanical-to-electrical conversion stage requires further experimental validation. This paper introduces a validated electrical test bench emulator and a torque-ripple-optimized model predictive control (MPC) strategy, evaluated using two real AWES flight datasets. The emulator reproduces variable tether forces and reeling dynamics under optimal figure-eight crosswind flight. Two DC-bus topologies are compared: a separated bus that accurately mimics AWES storage dynamics (≈ 98 % fidelity) but demands 45 %–55 % more battery capacity and a common bus that recirculates energy, reducing storage needs by two-thirds. When realistic storage dynamic emulation is required, the separated-bus configuration is the only suitable option. The proposed MPC ensures precise generator speed and torque regulation, achieving torque-tracking root mean squared errors (RMSEs) below 0.11 % (Dataset 1) and 0.14 % (Dataset 2) and speed-tracking RMSEs of 0.44 % and 0.82 %, respectively. Overall energy efficiencies reach 82 % with Dataset 1 and 60 % with Dataset 2, with peak instantaneous efficiencies of 93 % and 88 %. Permanent magnet synchronous generators (PMSGs) outperform induction machines (IMs) by 4 % in Dataset 1 and up to 20 % in Dataset 2, with instantaneous gains of 2 %–10 % at high power. Off-nominal operation degrades cycle efficiency and drives higher battery cycling even in a common-bus setup, highlighting the importance of correct machine dimensioning. However, when storage dynamics are not under study, the common-bus configuration is the most cost-effective option, requiring less hardware and imposing lower peak discharge stresses. These results establish electrical test bench emulators as essential platforms for systematic evaluation and optimization of AWES power conversion, informing both machine design and control strategies for scalable, efficient AWES deployment.

Cyclic alternating patterns (CAP) of sleep can be observed through electroencephalogram (EEG) signals. Analyzing CAP can provide valuable insights into different abnormalities relating to sleep. CAP comprises of two phases: A and B, characterized by the brain response to different types of stimuli. In this study, we propose an efficient and accurate system to segregate the CAP phases by considering the EEG signals from two different categories, i.e., healthy individuals and those suffering from insomnia. The input signal is divided into short sequences, which are analyzed using Gaussian filters to generate the frequency band (FB) components. Forward ternary encoding (FTE) is applied to each of the FB components and the encoded values are represented using histograms to capture the intrinsic signal patterns. The feature vector is constructed by combining the histograms obtained from all FB components, while the feature selection is achieved using the Kruskal-Wallis test. We evaluate the performance of four different machine learning classifiers and compare their results. The bagged tree (BT) classifier yields accuracy of 80.16% and 81.12% for the healthy and insomnia datasets, respectively. The proposed method performs better than the existing studies on CAP classification for two different datasets. It is accurate and easy to implement, and hence, it holds promise for efficient real-time deployment.

As the complexity of convolutional neural networks (CNN) continues to increase, efficient deployment on computationally constrained hardware platforms has become a significant challenge. Against this backdrop, field-programmable gate arrays (FPGA) emerge as an up-and-coming CNN acceleration platform due to their inherent energy efficiency, reconfigurability, and parallel processing capabilities. This paper establishes a systematic analytical framework to explore CNN optimization strategies on FPGA from both algorithmic and hardware perspectives. It emphasizes co-design methodologies between algorithms and hardware, extending these concepts to other embedded system applications. Furthermore, the paper summarizes current performance evaluation frameworks to assess the effectiveness of acceleration schemes comprehensively. Finally, building upon existing work, it identifies key challenges in this field and outlines future research directions.

Fourier transform infrared spectroscopy enables rapid, nondestructive identification of mixture components through characteristic absorption peaks. However, in practical applications, challenges such as instrument line shape variations, overlapping absorption peaks, and various measurement errors significantly complicate the identification of mixtures. To address this, we developed an innovative deep learning framework based on an attention mechanism. Extensive experiments were conducted on a self-constructed data set comprising ten distinct instrument line shapes and eight gas components. Remarkably, it attained exact match ratios exceeding 91.7% when applied to the other nine instrument line shapes, outperforming existing methods by margins ranging from 25% to 88%. These findings demonstrate the model's robust generalization capability and efficient deployment flexibility, while more importantly highlighting its significant potential for cross-device applications, other FTIR mixture analyses, and similar spectroscopic challenges, such as transfer function in near-infrared spectroscopy.

Edge computing environments face unprecedented challenges in deploying large language models due to severe resource constraints, latency requirements, and privacy concerns that traditional cloud-based solutions cannot address. Current approaches struggle with the fundamental mismatch between LLMs' computational demands - requiring gigabytes of memory and billions of operations - and edge devices' limited capabilities, resulting in either degraded performance or infeasible deployments. This survey presents a systematic analysis of emerging techniques that enable efficient LLM deployment at the edge through four complementary strategies: model compression via quantisation and pruning that reduces memory footprint by up to 75% while maintaining accuracy, knowledge distillation frameworks achieving 4000× parameter reduction with comparable performance, edge-cloud collaborative architectures like EdgeShard delivering 50% latency reduction through intelligent workload distribution, and hardware-specific optimisations leveraging specialised accelerators. Extensive evaluation across multiple real-world testbeds demonstrates that hybrid edge-microservices architectures achieve 46% lower P99 latency and 67% higher throughput compared to monolithic approaches, while supporting 10,000 concurrent users with 100 ms latency constraints and reducing bandwidth consumption by 99.5% through selective cloud offloading. These advancements enable transformative applications in healthcare monitoring, autonomous systems, real-time IoT analytics, and personalised AI services, fundamentally reshaping how intelligence is delivered at the network edge while preserving privacy and ensuring responsiveness critical for next-generation computing paradigms.

Street view imagery (SVI) can capture urban physical space features, residents’ sentiments, and soundscape characteristics, providing insights into complex relationships between human activities and the built environment. However, existing SVI-based urban perception models have limitations, including insufficient model generality, neglect of element interactions, and lack of logical reasoning capabilities. The study proposes a multimodal large language model, StreetSenser, with powerful SVI analysis and general capabilities learned from human perception. First, an optimized dataset was designed and constructed, encompassing diverse street scenarios and annotations from human perception. Subsequently, StreetSenser was developed through a two-stage Chain-of-Perception (CoP) fine-tuning method, with the ‘Object - Attribute - Relation’ framework applied to the Qwen2.5-VL model to enable perception of complex information across visual, emotional and acoustic dimensions. Results revealed that modules such as SVI description and CoP enhance urban perception capacity of StreetSenser significantly. Furthermore, with only seven billion parameters, StreetSenser demonstrates a performance comparable to or exceeding that of large models such as GPT-4o across various tasks, while also approaching or surpassing traditional models. In addition, the compact model size enables efficient and cost-effective deployment for large-scale urban applications. The results validate the potential of StreetSenser as a key decision-support tool for urban planning.

Since the rise of the Internet, several IT services and applications have become widely accessible, making cloud computing a vital solution for its deployment for corporate and personal use. Additionally, the Internet of Things (IoT) has accelerated large-scale data generation, e.g., for monitoring applications in medical and industrial environments. Cloud computing and IoT are seamlessly integrated: IoT devices generate data later stored and accessed in the cloud, enabling efficient data use across multiple applications and processing models. Consequently, cloud services are increasingly being used for outsourcing the high processing and storage requirements demanded by IoT applications. While this integration offers significant advantages, it also presents major data security challenges, particularly concerning the confidentiality and access control of outsourced sensitive data. It is especially relevant because cloud service providers are typically assumed to be honest but curious and, hence, untrustworthy. In this context, Ciphertext-Policy Attribute-Based Encryption (CP-ABE) can successfully enforce complex access control over outsourced data. It is achieved by encrypting it using fine-grained access policies and delegating access control to decryption keys dependent on end-user attributes. Although CP-ABE offers several advantages, its wide adoption and efficient deployment in practical applications is still hindered by some issues. One of the major concerns involves the strong dependency on a centralized trusted authority setting and managing CP-ABE’s access control policies and attribute sets. This dependency represents a single point of failure that threatens the system’s continuous operation. In this paper, we eliminate CP-ABE’s dependency on a single trusted authority by adopting a decentralization strategy relying on blockchain’s main features. Therefore, we propose a blockchain-based approach to distribute among multiple peers the users’ secret keys generation and management tasks performed by the trusted authority, solving CP-ABE’s centralization problem. By combining blockchain, CP-ABE, and Elliptic Curve Integrated Encryption Scheme (ECIES), we ensure the confidentiality of CP-ABE critical components regardless of their heterogeneous privacy requirements. We evaluated our proposal considering a case study in the eHealth domain, whose results confirm its deployment feasibility in practical applications, where confidentiality and access control hold while resiliency and the system’s continuous operation are achieved.

Accurately detecting and counting potatoes during early harvest is essential for estimating yield, automating sorting, and supporting data-driven agricultural decisions. However, field environments often present practical challenges-such as soil occlusion, overlapping tubers, and inconsistent lighting-that hinder robust visual recognition. In response, we introduce SCG-YOLOv8n, a compact and field-adapted detection framework built upon the YOLOv8n architecture and specifically tailored for small-object detection in real-world farming conditions. The model incorporates three practical enhancements: a C-SPD module that preserves spatial detail to improve recognition of partially buried tubers; an S-CARAFE operator that reconstructs fine-scale features during upsampling; and GhostShuffleConv layers that reduce computational overhead without sacrificing accuracy. Through extensive field-based experiments, SCG-YOLOv8n consistently outperforms YOLOv5n and its base version across all key metrics. Float16 quantization compresses the model to 3.2 MB, enabling real-time inference on Android devices. We also developed PotatoDetector, a mobile application that demonstrates stable performance in field trials, achieving an RMSE of 1.38 and [Formula: see text] of 0.96 in counting tasks. These results suggest that SCG-YOLOv8n offers a practical and scalable tool for precision agriculture, with potential applicability to other root and tuber crop monitoring scenarios.

This study presents a heuristic approach to optimize the charging and rebalancing of automatic micromobility devices with battery constraints. The methodology integrates a vehicle routing problem to reposition and charge automatic micromobility devices and a facility location problem to ensure efficient deployment of charging locations. We first define density-based homogeneous regions through a clustering technique and then employ a continuous approximation technique to estimate the average distance between the nodes in each cluster, which is then used to assess the routing objective value. By estimating total travel distance and cost, the heuristic accommodates both known and potential repositioning needs. Using real-world data from Chicago, IL, our findings indicate that the heuristic achieves near-optimal solutions with substantial reductions in computational time, highlighting its practical applicability in real-world scenarios compared with traditional methods. Additionally, sensitivity analyses reveal the impact of battery levels and facility costs on overall performance, providing valuable insights for decision makers. The proposed approach offers a robust framework for enhancing the efficiency of micromobility systems, with promising applications in improving system resilience in disaster-affected areas and improving equitable access to underserved communities.

Among the main causes of traffic accidents and fatalities worldwide is sleepiness among drivers. Driving safety is seriously compromised by prolonged driving without enough rest because it causes microsleeps, reduces reaction times, and diminished awareness. Steering pattern analysis and physiological sensors are two examples of conventional monitoring methods that are frequently costly, invasive, or unreliable in practical settings. To get around restrictions, these studies propose deep learning and non-intrusive computer vision methods for real-time driver drowsiness detection. The device takes frontal pictures of the driver using a camera and facial landmark identification to locate and extract eye regions. A convolutional neural network model is then used to classify the eyes as either open or closed. When eye closure lasts longer than a certain threshold a indication of drowsiness and alarm is set off to alert the driver. TensorFlow, OpenCV, and Python frameworks have been used to implement the proposed system. Experimental results show that the model is robust against factors like as the presence of spectacles and achieves an overall accuracy of more than 83% under a range of scenarios, including driving during the day and at night. Furthermore, the CNNs lightweight architecture which has a maximum model size of 75 KB ensures efficient deployment on mobile devices and embedded platforms. Compared to existing systems, the suggested approach significantly reduces false alarms while maintaining real-time performance. This study demonstrates the possibilities of CNN-based methods to offer a practical, cost-effective, and scalable solution for integration enter ADAS (Advanced Driver Assistance) to improve road safety and prevent fatigue- related accidents.

The proliferation of social network data has unlocked unprecedented opportunities for extensive, data-driven exploration of human behavior. The structural intricacies of social networks offer insights into various computational social science issues, particularly concerning social influence and information diffusion. However, modeling large-scale social network data comes with computational challenges. Though large language models make it easier than ever to model textual content, any advanced network representation method struggles with scalability and efficient deployment to out-of-sample users. In response, we introduce a novel approach tailored for modeling social network data in user-detection tasks. This innovative method integrates localized social network interactions with the capabilities of large language models. Operating under the premise of social network homophily, which posits that socially connected users share similarities, our approach is designed with scalability and inductive capabilities in mind, avoiding the need for full-graph training. We conduct a thorough evaluation of our method across seven real-world social network datasets, spanning a diverse range of topics and detection tasks, showcasing its applicability to advance research in computational social science.

Abstract Prolonged storage of red blood cells (RBCs) induces morphological degradation that can compromise transfusion efficacy. Traditional quality assessment methods are often labor‐intensive and time‐consuming, limiting their utility in real‐time settings. Although deep learning has been applied to RBC imaging, most approaches require large datasets and complex architectures, making them impractical for efficient deployment. This study introduces a holographic sensor‐integrated deep learning framework for noninvasive RBC quality assessment using small datasets. A diffusion model is employed to synthetically generate phase images and segmentation masks, augmenting limited data. Self‐supervised learning with pre‐trained models further enhances classification performance while maintaining a streamlined model architecture. Compared to conventional segmentation methods, the proposed framework achieves higher accuracy and significantly faster inference. It also enables reliable detection of storage‐induced morphological changes, providing proportional indicators of transfusion viability. Experimental results validate its effectiveness as a practical tool for real‐time, sensor‐driven monitoring of RBC quality.

Nonastreda multimodal dataset for efficient tool wear state monitoring.

BFP: Balanced filter pruning via knowledge distillation for efficient deployment of CNNs on edge devices