Computational Resources Research Articles

Enhancing Data Processing Efficiency : The Synergy of Edge Computing and Hybrid Cloud Storage

This comprehensive article explores the transformative integration of edge computing and hybrid cloud storage, a technological convergence that is reshaping data processing architectures in the era of exponential data growth. The research delves into the fundamental principles of edge computing and hybrid cloud storage, examining their synergistic relationship in addressing the limitations of traditional centralized cloud computing. By bringing computational resources closer to data sources, this integrated approach significantly reduces latency, enhances processing efficiency by up to 50%, and improves overall system reliability. The article presents detailed case studies in autonomous driving and smart city infrastructure, showcasing real-world applications and benefits. It critically analyzes the challenges inherent in this integration, including security concerns in decentralized architectures, data consistency issues, and cost implications. Furthermore, the article explores future directions, discussing emerging technologies such as AI-powered edge devices, evolving hybrid cloud solutions, and the potential for further optimization. This research provides valuable insights for organizations and researchers navigating the complex landscape of distributed computing, offering a roadmap for leveraging edge computing and hybrid cloud storage to achieve unprecedented levels of performance, scalability, and flexibility in data management and processing.

Typical Scenario Load Identification Based on Feature Fusion and Transfer Learning

Background: The electricity demand is continuously increasing. However, various institutions, enterprises, and individuals exhibit many irregularities in their electricity usage, leading to significant wastage of electricity. To achieve effective energy management, researchers are attempting to analyze and regulate users' electricity demands by monitoring their load usage through Non- Intrusive Load Monitoring (NILM) technology. The accuracy of load identification in this technology will greatly impact the results of load monitoring. Although there are currently many articles and patents related to NILM, they utilize a large amount of computational resources and require high sampling rates from devices, yet the results are still unsatisfactory. Therefore, it is necessary to improve the accuracy of load identification in data with relatively low sampling frequencies. Objective: To improve the accuracy of load identification with low sampling frequency data, this paper proposes a typical scenario load identification method based on feature fusion and transfer learning. Methods: This method adopts the fusion of current and power factor angles to provide abundant identification information for NILM, effectively reducing the situation of single-feature overlap of different loads. By inputting the fused feature data into GoogLeNet and utilizing transfer learning for training, not only is the accuracy improved, but also the training time and the requirement for the sampling rate of training data are greatly reduced. In addition, selecting typical scenario loads can monitor loads in a targeted manner, reduce the waste of computing resources caused by irrelevant loads, and more effectively guide electricity usage strategies. Results: The proposed load identification method was tested on the low sampling frequency dataset used in this paper. It achieved an overall load identification accuracy of 94.61% across three scenarios, improving accuracy by 3% to 7% compared to other models. Conclusion: The simulation results indicate that this method achieves high load identification accuracy at low sampling frequencies. It also exhibits good generalization ability. This method not only reduces the performance requirements for monitoring equipment but also enhances monitoring efficiency.

GCS-YOLOv8: A Lightweight Face Extractor to Assist Deepfake Detection

To address the issues of target feature blurring and increased false detections caused by high compression rates in deepfake videos, as well as the high computational resource requirements of existing face extractors, we propose a lightweight face extractor to assist deepfake detection, GCS-YOLOv8. Firstly, we employ the HGStem module for initial downsampling to address the issue of false detections of small non-face objects in deepfake videos, thereby improving detection accuracy. Secondly, we introduce the C2f-GDConv module to mitigate the low-FLOPs pitfall while reducing the model’s parameters, thereby lightening the network. Additionally, we add a new P6 large target detection layer to expand the receptive field and capture multi-scale features, solving the problem of detecting large-scale faces in low-compression deepfake videos. We also design a cross-scale feature fusion module called CCFG (CNN-based Cross-Scale Feature Fusion with GDConv), which integrates features from different scales to enhance the model’s adaptability to scale variations while reducing network parameters, addressing the high computational resource requirements of traditional face extractors. Furthermore, we improve the detection head by utilizing group normalization and shared convolution, simplifying the process of face detection while maintaining detection performance. The training dataset was also refined by removing low-accuracy and low-resolution labels, which reduced the false detection rate. Experimental results demonstrate that, compared to YOLOv8, this face extractor achieves the AP of 0.942, 0.927, and 0.812 on the WiderFace dataset’s Easy, Medium, and Hard subsets, representing improvements of 1.1%, 1.3%, and 3.7% respectively. The model’s parameters and FLOPs are only 1.68 MB and 3.5 G, reflecting reductions of 44.2% and 56.8%, making it more effective and lightweight in extracting faces from deepfake videos.

Optimizing Flow and Heat Transfer With Guiding Pin Fins in a Wedge-Shaped Cooling Channel for Gas Turbine Blade Trailing Edge

Abstract In this study, an innovative guiding pin-fin array optimization has been developed through three-dimensional numerical simulations to enhance the cooling efficiency of gas turbine blade trailing edge. The guiding pin-fin array is optimized using the Kriging surrogate model and Genetic Algorithm (GA), aiming to significantly improve the heat transfer rates and uniformity within the wedge-shaped channel. The design parameter chosen for optimization is the deflection angle of each guiding pin fin, and the optimization process is conducted in two rounds. The first-round optimization yields a first-optimized guiding pin-fin array, which exhibits superior overall heat transfer performance and reasonable pressure loss compared to conventional circular, oblong, and parallel pin-fin arrays. For the first-optimized guiding pin-fin (1st-OGP) channel at the Reynolds number of Re = 50,000, the total Nusselt number and pressure loss are 44.1% higher and 9.9% lower than those of the baseline circular pin-fin array (CP), respectively. An experimental validation using the transient liquid crystal (TLC) thermography method is carried out and proves the effectiveness of the optimization process. However, it is noted that the first-optimized guiding pin-fin exhibits even lower heat transfer at a low Reynolds number of Re = 10,000, particularly in the channel middle region, which is mainly due to the incapable turning flow control in the root region of the wedged channel. To address this issue and further improve the heat transfer performance at low Reynolds numbers, a second-round optimization is performed by specifically adjusting the deflection angle of the selected guiding pin fins near the root region of the wedged channel. This secondary optimization demonstrates significant heat transfer improvements over the whole studied Reynolds number range with a reasonably reduced consumption of computational resources. The total Nusselt number and pressure loss are 69.3% higher and 11.9% lower than those of the baseline circular pin-fin array, respectively, at Re = 50,000. The optimization process proposed in this paper produces a high-performance cooling structure design with elaborate guiding pin-fin arrangements in the wedge-shaped channel, which indicates high heat transfer enhancement and relatively lower pressure loss in the wedged channel for the turbine blade trailing edge.

EKV-VBQ: Ensuring Verifiable Boolean Queries in Encrypted Key-Value Stores

To address the deficiencies in privacy-preserving expressive query and verification mechanisms in outsourced key-value stores, we propose EKV-VBQ, a scheme designed to ensure verifiable Boolean queries over encrypted key-value data. We have integrated blockchain and homomorphic Xor operations and pseudo-random functions to create a secure and verifiable datastore, while enabling efficient encrypted Boolean queries. Additionally, we have designed a lightweight verification protocol using bilinear map accumulators to guarantee the correctness of Boolean query results. Our security analysis demonstrates that EKV-VBQ is secure against adaptive chosen label attacks (IND-CLA) and guarantees Integrity and Unforgeability under the bilinear q-strong Diffie–Hellman assumption. Our performance evaluations showed reduced server-side storage overhead, efficient proof generation, and a significant reduction in user-side computational complexity by a factor of log n. Finally, GPU-accelerated optimizations significantly enhance EKV-VBQ’s performance, reducing computational overhead by up to 50%, making EKV-VBQ highly efficient and suitable for deployment in environments with limited computational resources.

Deep Earth System Data Laboratory (DeepESDL)

Abstract. The Deep Earth System Data Lab (DeepESDL) provides an AI-ready, collaborative environment for researchers aiming to study the Earth's complex dynamics using various datasets and empirical approaches. Recently opened to Early Adopters, it builds on projects like CAB-LAB and ESDL, utilizing well-established Python and Julia technology stacks. DeepESDL offers programmatic access to extensive analysis-ready data cubes and computational resources, enabling researchers to focus on analysis without extensive preparations. Scientists can use persistently available data cubes or generate user-tailored cubes from own data or publicly available datasets. The goal is to streamline data processing through empirical or AI methods within high-dimensional Earth Observation workflows. DeepESDL addresses the complete research cycle, from discovery of earth data to powerful analyses, collaborative scientific research, advanced data visualisation and publication of results, promoting FAIR and Open Science. Apart from serving as a research environment, DeepESDL showcases scientific use cases and supports educational purposes through capacity building, academic programs, and Open Science initiatives. This paper presents an overview of DeepESDL.

Thought-Controlled Computer Applications: A Brain–Computer Interface System for Severe Disability Support

This study introduces an integrated computational environment that leverages Brain–Computer Interface (BCI) technology to enhance information access for individuals with severe disabilities. Traditional assistive technologies often rely on physical interactions, which can be challenging for this demographic. Our innovation focuses on creating new assistive technologies that use novel Human–Computer interfaces to provide a more intuitive and accessible experience. The proposed system offers four key applications to users controlled by four thoughts: an email client, a web browser, an e-learning tool, and both command-line and graphical user interfaces for managing computer resources. The BCI framework translates ElectroEncephaloGraphy (EEG) signals into commands or events using advanced signal processing and machine learning techniques. These identified commands are then processed by an integrative strategy that triggers the appropriate actions and provides real-time feedback on the screen. Our study shows that our framework achieved an 82% average classification accuracy using four distinct thoughts of 62 subjects and a 95% recognition rate for P300 signals from two users, highlighting its effectiveness in translating brain signals into actionable commands. Unlike most existing prototypes that rely on visual stimulation, our system is controlled by thought, inducing brain activity to manage the system’s Application Programming Interfaces (APIs). It switches to P300 mode for a virtual keyboard and text input. The proposed BCI system significantly improves the ability of individuals with severe disabilities to interact with various applications and manage computer resources. Our approach demonstrates superior performance in terms of classification accuracy and signal recognition compared to existing methods.

An Echo State Network-Based Light Framework for Online Anomaly Detection: An Approach to Using AI at the Edge

Production efficiency is used to determine the best conditions for manufacturing goods at the lowest possible unit cost. When achieved, production efficiency leads to increased revenues for the manufacturer, enhanced employee safety, and a satisfied customer base. Production efficiency not only measures the amount of resources that are needed for production but also considers the productivity levels and the state of the production lines. In this context, online anomaly detection (AD) is an important tool for maintaining the reliability of the production ecosystem. With advancements in artificial intelligence and the growing significance of identifying and mitigating anomalies across different fields, approaches based on artificial neural networks facilitate the recognition of intricate types of anomalies by taking into account both temporal and contextual attributes. In this paper, a lightweight framework based on the Echo State Network (ESN) model running at the edge is introduced for online AD. Compared to other AD methods, such as Long Short-Term Memory (LSTM), it achieves superior precision, accuracy, and recall metrics while reducing training time, CO2 emissions, and the need for high computational resources. The preliminary evaluation of the proposed solution was conducted using a low-resource computing device at the edge of the real production machine through an Industrial Internet of Things (IIoT) smart meter module. The machine used to test the proposed solution was provided by the Italian company SIFIM Srl, which manufactures filter mats for industrial kitchens. Experimental results demonstrate the feasibility of developing an AD method that achieves high accuracy, with the ESN-based framework reaching 85% compared to 80.88% for the LSTM-based model. Furthermore, this method requires minimal hardware resources, with a training time of 9.5 s compared to 2.100 s for the other model.

Quantum classical hybrid convolutional neural networks for breast cancer diagnosis.

The World Health Organization states that early diagnosis is essential to increasing the cure rate for breast cancer, which poses a danger to women's health worldwide. However, the efficacy and cost limitations of conventional diagnostic techniques increase the possibility of misdiagnosis. In this work, we present a quantum hybrid classical convolutional neural network (QCCNN) based breast cancer diagnosis approach with the goal of utilizing quantum computing's high-dimensional data processing power and parallelism to increase diagnosis efficiency and accuracy. When working with large-scale and complicated datasets, classical convolutional neural network (CNN) and other machine learning techniques generally demand a large amount of computational resources and time. Their restricted capacity for generalization makes it challenging to maintain consistent performance across multiple data sets. To address these issues, this paper adds a quantum convolutional layer to the classical convolutional neural network to take advantage of quantum computing to improve learning efficiency and processing speed. Simulation experiments on three breast cancer datasets, GBSG, SEER and WDBC, validate the robustness and generalization of QCCNN and significantly outperform CNN and logistic regression models in classification accuracy. This study not only provides a novel method for breast cancer diagnosis but also achieves a breakthrough in breast cancer diagnosis and promotes the development of medical diagnostic technology.

Modeling of an inductive displacement sensor based on 1DCNN-LSTM-AT

Abstract The input-output relationship of inductive displacement sensors is typically nonlinear, which demands numerous computational resources for precise calculation. Therefore, the traditional analytical methods for the inductive displacement sensors cannot meet the real-time high-precision measurement requirements. In response to the aforementioned issues, this paper proposes an optimized Long Short-Term Memory (LSTM) neural networks based on one-dimensional convolutional neural networks (1DCNN) to modeling the input-output relationship of the inductive displacement sensor. First, the measurement principle of the inductive displacement sensors was analyzed, and the analytical model of the sensor output was derived. And the influences of the key parameters of sensors on the relationship between the induced voltage and the displacement were studied. Then, the 1DCNN-LSTM-AT network for modelling the input-output relationship of the inductive displacement sensor was studied. The spatial features of historical induced electromotive force (EMF) data generated by the induction coil were initially extracted using the 1DCNN network. And these spatial features were then utilized as input to the LSTM neural network to capture the temporal features of the historical induced EMF data. Subsequently, the spatiotemporal features of the induced EMF data were fed into the regression prediction layer to compute the displacement measurement results corresponding to the current input. Moreover, the attention mechanism was used for the 1DCNN-LSTM model to enhance the prediction accuracy and stability of the model. Finally, the experimental results demonstrate that the proposed 1DCNN-LSTM-AT model achieves an average absolute percentage error (MAPE) of 3.1%, significantly outperforming traditional models such as LSTM (29.3%), CNN (4.7%), and ANN (4.3%). This paper provides a new method for modelling the nonlinear relationships of the inductive displacement sensor, and presents a fresh perspective for research in the data processing of nonlinear sensors.

Scene Classification of Remote Sensing Image Based on Multi-Path Reconfigurable Neural Network

Land image recognition and classification and land environment detection are important research fields in remote sensing applications. Because of the diversity and complexity of different tasks of land environment recognition and classification, it is difficult for researchers to use a single model to achieve the best performance in scene classification of multiple remote sensing land images. Therefore, to determine which model is the best for the current recognition classification tasks, it is often necessary to select and experiment with many different models. However, finding the optimal model is accompanied by an increase in trial-and-error costs and is a waste of researchers’ time, and it is often impossible to find the right model quickly. To address the issue of existing models being too large for easy selection, this paper proposes a multi-path reconfigurable network structure and takes the multi-path reconfigurable residual network (MR-ResNet) model as an example. The reconfigurable neural network model allows researchers to selectively choose the required modules and reassemble them to generate customized models by splitting the trained models and connecting them through modules with different properties. At the same time, by introducing the concept of a multi-path input network, the optimal path is selected by inputting different modules, which shortens the training time of the model and allows researchers to easily find the network model suitable for the current application scenario. A lot of training data, computational resources, and model parameter experience are saved. Three public datasets, NWPU-RESISC45, RSSCN7, and SIRI-WHU datasets, were used for the experiments. The experimental results demonstrate that the proposed model surpasses the classic residual network (ResNet) in terms of both parameters and performance.

Co-Estimating State of Charge and Capacity of Automotive Lithium-Ion Batteries Under Deep Degradation Using Multiple Estimators

Since battery systems typically account for over 40% of the cost of an electric vehicle, their mid-life replacements are exceptional. Therefore, the battery’s lifespan must exceed that of the vehicle. To ensure long-term and safe use, accurate state-of-charge (SOC) estimation must be maintained throughout the battery’s lifespan. This requires appropriate updates to parameters, such as capacity, in the battery model. In this context, dual extended Kalman filters, which simultaneously estimate both states and parameters, have gained interest. While existing reports on simultaneous estimators seemed promising, our study found that they performed well under low levels of battery aging but encountered issues at higher levels. Accurately reflecting the actual physicochemical changes of the parameters in aging cells is challenging for two reasons: the limited number of measurements of terminal voltage available for numerous parameters, and the weak observability of the capacity. Therefore, we combined the simultaneous estimator with a capacity estimator operated separately during charging and a sequential estimator specialized for an enhanced self-correcting model, achieving SOC accuracy within 5% even when the SOH decreased by 30%. However, there is still much work to be carried out to implement sequential estimators in battery management systems operating in real time with limited computational resources.

3D LVCN: A Lightweight Volumetric ConvNet

ABSTRACTIn recent years, with the significant increase in the volume of three‐dimensional medical image data, three‐dimensional medical models have emerged. However, existing methods often require a large number of model parameters to deal with complex medical datasets, leading to high model complexity and significant consumption of computational resources. In order to address these issues, this paper proposes a 3D Lightweight Volume Convolutional Neural Network (3D LVCN), aiming to achieve efficient and accurate volume segmentation. This network architecture combines the design principles of convolutional neural network modules and hierarchical transformers, using large convolutional kernels as the basic framework for feature extraction, while introducing 1 × 1 × 1 convolutional kernels for deep convolution. This improvement not only enhances the computational efficiency of the model but also improves its generalization ability. The pro‐posed model is tested on three challenging public datasets, namely spleen, liver, and lung, from the medical segmentation decathlon. Experimental results show that the proposed model performance has in‐creased from 0.8315 to 0.8673, with a reduction in parameters of approximately 5%. This indicates that compared to currently advanced model structures, our proposed model architecture exhibits significant advantages in segmentation performance.

SVQ-MAE: an efficient speech pre-training framework with constrained computational resources

Self-supervised learning for speech pre-training models has achieved remarkable success in acquiring superior speech contextual representations by learning from unlabeled audio, excelling in numerous downstream speech tasks. However, the pre-training of these models necessitates significant computational resources and training duration, presenting a high barrier to entry into the realm of pre-training learning. In our efforts, by amalgamating the resource-efficient benefits of the generative learning model, Masked Auto Encoder, with the efficacy of the vector quantization method in discriminative learning, we introduce a novel pre-training framework: Speech Vector Quantization Masked Auto Encoder (SVQ-MAE). Distinct from the majority of SSL frameworks, which require simultaneous construction of speech contextual representations and mask reconstruction within an encoder-only module, we have exclusively designed a decoupled decoder for pre-training SVQ-MAE. This allows the additional decoupled decoder to undertake the mask reconstruction task solely, reducing the learning complexity of pretext tasks and enhancing the encoder’s efficiency in extracting speech contextual representations. Owing to this innovation, by using only 4 GPUs, SVQ-NAE can achieve high performance comparable to wav2vec 2.0, which requires 64 GPUs for training. In the Speech Processing Universal Performance Benchmark, SVQ-MAE surpasses wav2vec 2.0 in both keyword spotting and emotion recognition tasks. Furthermore, in cross-lingual ASR for Mandarin, upon fine-tuning on AISHELL-1, SVQ-MAE achieves a Character Error Rate of 4.09%, outperforming all supervised ASR models.

Enhancing Machine Learning-Based Autoscaling for Cloud Resource Orchestration

Performance and cost-effectiveness are sustained by efficient management of resources in cloud computing. Current autoscaling approaches, when trying to balance between the consumption of resources and QoS requirements, usually fall short and end up being inefficient and leading to service disruptions. The existing literature has primarily focuses on static metrics and/or proactive scaling approaches which do not align with dynamically changing tasks, jobs or service calls. The key concept of our approach is the use of statistical analysis to select the most relevant metrics for the specific application being scaled. We demonstrated that different applications require different metrics to accurately estimate the necessary resources, highlighting that what is critical for an application may not be for the other. The proper metrics selection for control mechanism which regulates the requried recources of application are described in this study. Introduced selection mechanism enables us to improve previously designed autoscaler by allowing them to react more quickly to sudden load changes, use fewer resources, and maintain more stable service QoS due to the more accurate machine learning models. We compared our method with previous approaches through a carefully designed series of experiments, and the results showed that this approach brings significant improvements, such as reducing QoS violations by up to 80% and reducing VM usage by 3% to 50%. Testing and measurements were conducted on the Hungarian Research Network (HUN-REN) Cloud, which supports the operation of over 300 scientific projects.

The Climate Data for Adaptation and Vulnerability Assessments and the Spatial Interactions Downscaling Method

This study presents the spatial interactions downscaling (SPID) method and introduces the climate data for adaptation and vulnerability assessments (ClimAVA) dataset. SPID employs random forest models to capture the relationship between spatial patterns at global circulation model (GCM) resolution and fine-resolution pixel values. In summary, a random forest model is trained for each fine spatial resolution pixel of the reference data as the predictand, and nine pixels from the spatially resampled (coarser) version of the reference data at the GCM’s resolutions as predictors. Models are then utilized to downscale the bias-corrected GCM data. The ClimAVA-SW dataset offers a high-resolution (4 km), bias-corrected, downscaled future climate projection derived from seventeen CMIP6 GCMs. It includes three variables (daily precipitation, minimum and maximum temperature) for three shared socioeconomic pathways (SSP245, SSP370, SSP585) across the U.S. Southwest region. The ClimAVA dataset sets itself apart with the SPID method’s capacity to provide remarkable climate realism, high physical plausibility of change, and excellent representation of extreme events while maintaining user-friendliness and requiring relatively low computational resources.

YOLOv8n-CSD: A Lightweight Detection Method for Nectarines in Complex Environments

At present, the picking of nectarines mainly relies on manual completion in China, and the process involves high labor intensity during picking and low picking efficiency. Therefore, it is necessary to introduce automated picking. To improve the accuracy of nectarine fruit recognition in complex environments and to increase the efficiency of automatic orchard-picking robots, a lightweight nectarine detection method, YOLOv8n-CSD, is proposed in this study. This model improves on YOLOv8n by first proposing a new structure, C2f-PC, to replace the C2f structure used in the original network, thus reducing the number of model parameters. Second, the SEAM is introduced to improve the model’s recognition of the occluded part. Finally, to realize real-time detection of nectarine fruits, the DySample Lightweight Dynamic Upsampling Module is introduced to save computational resources while effectively enhancing the model’s anti-interference ability. With a compact size of 4.7 MB, this model achieves 95.1% precision, 84.9% recall, and a mAP@0.5 of 93.2%—the model’s volume has been reduced while the evaluation metrics have all been improved over the baseline model. The study shows that the YOLOv8n-CSD model outperforms the current mainstream target detection models, and can recognize nectarines in different environments faster and more accurately, which lays the foundation for the field application of automatic picking technology.

Multi-Class Detection of Neurodegenerative Diseases from EEG Signals Using Lightweight LSTM Neural Networks.

Neurodegenerative diseases severely impact the life of millions of patients worldwide, and their occurrence is more and more increasing proportionally to longer life expectancy. Electroencephalography has become an important diagnostic tool for these diseases, due to its relatively simple procedure, but it requires analyzing a large number of data, often carrying a small fraction of informative content. For this reason, machine learning tools have gained a considerable relevance as an aid to classify potential signs of a specific disease, especially in its early stages, when treatments can be more effective. In this work, long short-term memory-based neural networks with different numbers of units were properly designed and trained after accurate data pre-processing, in order to perform a multi-class detection. To this end, a custom dataset of EEG recordings from subjects affected by five neurodegenerative diseases (Alzheimer's disease, frontotemporal dementia, dementia with Lewy bodies, progressive supranuclear palsy, and vascular dementia) was acquired. Experimental results show that an accuracy up to 98% was achieved with data belonging to different classes of disease, up to six including the control group, while not requiring particularly heavy computational resources.

Acceleration of Numerical Modeling of Uranium In Situ Leaching: Application of IDW Interpolation and Neural Networks for Solving the Hydraulic Head Equation

The application of In Situ Leaching (ISL) has significantly boosted uranium production in countries like Kazakhstan. Given that hydrodynamic and chemical processes occur underground, mining enterprises worldwide have developed models of reactive transport. However, modeling these complex processes demands considerable computational resources. This issue is particularly significant in the context of numerical analyses of mining processes or when modeling production scenarios in uranium mining by the ISL technique, given that a substantial portion of computational resources is allocated to solving the hydraulic head equation. This work aims to explore the applicability of PINNs to accelerate hydrodynamic simulations of the ISL process. The solution of the Poisson equation is accelerated by generating an initial approximation for the iterative method through the application of the Inverse Distance Weighting (IDW) interpolation and PINNs. The impact of various factors, including the computational grid and the spacing between wells, on both the accuracy and efficiency of initial approximation and the overall solution of the elliptic equation are explored. Employing the hydraulic head distribution obtained through PINNs as the initial approximation led to a significant reduction in computation time and a decrease in the number of iterations by a factor of 2.8 to 7.10.

FGYOLO: An Integrated Feature Enhancement Lightweight Unmanned Aerial Vehicle Forest Fire Detection Framework Based on YOLOv8n

To address the challenges of complex backgrounds and small, easily confused fire and smoke targets in Unmanned Aerial Vehicle (UAV)-based forest fire detection, we propose an improved forest smoke and fire detection algorithm based on YOLOv8. Considering the limited computational resources of UAVs and the lightweight property of YOLOv8n, the original model of YOLOv8n is improved, the Bottleneck module is reconstructed using Group Shuffle Convolution (GSConv), and the residual structure is improved, thereby enhancing the model’s detection capability while reducing network parameters. The GBFPN module is proposed to optimize the neck layer network structure and fusion method, enabling the more effective extraction and fusion of pyrotechnic features. Recognizing the difficulty in capturing the prominent characteristics of fire and smoke in a complex, tree-heavy environment, we implemented the BiFormer attention mechanism to boost the model’s ability to acquire multi-scale properties while retaining fine-grained features. Additionally, the Inner-MPDIoU loss function is implemented to replace the original CIoU loss function, thereby improving the model’s capacity for detecting small targets. The experimental results of the customized G-Fire dataset reveal that FGYOLO achieves a 3.3% improvement in mean Average Precision (mAP), reaching 98.8%, while reducing the number of parameters by 26.4% compared to the original YOLOv8n.