Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3653110
Tymoteusz Zwierzchowski
This paper features a novel approach to modelling an inventory management system. The work takes into account a single product warehouse with multiple suppliers. Each supplier delivers its product with a certain lead time. Furthermore, the amount of product each supplier can deliver at a single time instant is limited by the supplier’s maximum order quantity. The system prioritizes faster suppliers (i.e. suppliers with shorter lead times), which can result in them delivering large portions of the full resupply orders at time instants with low enough values of the control signal. The warehouse is subject to a demand of dual nature: the first type of demand is contractual, resulting from a priori known obligations to its customers. The second type is a random, unknown term, bounded by a maximum value, realized by selling product leftover from trading with contracted customers. The controller’s goal is to ensure full demand satisfaction. We begin by employing a reference model with just one supplier and no random demand. Then, a sliding mode controller is applied to generate a desired resupply order profile capable of fulfilling the contractual demand at any time instant. This control scheme is designed to keep the amount of goods in the warehouse at its absolute minimum–in other words, the effect of the demand will always empty the warehouse at each time instant. We then continue by using this resupply order profile as a desired trajectory in a sliding mode controller for the real system. Finally, it is proven that with appropriate compensation for the random demand present in the system, this approach can achieve full demand satisfaction at all time instants in a system with multiple suppliers with varying lead times.
{"title":"Model Reference-Based Sliding Mode Control of Supply Chains With Defined Suppliers’ Delivery Capabilities","authors":"Tymoteusz Zwierzchowski","doi":"10.1109/ACCESS.2026.3653110","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3653110","url":null,"abstract":"This paper features a novel approach to modelling an inventory management system. The work takes into account a single product warehouse with multiple suppliers. Each supplier delivers its product with a certain lead time. Furthermore, the amount of product each supplier can deliver at a single time instant is limited by the supplier’s maximum order quantity. The system prioritizes faster suppliers (i.e. suppliers with shorter lead times), which can result in them delivering large portions of the full resupply orders at time instants with low enough values of the control signal. The warehouse is subject to a demand of dual nature: the first type of demand is contractual, resulting from a priori known obligations to its customers. The second type is a random, unknown term, bounded by a maximum value, realized by selling product leftover from trading with contracted customers. The controller’s goal is to ensure full demand satisfaction. We begin by employing a reference model with just one supplier and no random demand. Then, a sliding mode controller is applied to generate a desired resupply order profile capable of fulfilling the contractual demand at any time instant. This control scheme is designed to keep the amount of goods in the warehouse at its absolute minimum–in other words, the effect of the demand will always empty the warehouse at each time instant. We then continue by using this resupply order profile as a desired trajectory in a sliding mode controller for the real system. Finally, it is proven that with appropriate compensation for the random demand present in the system, this approach can achieve full demand satisfaction at all time instants in a system with multiple suppliers with varying lead times.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7764-7775"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11346488","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982230","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3651811
Tarek Sallam;Ahmed M. Attiya
Designing efficient and reliable two-dimensional (2D) beamforming for phased array antennas (PAAs) remains a significant challenge because of the heavy computational demands involved. In this study, we introduce a beamforming framework that adopts a physics-informed deep neural network (PIDNN). The proposed model incorporates physical principles directly into the training process through a customized loss function, which minimizes the mean squared error between the array response and a target reference signal. The beamforming weights produced by the PIDNN are systematically evaluated against the theoretically optimal Wiener solution. To assess robustness, the method is implemented on an $8times 8$ PAA and evaluated against both a shallow architecture—the radial basis function neural network (RBFNN)—and a deeper model, the convolutional neural network (CNN). Moreover, the PIDNN is applied to a large PAA to assess its performance when the number of antenna elements increases. Experimental results demonstrate that the PIDNN closely approximates Wiener-optimal weights while maintaining robustness to array imperfections and achieving this with markedly reduced computational cost.
{"title":"Physics-Informed Deep Learning for 2-D Phased Array Beamforming With Imperfection Tolerance","authors":"Tarek Sallam;Ahmed M. Attiya","doi":"10.1109/ACCESS.2026.3651811","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3651811","url":null,"abstract":"Designing efficient and reliable two-dimensional (2D) beamforming for phased array antennas (PAAs) remains a significant challenge because of the heavy computational demands involved. In this study, we introduce a beamforming framework that adopts a physics-informed deep neural network (PIDNN). The proposed model incorporates physical principles directly into the training process through a customized loss function, which minimizes the mean squared error between the array response and a target reference signal. The beamforming weights produced by the PIDNN are systematically evaluated against the theoretically optimal Wiener solution. To assess robustness, the method is implemented on an <inline-formula> <tex-math>$8times 8$ </tex-math></inline-formula> PAA and evaluated against both a shallow architecture—the radial basis function neural network (RBFNN)—and a deeper model, the convolutional neural network (CNN). Moreover, the PIDNN is applied to a large PAA to assess its performance when the number of antenna elements increases. Experimental results demonstrate that the PIDNN closely approximates Wiener-optimal weights while maintaining robustness to array imperfections and achieving this with markedly reduced computational cost.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7841-7849"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339486","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982149","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3652353
N. Chotikakamthorn
This study addresses the problem of detecting disengagement in online low-stakes tests used in blended learning within higher education. The detection method was developed based on an analysis of item responses and associated response times. The method applied the ex-Gaussian mixture model to response times, rather than the conventional lognormal model. The mixture component with the smallest Gaussian mean was chosen to represent the response times distribution of early correct responses. The selected mixture component was used to obtain the model’s mode, which then served as the threshold for classifying item responses into early and subsequent response groups. Based on the two classified groups, descriptive statistics and graphical visualizations were introduced to support manual inspection and provide insight into item- and person-level characteristics. A test statistic for disengagement detection was formulated based on the distribution of the number of early responses. Drawing on prior knowledge of the success probabilities associated with disengaged responses, two detection boundaries were defined to classify item-preknowledge and rapid-guessing behaviors. Unlike existing model-based methods for rapid guessing and item preknowledge behavior detections, the proposed non-parametric method does not require prior knowledge of item or person parameters, nor does it involve modeling or estimating such characteristics. The method’s performance was assessed using both real and simulated data, and results for true positive rates and false positive rates were reported under various test conditions. The findings indicate that the method’s performance improves with an increasing number of test items and a higher proportion of disengaged responses. Simulation results further demonstrated the method’s robustness to measurement error and small variations in response times, in contrast to the person-level adaptation of the NT10 and CUMP methods, whose performance varied significantly under the same conditions.
{"title":"Item Response Time Analysis Using Ex-Gaussian Distribution for Disengagement Detection in Online Low-Stakes Tests","authors":"N. Chotikakamthorn","doi":"10.1109/ACCESS.2026.3652353","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3652353","url":null,"abstract":"This study addresses the problem of detecting disengagement in online low-stakes tests used in blended learning within higher education. The detection method was developed based on an analysis of item responses and associated response times. The method applied the ex-Gaussian mixture model to response times, rather than the conventional lognormal model. The mixture component with the smallest Gaussian mean was chosen to represent the response times distribution of early correct responses. The selected mixture component was used to obtain the model’s mode, which then served as the threshold for classifying item responses into early and subsequent response groups. Based on the two classified groups, descriptive statistics and graphical visualizations were introduced to support manual inspection and provide insight into item- and person-level characteristics. A test statistic for disengagement detection was formulated based on the distribution of the number of early responses. Drawing on prior knowledge of the success probabilities associated with disengaged responses, two detection boundaries were defined to classify item-preknowledge and rapid-guessing behaviors. Unlike existing model-based methods for rapid guessing and item preknowledge behavior detections, the proposed non-parametric method does not require prior knowledge of item or person parameters, nor does it involve modeling or estimating such characteristics. The method’s performance was assessed using both real and simulated data, and results for true positive rates and false positive rates were reported under various test conditions. The findings indicate that the method’s performance improves with an increasing number of test items and a higher proportion of disengaged responses. Simulation results further demonstrated the method’s robustness to measurement error and small variations in response times, in contrast to the person-level adaptation of the NT10 and CUMP methods, whose performance varied significantly under the same conditions.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7860-7878"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11343808","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982150","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3652328
Woohyeong Lee;Byungkwon Park
The rapid integration of distributed energy resources requires optimization and control of power systems with many controllable devices, driving a growing interest in efficient distributed optimization algorithms. To this end, this paper proposes and examines learning-augmented initialization to enhance the convergence speed of the Alternating Direction Method of Multipliers (ADMM) for solving the distributed DC and AC optimal power flow (OPF) problems. The core concept is to leverage deep learning techniques, designed with feedforward and recurrent neural networks, as auxiliary tools to accelerate the convergence of ADMM. We perform comprehensive numerical case studies and empirically validate the benefits of the proposed methods on the IEEE 14, 118, and 1888-bus test networks in the DC model and IEEE 14, 118, and 2746-bus test networks in the AC model under different loading scenarios. In particular, the proposed method has achieved up to a 78% reduction in the average number of ADMM iterations for the DC-OPF problem and a 48% reduction for the AC-OPF problem. These findings illustrate the significant potential of combining deep learning frameworks with ADMM and possibly other distributed optimization algorithms to enhance the efficiency and reliability of future power system and energy market operations.
{"title":"Examination of Learning-Augmented Approaches for Initializing Distributed Optimal Power Flow With Consensus ADMM","authors":"Woohyeong Lee;Byungkwon Park","doi":"10.1109/ACCESS.2026.3652328","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3652328","url":null,"abstract":"The rapid integration of distributed energy resources requires optimization and control of power systems with many controllable devices, driving a growing interest in efficient distributed optimization algorithms. To this end, this paper proposes and examines learning-augmented initialization to enhance the convergence speed of the Alternating Direction Method of Multipliers (ADMM) for solving the distributed DC and AC optimal power flow (OPF) problems. The core concept is to leverage deep learning techniques, designed with feedforward and recurrent neural networks, as auxiliary tools to accelerate the convergence of ADMM. We perform comprehensive numerical case studies and empirically validate the benefits of the proposed methods on the IEEE 14, 118, and 1888-bus test networks in the DC model and IEEE 14, 118, and 2746-bus test networks in the AC model under different loading scenarios. In particular, the proposed method has achieved up to a 78% reduction in the average number of ADMM iterations for the DC-OPF problem and a 48% reduction for the AC-OPF problem. These findings illustrate the significant potential of combining deep learning frameworks with ADMM and possibly other distributed optimization algorithms to enhance the efficiency and reliability of future power system and energy market operations.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7600-7615"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11343750","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982265","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3653462
Omkar Vilas Sawant;Anirban Bhowmick
Spoken language recognition in low-resource settings is hindered by domain shift and limited labeled data. We propose ProtoAlign, a teacher–student few-shot prototype alignment framework that learns domain-invariant, language-discriminative representations with minimal target supervision. The student uses a compact transformer-style backbone with Feature Reweighting Layer (FRL). Source-domain class prototypes are maintained as exponential moving averages and serve as stable anchors. A target-to-source Information Noise Contrastive Estimate(InfoNCE) alignment term pulls few-shot target embeddings toward their language-matched source prototypes, while a lightweight knowledge-distillation loss from a source-only teacher preserves source accuracy. Warm-start schedules for the alignment and distillation weights stabilize optimization, and a pairing sampler ensures each batch contains target samples with same-language source counterparts. We evaluate language recognition performance across five heterogeneous domains, namely All India Radio (AIR), Common Voice (CV), Kaggle, the Indian Institute of Technology Hyderabad (IIT-H), and Indic TTS datasets. With at most ten labeled target examples per language, ProtoAlign consistently outperforms a strong transformer baseline in cross-domain tests and produces visibly tighter, more domain-invariant clusters in the embedding space. These results indicate that prototype anchoring combined with gentle teacher guidance provides a simple, scalable, and label-efficient path to robust cross-domain spoken language recognition.
{"title":"Bridging Domain Gaps With ProtoAlign: Teacher–Student Few-Shot Prototype Alignment for Cross-Domain Spoken Language Recognition","authors":"Omkar Vilas Sawant;Anirban Bhowmick","doi":"10.1109/ACCESS.2026.3653462","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3653462","url":null,"abstract":"Spoken language recognition in low-resource settings is hindered by domain shift and limited labeled data. We propose ProtoAlign, a teacher–student few-shot prototype alignment framework that learns domain-invariant, language-discriminative representations with minimal target supervision. The student uses a compact transformer-style backbone with Feature Reweighting Layer (FRL). Source-domain class prototypes are maintained as exponential moving averages and serve as stable anchors. A target-to-source Information Noise Contrastive Estimate(InfoNCE) alignment term pulls few-shot target embeddings toward their language-matched source prototypes, while a lightweight knowledge-distillation loss from a source-only teacher preserves source accuracy. Warm-start schedules for the alignment and distillation weights stabilize optimization, and a pairing sampler ensures each batch contains target samples with same-language source counterparts. We evaluate language recognition performance across five heterogeneous domains, namely All India Radio (AIR), Common Voice (CV), Kaggle, the Indian Institute of Technology Hyderabad (IIT-H), and Indic TTS datasets. With at most ten labeled target examples per language, ProtoAlign consistently outperforms a strong transformer baseline in cross-domain tests and produces visibly tighter, more domain-invariant clusters in the embedding space. These results indicate that prototype anchoring combined with gentle teacher guidance provides a simple, scalable, and label-efficient path to robust cross-domain spoken language recognition.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10635-10653"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11346933","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026377","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3651029
Timo de Waele;Jaron Fontaine;Eli de Poorter;Adnan Shahid
This research investigates Time-Series Transformer architectures for Electrocardiogram (ECG) heartbeat classification, particularly focusing on their generalization capabilities towards new patients and varying signal sampling rates, a critical challenge in real-world clinical applications. This study conducts a systematic comparison between Transformers and Convolutional Neural Network (CNN) models using the St. Petersburg Institute of Cardiological Technics 12-lead Arrhythmia Database (INCART). Key aspects explored include the impact of different input modalities (raw ECG, Continuous Wavelet Transform (CWT) scalograms, and their combination), various Positional Encoding (PE) schemes for Transformers, and the effect of integrating expert-derived RR interval features through different feature fusion techniques. Transformers, especially with concatenation-based PE schemes and CWT or combined ECG+CWT inputs, consistently outperformed CNNs in classification accuracy and generalization when expert features were not used. They demonstrated more than 30% better generalization to unseen patients, and 20% better generalization to unseen patients and sampling rates. Ultimately, this study emphasizes that robust ECG classifiers depend heavily on deliberate architectural choices, positional encoding schemes, input representations, and the integration of expert features to handle inter-patient and sampling rate variations. Importantly, it also demonstrates that Time-Series Transformers can achieve strong results even with relatively modest model sizes and datasets.
{"title":"Generalized ECG Heartbeat Classification Using Time-Series Transformers","authors":"Timo de Waele;Jaron Fontaine;Eli de Poorter;Adnan Shahid","doi":"10.1109/ACCESS.2026.3651029","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3651029","url":null,"abstract":"This research investigates Time-Series Transformer architectures for Electrocardiogram (ECG) heartbeat classification, particularly focusing on their generalization capabilities towards new patients and varying signal sampling rates, a critical challenge in real-world clinical applications. This study conducts a systematic comparison between Transformers and Convolutional Neural Network (CNN) models using the St. Petersburg Institute of Cardiological Technics 12-lead Arrhythmia Database (INCART). Key aspects explored include the impact of different input modalities (raw ECG, Continuous Wavelet Transform (CWT) scalograms, and their combination), various Positional Encoding (PE) schemes for Transformers, and the effect of integrating expert-derived RR interval features through different feature fusion techniques. Transformers, especially with concatenation-based PE schemes and CWT or combined ECG+CWT inputs, consistently outperformed CNNs in classification accuracy and generalization when expert features were not used. They demonstrated more than 30% better generalization to unseen patients, and 20% better generalization to unseen patients and sampling rates. Ultimately, this study emphasizes that robust ECG classifiers depend heavily on deliberate architectural choices, positional encoding schemes, input representations, and the integration of expert features to handle inter-patient and sampling rate variations. Importantly, it also demonstrates that Time-Series Transformers can achieve strong results even with relatively modest model sizes and datasets.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7699-7714"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11347064","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982132","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3651992
Anna Ledwoń;Marek Natkaniec
The rapid advancement of artificial intelligence (AI) has sparked extensive discussions regarding its potential to enhance daily life through various applications. Among these, machine learning (ML) has emerged as a promising tool for detecting threats in Wi-Fi networks, a domain increasingly vulnerable to attacks due to the widespread use of wireless communication. This paper addresses the limitations of existing research, which often relies on standard metrics without a comprehensive analysis of model performance, explainability, and robustness. The study aims to provide an in-depth evaluation of ML models for Wi-Fi threat detection by employing metrics such as accuracy, precision, recall, and F1-score, alongside confusion matrices to assess classification effectiveness. Additionally, the research will analyze training and inference times, model sizes, and the impact of features using Shapley Additive Explanations (SHAP) values. Misclassifications will be scrutinized to identify potential errors stemming from dataset properties, emphasizing the necessity of thorough dataset preprocessing for broader applicability. Furthermore, the robustness of the models will be tested against adversarial attacks tailored for Wi-Fi detection. The findings will culminate in a comparative analysis of the models, underscoring the significance of each methodological step and the potential consequences of neglecting critical aspects. This work aims to contribute to the field by enhancing the understanding of ML applications in cybersecurity and promoting the development of more reliable and explainable detection systems.
{"title":"Machine Learning for Wi-Fi Intrusion Detection: A Comparative Study of Accuracy, Explainability, and Adversarial Robustness","authors":"Anna Ledwoń;Marek Natkaniec","doi":"10.1109/ACCESS.2026.3651992","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3651992","url":null,"abstract":"The rapid advancement of artificial intelligence (AI) has sparked extensive discussions regarding its potential to enhance daily life through various applications. Among these, machine learning (ML) has emerged as a promising tool for detecting threats in Wi-Fi networks, a domain increasingly vulnerable to attacks due to the widespread use of wireless communication. This paper addresses the limitations of existing research, which often relies on standard metrics without a comprehensive analysis of model performance, explainability, and robustness. The study aims to provide an in-depth evaluation of ML models for Wi-Fi threat detection by employing metrics such as accuracy, precision, recall, and F1-score, alongside confusion matrices to assess classification effectiveness. Additionally, the research will analyze training and inference times, model sizes, and the impact of features using Shapley Additive Explanations (SHAP) values. Misclassifications will be scrutinized to identify potential errors stemming from dataset properties, emphasizing the necessity of thorough dataset preprocessing for broader applicability. Furthermore, the robustness of the models will be tested against adversarial attacks tailored for Wi-Fi detection. The findings will culminate in a comparative analysis of the models, underscoring the significance of each methodological step and the potential consequences of neglecting critical aspects. This work aims to contribute to the field by enhancing the understanding of ML applications in cybersecurity and promoting the development of more reliable and explainable detection systems.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7582-7599"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11340615","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982134","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3652343
Seda Efendioglu;Huseyin Pehlivan
How formal should a sentence sound? The answer is rarely limited to formal or informal. In natural communication, formality changes gradually from casual conversation to professional writing and highly academic prose. However, for many low-resource languages, this continuum of graded stylistic shifts remains largely unmodeled. Turkish, despite its rich stylistic variation, still lacks a systematic framework for capturing such gradual shifts. This study introduces LyreSense, a framework designed to represent the full spectrum of formality in Turkish. We integrate human-written texts with annotation assisted by large language models (LLMs) and controlled synthetic text generation to construct a stylistically diverse corpus. Building on this, we propose a style-intensity–calibrated triplet loss that adapts its margin to differences in formality, enabling embeddings to disentangle subtle stylistic variation independently of semantic content. To train efficiently while preserving model capacity, we apply Low-Rank Adaptation (LoRA) during fine-tuning. Experiments across four incremental formality classes (informal, neutral, formal, and highly formal) demonstrate that LyreSense achieves Macro-F1 of 0.69. Misclassifications are concentrated between adjacent categories, reflecting the natural continuity of formality, while extreme classes are consistently distinguished. LyreSense is more than a framework for Turkish: it establishes a scalable, language-agnostic pipeline for style-sensitive NLP in low-resource settings. By moving beyond binary style distinctions, it demonstrates how lightweight, efficient models can provide nuanced, human-like style awareness for both research and practical applications.
{"title":"From Chat to Academia: Calibrating Formality in Low-Resource Languages","authors":"Seda Efendioglu;Huseyin Pehlivan","doi":"10.1109/ACCESS.2026.3652343","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3652343","url":null,"abstract":"How formal should a sentence sound? The answer is rarely limited to formal or informal. In natural communication, formality changes gradually from casual conversation to professional writing and highly academic prose. However, for many low-resource languages, this continuum of graded stylistic shifts remains largely unmodeled. Turkish, despite its rich stylistic variation, still lacks a systematic framework for capturing such gradual shifts. This study introduces LyreSense, a framework designed to represent the full spectrum of formality in Turkish. We integrate human-written texts with annotation assisted by large language models (LLMs) and controlled synthetic text generation to construct a stylistically diverse corpus. Building on this, we propose a style-intensity–calibrated triplet loss that adapts its margin to differences in formality, enabling embeddings to disentangle subtle stylistic variation independently of semantic content. To train efficiently while preserving model capacity, we apply Low-Rank Adaptation (LoRA) during fine-tuning. Experiments across four incremental formality classes (informal, neutral, formal, and highly formal) demonstrate that LyreSense achieves Macro-F1 of 0.69. Misclassifications are concentrated between adjacent categories, reflecting the natural continuity of formality, while extreme classes are consistently distinguished. LyreSense is more than a framework for Turkish: it establishes a scalable, language-agnostic pipeline for style-sensitive NLP in low-resource settings. By moving beyond binary style distinctions, it demonstrates how lightweight, efficient models can provide nuanced, human-like style awareness for both research and practical applications.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7776-7791"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11343771","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982133","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3653414
Luoyi Feng;Sha Zong
With the advancement of agricultural modernization, the role of large-scale crop image recognition in precision agriculture, pest and disease monitoring, and crop yield prediction has become increasingly significant. To enhance the accuracy and efficiency of large-scale crop image recognition, this study introduces coordinate attention mechanism and ASPP module to improve the ConvLSTM model. It further combines the Localized Convolutional Multi-Scale Temporal Network (LCMST-Net) model to efficiently classify and recognize the extracted spatiotemporal features. In the results, the improved ConvLSTM achieved a 0.979 accuracy and a 0.926 recall. The average classification accuracy of LCMST-Net was 0.982, significantly higher than the control model. LCMST-Net performed well in metrics including MSE and MAE, further validating its advantages in prediction accuracy and classification performance. Research has shown that improved ConvLSTM and LCMST-Net models have significant advantages in feature extraction and classification performance, especially when dealing with complex spatiotemporal features, they can more accurately identify different types of crops. This study contributes to the automation and intelligence of crop growth monitoring, improving agricultural production efficiency and resource utilization efficiency.
{"title":"Large-Scale Crop Image Recognition Based on ConvLSTM Spatiotemporal Feature Extraction and LCMST-Net Model","authors":"Luoyi Feng;Sha Zong","doi":"10.1109/ACCESS.2026.3653414","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3653414","url":null,"abstract":"With the advancement of agricultural modernization, the role of large-scale crop image recognition in precision agriculture, pest and disease monitoring, and crop yield prediction has become increasingly significant. To enhance the accuracy and efficiency of large-scale crop image recognition, this study introduces coordinate attention mechanism and ASPP module to improve the ConvLSTM model. It further combines the Localized Convolutional Multi-Scale Temporal Network (LCMST-Net) model to efficiently classify and recognize the extracted spatiotemporal features. In the results, the improved ConvLSTM achieved a 0.979 accuracy and a 0.926 recall. The average classification accuracy of LCMST-Net was 0.982, significantly higher than the control model. LCMST-Net performed well in metrics including MSE and MAE, further validating its advantages in prediction accuracy and classification performance. Research has shown that improved ConvLSTM and LCMST-Net models have significant advantages in feature extraction and classification performance, especially when dealing with complex spatiotemporal features, they can more accurately identify different types of crops. This study contributes to the automation and intelligence of crop growth monitoring, improving agricultural production efficiency and resource utilization efficiency.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10719-10734"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11347006","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026405","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-06DOI: 10.1109/ACCESS.2025.3644494
Shahamat Mustavi Tasin;Muhammad E. H. Chowdhury;Shona Pedersen;Malek Chabbouh;Diala Bushnaq;Raghad Aljindi;Saidul Kabir;Anwarul Hasan
Inner speech recognition has gained enormous interest in recent years due to its applications in rehabilitation, developing assistive technology, and cognitive assessment. However, since language and speech productions are a complex process, for which identifying speech components has remained a challenging task. Different approaches were taken previously to reach this goal, but new approaches remain to be explored. Also, a subject-oriented analysis is necessary to understand the underlying brain dynamics during inner speech production, which can bring novel methods to neurological research. A publicly available dataset, “Thinking Out Loud Dataset,” has been used to develop a Machine Learning (ML)-based technique to classify inner speech using 128-channel surface Electroencephalography (EEG) signals. The dataset is collected on a Spanish cohort of ten subjects while uttering four words (“Arriba,” “Abajo,” “Derecha,” and “Izquierda”) by each participant. Statistical methods were employed to detect and remove motion artifacts from the signals. A large number (191 per channel) of time-, frequency- and time-frequency-domain features were extracted. Eight feature selection algorithms are explored, and the best feature selection technique is selected for subsequent evaluations. The performance of six ML algorithms is evaluated, and an ensemble model is proposed. Deep Learning models are also explored, and the results are compared with the classical ML approach. The proposed ensemble model, by stacking the five best logistic regression models, generated a promising result with an overall accuracy of 81.13% and an F1 score of 81.12% in the classification of four inner speech words using surface EEG signals.
近年来，内部语音识别因其在康复、开发辅助技术和认知评估方面的应用而获得了极大的关注。然而，由于语言和语音产生是一个复杂的过程，因此识别语音成分仍然是一项具有挑战性的任务。以前采取了不同的方法来实现这一目标，但新的方法仍有待探索。此外，以主体为导向的分析是了解内在言语产生过程中潜在的大脑动态的必要条件，这可以为神经学研究带来新的方法。一个公开可用的数据集“Thinking Out Loud dataset”已被用于开发一种基于机器学习（ML）的技术，该技术使用128通道表面脑电图（EEG）信号对内部语音进行分类。数据集是在一个西班牙队列中收集的，每个参与者说四个词（“Arriba”，“Abajo”，“Derecha”和“Izquierda”）。采用统计方法检测和去除信号中的运动伪影。提取了大量的时间域、频率域和时频域特征（每通道191个）。探索了八种特征选择算法，并选择了最佳特征选择技术进行后续评价。对六种机器学习算法的性能进行了评价，并提出了一个集成模型。还对深度学习模型进行了探索，并将结果与经典ML方法进行了比较。本文提出的集成模型将5个最佳逻辑回归模型进行叠加，得到了利用表面脑电信号对4个内部语音词进行分类的总体准确率为81.13%，F1得分为81.12%的结果。
{"title":"Ensemble Machine Learning Model for Inner Speech Recognition: A Subject-Specific Investigation","authors":"Shahamat Mustavi Tasin;Muhammad E. H. Chowdhury;Shona Pedersen;Malek Chabbouh;Diala Bushnaq;Raghad Aljindi;Saidul Kabir;Anwarul Hasan","doi":"10.1109/ACCESS.2025.3644494","DOIUrl":"https://doi.org/10.1109/ACCESS.2025.3644494","url":null,"abstract":"Inner speech recognition has gained enormous interest in recent years due to its applications in rehabilitation, developing assistive technology, and cognitive assessment. However, since language and speech productions are a complex process, for which identifying speech components has remained a challenging task. Different approaches were taken previously to reach this goal, but new approaches remain to be explored. Also, a subject-oriented analysis is necessary to understand the underlying brain dynamics during inner speech production, which can bring novel methods to neurological research. A publicly available dataset, “Thinking Out Loud Dataset,” has been used to develop a Machine Learning (ML)-based technique to classify inner speech using 128-channel surface Electroencephalography (EEG) signals. The dataset is collected on a Spanish cohort of ten subjects while uttering four words (“Arriba,” “Abajo,” “Derecha,” and “Izquierda”) by each participant. Statistical methods were employed to detect and remove motion artifacts from the signals. A large number (191 per channel) of time-, frequency- and time-frequency-domain features were extracted. Eight feature selection algorithms are explored, and the best feature selection technique is selected for subsequent evaluations. The performance of six ML algorithms is evaluated, and an ensemble model is proposed. Deep Learning models are also explored, and the results are compared with the classical ML approach. The proposed ensemble model, by stacking the five best logistic regression models, generated a promising result with an overall accuracy of 81.13% and an F1 score of 81.12% in the classification of four inner speech words using surface EEG signals.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10811-10827"},"PeriodicalIF":3.6,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11333930","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026289","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：