Failure of drone motors is one of the imminent risks both for drones themselves and humans around. This research aims to find out whether raw IMU readings alone suffice to produce sufficiently reliable alarms allowing flight systems (or operators, in case of manual control) sufficient time to prevent the failure. Three popular machine learning models-Random Forest, XGBoost, and a multilayer perceptron (MLP)-are compared in terms of their ability to perform binary classification using RflyMAD dataset. In order to avoid any kind of temporal leakage of information, the traintest split is performed based on flight-level splitting rather than chronological order of measurements. All experiments involve severe class imbalance of $111.6: 1$ as well as two prediction horizons- 0.5 second and 3.0 seconds prior to the failure onset. At $h=0.5$, the best results in terms of PR-AUC (0.753), precision (0.919), and Brier score (0.003) are shown by MLP. As predicted, performance worsens for a longer prediction horizon; XGBoost achieves the best scores (PR-AUC 0.550) for $\boldsymbol{h}=\mathbf{3} \boldsymbol{.} \mathbf{0}$. Most important features include acceleration along zaxis ($a_{z}$).
The growth of the Internet of Things (IoT) means that we are seeing many resource-constrained devices deployed in smart homes, health care, industry and beyond. But these environments are still vulnerable to advanced botnet attacks because of their constrained computing resources and their large cyberattack surface. To address this issue, this paper proposes an intrusion detection system (IDS) incorporating an optimized machine learning approach to protect IoT environments. Our method is evaluated on the Bot-IoT dataset with data pre-processing, class imbalance (SMOTE), and Z-score normalisation. Several classifiers, such as Logistic Regression, Naive Bayes, K-Nearest Neighbors, Decision Tree, Support Vector Machine, and Random Forest, are built and evaluated. The performance of these models is evaluated using accuracy, precision, recall and F1-score, achieving detection performance of up to $99.9 \%$. The best performance is achieved by the optimized Random Forest model. Moreover, we investigate hardware resource usage, and the inference time for the proposed approach, showing that the proposed method ensures a short response time and high performance, making it an ideal solution for real-time applications on resource-constrained IoT devices. Our approach provides a performance-complexity trade-off to enable the use of larger deep network models.
The absence of datasets containing emotional information and models tailored for the Arabic language significantly hinders the ability to develop empathetic Arabic chatbots. In this context, we present a model for building an empathetic Arabic chatbot, based on an encoder-decoder model, with the core part being AraBERT. The model was trained on a harmonized version of the Empathetic Dialogues dataset to learn generating emotional responses. The training was carried out on Google Colab using an A100 GPU and the evaluation of the model was based on BLEU, perplexity, and validation loss. For the purpose of evaluating the model in real-life conditions, an interactive Gradio interface was also developed. The chatbot was able to produce coherent and empathetic responses, although it did perform poorly for neutral stimuli and was not very diverse in responses.
Wireless communication systems continue to evolve, with 5 G and emerging 6 G networks demanding antennas that cover multiple frequency bands without hardware changes. This paper presents a frequency-reconfigurable rectangular microstrip patch antenna for dual-band operation at 3.5 GHz and 6.5 GHz, targeting 5G and prospective 6 G applications. Two PIN diodes selectively connect the main radiating patch to additional slots, altering the effective electrical length and shifting the resonant frequency. The antenna is designed and simulated in CST Microwave Studio, with PIN diode behavior modeled in LTSpice. An FR-4 substrate with relative permittivity $\varepsilon_{r}=4.4$ and thickness 1.6 mm is used throughout. Simulations show stable switching between 3.59 GHz and 6.52 GHz. At both frequencies, the antenna exhibits good impedance matching, with VSWR below 1.2, realized gains of 5.32 dBi and 5.29 dBi, and radiation efficiencies of 87.0% and 69.7%, respectively.
Language impairments in Alzheimer’s disease (AD) often appear in spontaneous narratives as reduced informativeness, lexical retrieval difficulties, and disrupted discourse organization. While recent text-based approaches leverage pretrained language models, many pipelines still treat narratives as purely sequential signals, which may under-represent nonlocal semantic relations such as topic returns and repetitions. In this paper, we propose a text graph-based framework that explicitly models discourse structure from transcripts. Starting from subject audio, we obtain timestamped transcripts using Whisper and segment them into short discourse units. Each segment is encoded with a frozen Transformer sentence embedding. To reduce overfitting in small clinical datasets, embeddings are projected using PCA, robust-scaled, and $\ell_{2}$-normalized. We then construct a deterministic small-world discourse graph by combining sequential edges (temporal adjacency) with semantic recurrence edges defined by a cosine similarity above a threshold. A Graph Convolutional Network (GCN) performs graph-level classification. Experiments on the ADReSSo benchmark using the official split show that incorporating semantic recurrence consistently improves performance over temporal-only graphs consequently, our best configuration achieves $\mathbf{9 2. 2 \%}$ accuracy. The proposed topology is simple, interpretable, and enables controlled ablations to quantify the contribution of discourse connectivity for AD screening from text.
Instead of relying solely on terrestrial fixed infrastructure, non-terrestrial networks (NTNs) use aerial or space-based platforms to support terrestrial communication facilities. NTNs can provide connectivity in emergency situations and in remote areas such as maritime, desert, mountainous, or polar regions. On the other hand, the open radio access network (ORAN) facilitates the functional disaggregation of the central unit (CU), distributed unit (DU), and radio unit (RU), enabling flexible deployment, reduced power consumption, support for multi-layer architectures, scalability, network slicing, AI-driven control, multi-vendor interoperability, cost efficiency, and improved latency. These advantages are particularly critical for NTNs, for which power consumption and energy constraints are additionally major concerns, and where latency minimization is a key challenge due to the long-distance nature of satellite and aerial communication systems. Accordingly, this paper proposes an ORAN functional splitting architecture for NTNs and evaluates the performance of various functional split options in terms of power consumption and latency for each functional unit and transmission link. Based on the required services, the results for power, latency, and outage probability in a Rician fading channel environment are presented for all options.
The rapid emergence of new wireless technologies has driven increasing demand for fifth-generation mobile communication systems. In this work, a machine learning-assisted design approach is proposed for a microstrip patch antenna with an inset feed, specifically configured to operate in the 26 GHz 5 G band. This study focuses on predicting the geometrical parameters of the radiating patch to ensure appropriate resonant frequency and good impedance matching. Four machine learning models were selected based on their reliable predictive performances: a hybrid approach combining Gaussian Process Regression (GPR) for resonance frequency prediction and a Multilayer Perceptron (MLP) for $S_{11}$ prediction, as well as Extra Trees (ET) and Random Forest (RF) models used independently to predict both resonance frequency and reflection coefficient. To validate the proposed approaches, 20 candidate antenna configurations are selected and evaluated using CST simulations. Among these configurations, the best result is obtained at the target frequency band of 26.41 GHz, with a reflection coefficient of -56.21 dB for the ET and RF models with the same predicted parameters, confirming the effectiveness of the proposed machine learningassisted antenna design methodology. The optimized antenna structure is based on a simple rectangular geometry with compact dimensions, which facilitates its integration into miniaturized communication devices.
As digital learning environments grow in number, education professionals and platform designers increasingly are faced with the challenge of catering to diverse learning paths. Adaptive learning systems have been developed to address this challenge, and skill/knowledge graphs (KG) are used to model these systems. Simultaneously, machine learning (ML) methods are being applied to leverage these structures to recommend personalized learning paths (PLP), and to update such recommendations as learners’ needs unfold over time. This paper presents a systematic literature review (SLR). In adherence to PRISMA recommendations, this paper presents a systematic literature review (SLR) of 23 most relevant and high-impact publications in the years 2024-2026. The review identifies key algorithmic trends, including graph neural networks (GNNs), reinforcement learning (RL), and hybrid architectures. The analysis shows that although modern systems have gone further than the traditional rulebased systems, issues related to explainability, scalability, sparsity and sensitivity to the educational context remain. Synthesis of these findings in this review is used to characterize the current state of the field and identify research priority for developing more transparent, robust, and context-aware adaptive learning systems.
Person Re-identification (ReID) is the process of identifying a person based on their visual appearance across multiple cameras or sensors. In recent years, this field has significantly benefited from the rapid development of deep learning techniques. State-of-the-art methods are mainly based on Convolutional Neural Networks (CNNs), Siamese and Triplet architectures, Vision Transformers (ViT), and hybrid CNNTransformer models. In addition, contrastive learning, attention mechanisms, and multimodal approaches have improved robustness under difficult conditions, such as occlusion and lowresolution images. However, major security and privacy concerns remain, including adversarial attacks, identity spoofing, and data leakage. To mitigate these risks, several defense mechanisms have been proposed, such as adversarial training, federated learning, differential privacy, and data anonymization. This paper highlights the need to develop efficient, robust, and trustworthy ReID systems for secure deployment in realistic environments.
The proliferation of IoT and vehicular applications is increasing the volume of delay-sensitive workloads generated at the network edge. Although Multi-access Edge Computing (MEC) brings compute resources closer to users, performance still depends heavily on orchestration choices made under fluctuating load and network conditions. This paper introduces a learning-based matching strategy aimed at minimizing end-toend task completion time in a three-tier architecture (device-edge-cloud). The proposed orchestrator follows a two-stage decision process: an MLP classifier first filters offloading options that are unlikely to satisfy delay constraints, then a time estimator ranks the remaining candidates to select the most suitable destination. The learning models are trained offline from EdgeCloudSim traces collected under different vehicular load conditions, while runtime decisions rely only on lightweight online inference. We implement the method in EdgeCloudSim and assess its behavior under increasing vehicular load. Comparisons with game-theoretic and moving-average baselines indicate that the proposed matching approach preserves more stable completiontime trends as congestion grows, while keeping the orchestration process suitable for time-sensitive edge and vehicular services.
The rise of Industry 4.0 has introduced a new generation of intelligent manufacturing systems that integrate Artificial Intelligence, the Internet of Things, and Digital Twins. Among these, Digital twins play a pivotal role by providing real-time virtual representations of physical processes, enabling monitoring, optimization, and predictive maintenance. However, their effectiveness is highly dependent on the availability of highquality data. Using computer vision is a great candidate to obtain such data, but the deployment of computer vision in an industrial context remains difficult. Collecting and annotating large-scale industrial datasets remains costly, time-consuming, and often constrained by privacy and safety concerns. To address this limitation, this work explores the use of synthetic data for object detection in a tidal turbine assembly context, as a foundational step toward vision-based digital twins. Using Unity’s Perception Package and CAD models of turbine components, we generate realistic synthetic images with domain randomization and industrial background integration. Two state-of-theart detectors, YOLOv8 and YOLOv9, are fine-tuned exclusively on synthetic data and compared with models trained on real images acquired from a workstation of the production line. Experimental results show that while models trained on real data achieve higher performance in similar visual conditions, synthetic data with sufficient diversity and randomization supports strong generalization to unseen real scenes, confirming its potential as a scalable and privacy-preserving alternative for industrial object detection.
Deep learning models often require substantial amounts of labeled data to realize their full generalization potential. However, there are many scenarios where the collection of labeled data is prohibitively expensive or labor-intensive, which may lead to low-data conditions. Generative data augmentation is one proposed technique to overcome this data scarcity, and its success heavily relies on the quality of the generated data. In this work, we study conditional generative adversarial networks (cGANs) for data augmentation in image classification problems where the available data is extremely limited. We propose a simple framework that integrates class-conditional image generation with discriminator-based filtering to identify high-quality synthetic samples. We conduct experiments on the MNIST dataset using only $0.5 \%$ of the training data. The baseline CNN in this setup yields an accuracy of 0.8245, whereas data augmentation with discriminator-filtered synthetic samples reaches 0.8755, thus providing an absolute gain of $\mathbf{5. 1}$ percentage points. Importantly, this gain is realized by using only a fraction of the generated samples, illustrating that the success of generative data augmentation largely depends on the quality, rather than the quantity, of the samples. These results show that quality-aware generative data augmentation has the potential to greatly improve deep learning in data-scarce regimes.
In semi-arid regions, where crop yield is increasingly threatened by climatic unpredictability and rising temperatures, monitoring agricultural water stress is a significant concern. For the long-term identification and mapping of agricultural water stress in the Ourika Basin (Morocco) between 1990 and 2024, this study proposes an integrated remote sensing and artificial intelligence framework. Google Earth Engine was used to handle multi-source environmental variables, such as topography data, meteorological variables (precipitation and evapotranspiration), and Landsat-derived vegetation and thermal indices (NDVI, NDWI, MSI, LST). Five approaches: Random Forest, XGBoost, LightGBM, AdaBoost, and GRU were employed to estimate the objective variable, the Crop Water Stress Index (CWSI). Standard regression measures ($R^{2}$, RMSE, and MAE) and a temporal split approach were used to assess the model’s performance. Strong generalization ability for spatial stress prediction was demonstrated by Random Forest, which had the best predictive performance on the test set ($R^{2}=0.81$). A comparison with SPEI3 demonstrated the complimentary relation between agricultural and meteorological drought indicators, while seasonal analysis showed clear summer stress peaks. For long-term drought monitoring and climate-resilient agriculture management in semi-arid regions, the suggested framework offers a scalable, repeatable, and spatially explicit method.
Automatic text summarization shortens an input document to get important ideas and understandable representation of the original text that is less lengthy and more focused. ive summarization is very challenging as it needs a profound level of semantic understanding, besides the capability of generating new sentences that are not literally present in the initial text. The volume of Arabic research in this field is limited when compared to extractive summarization and other languages. A crucial step in automatic text summarization is Preprocessing. It directly affects the quality and consistency of the input data. In our study, we intend to investigate the effect of lemmatization as a preprocessing step on Arabic text summarization and examine whether it can improve performance, using the AraBART model. We applied two preprocessing pipelines: the first follows standard normalization and cleaning procedures, while the second extends this approach by incorporating Farasa lemmatization to reduce inflectional variation and map words to their canonical forms. Our evaluation on the XL-Sum dataset combines ROUGE (R1, R2, RL) and BERTScore showing competitive results.
This work investigates an adaptive terminal sliding mode control (ATSMC) approach for cyber-physical systems (CPS), with a focus on robot manipulators subject to cyberattacks on sensor measurements and actuator inputs. First, an adaptive technique is employed to online detecting different possible unknown attacks. Second, the concept of predefined performance is introduced within sliding variables design to force their terminal convergence. Third, the reaching control law is designed to ensure the convergence of the system states despite different cyber-attacks. The whole control closed-loop system seeks to guarantee stability and security of the cyber-physical systems under different cyber-attacks. Finally, the effectiveness of the system is validated through numerical simulation.
Traffic safety, vehicle operating costs, and maintenance planning are all directly impacted by road surface degradations such as potholes, cracks, rutting, stripping, raveling, and bleeding. This paper reformulates the study as a reproducible benchmark for multi-class road surface degradation detection using lightweight object detectors suitable for deployment on edge devices, UAVs, and smartphones. The benchmark is based on a carefully selected dataset of $\mathbf{1 1, 0 2 4}$ annotated road images covering eight degradation classes with COCO-style boundingbox annotations. By using common data splits, image size, optimization settings, hardware conditions, and evaluation metrics, a unified protocol is defined to guarantee equitable comparison across detectors. YOLOv8s, YOLO11s, and YOLO26s are fully trained and assessed under the same conditions in the most recent benchmark release. The experimental comparison shows that YOLO11s improves recall (0.752 vs. 0.741 vs. 0.732), F1-score (0.777 vs. 0.773 vs. 0.765), mAP @ 0.5 (0.806 vs. 0.794 vs. 0.784), and mAP@0.5:0.95 (0.525 vs. 0.516 vs. 0.522), while YOLOv8s remains faster at inference time $(2.4 \mathrm{~ms}$ vs. 4.4 ms vs. 2.7 ms per image). The analysis shows that important degradations, with texture or shape can be detected easily, thin crack-like anomalies are still hard to detect and the main challenge.
Frequency Modulated Continuous Wave (FMCW) radars are used as the main perception system for Advanced Driver Assistance Systems (ADAS) applications such as Adaptive Cruise Control (ACC) and Autonomous Emergency Braking (AEB), especially for object detection in harsh environments. A review of related works shows that most existing mmWave radar studies focus on short-range detection, highlighting the need for enhancements to extend range detection especially for human targets. This paper presents an experimental evaluation of two-millimeter wave (mmWave) radar modules from Texas Instruments (TI) i.e. AWR1843AOPEVM operating in 77-81 GHz frequency range and IWR6843ISK operating in $60-64 \mathrm{GHz}$ band. Moreover, a fine-tuning of the FMCW radar parameters is proposed to enhance range detection. The results obtained show a detection range that exceeds 100 m for human targets using the IWR6843ISK radar.
Driver inattention remains a major contributor to road accidents worldwide, motivating the development of reliable and real-time Driver Monitoring Systems (DMS). While deep learning-based vision models have shown promising performance, their effectiveness is often limited by severe class imbalance in publicly available driver monitoring datasets, where safety-critical behaviors are underrepresented. In our previous work, we benchmarked several unimodal architectures on the DMD-TFIW0 dataset and identified Convolutional Neural Networks (CNN) and ResNet-18 as the most effective models in terms of accuracy and computational efficiency. In this paper, we extend that study by integrating Generative Artificial Intelligence (GenAI)-based synthetic data augmentation to balance the DMD-TFIW0 dataset, employing a diffusion-based text-to-image model to generate high-quality synthetic driver images targeting minority classes such as Drinking, Yawning, and Sleepy Driving. The hybrid real-synthetic dataset is then used to retrain the CNN and ResNet-18 models, evaluating the performance on real test data to validate an unbiased assessment. Experimental results indicate that GenAI-based data augmentation enhances classification while minimally affecting generalization. The CNN achieved an overall accuracy of $96.15 \%$ and a balanced accuracy of $94.95 \%$. ResNet-18 demonstrated improved performance, achieving an accuracy of $95.68 \%$ and an inference time of 0.10 ms per frame, thus suggesting its applicability in real-time scenarios. These results, taken together, confirm the effectiveness of controlled synthetic data generation in addressing dataset imbalance and strengthening the reliability of driver monitoring systems.
Over the past several decades, we have observed three significant technological revolutions in medical imaging, specifically as it relates to AI, the digitization of images in medicine, natural language processing algorithms, and machine learning algorithms that can analyze large-scale radiological data as a result. They allow early detection and diagnosis of diseases like cancer and provide new ground for extending the treatment planning possibilities, with sometimes even better outcomes than conventional radiological images. In this study, a computer-aided detection and classification system of brain tumors, specifically low-grade (LG) and high-grade (HG) gliomas in MRI data is proposed on BraTS-2015 dataset. Different supervised machine learning algorithms and feature extraction methods were performed after the steps of image acquisition, preprocessing, feature selection, and extraction. The models’ performance was tested by different evaluation metrics. The experimental results show that the proposed methods are accurate, fast, and robust, indicating their potential to enhance diagnostic reproducibility and reduce the processing time in brain MRI analysis.
Recent advances in self-supervised learning (SSL) enable rich, high-dimensional speech representations useful for Autism Spectrum Disorder (ASD) detection, but their direct use is challenging when clinical data are scarce. We apply a multi-criteria statistical feature-selection framework to fused SSL embeddings: utterance-level HuBERT, Wav2Vec 2.0 and WavLM vectors are concatenated, pruned to the most informative dimensions, and evaluated under a subject-independent protocol with several classifiers - Multilayer Perceptron (MLP), Random Forest (RF), Gradient Boosting (GB), and k-Nearest Neighbors (KNN). The pipeline reduces dimensionality and improves performance; the best configuration reaches an accuracy of 91.67%. WavLM and HuBERT supply most of the discriminative signal and complement each other when combined.
