Industrial control systems are essential for ensuring the stable, efficient, and safe operation of complex physical processes. Fractional-order models have gained increasing relevance due to their ability to capture nonlinear and memory-dependent dynamics more accurately than classical integer-order formulations. At the same time, Programmable Logic Controllers (PLCs) remain the backbone of industrial automation thanks to their robustness, reliability, and suitability for harsh operating conditions. In this work we present a practical and fully PLC-based approach for fractional system identification, where the parameters of a fractional-order model are estimated directly on a Siemens S7–1517 PLC using Particle Swarm Optimization (PSO). Our method adopts the Grünwald-Letnikov (GL) formulation to enable an efficient real-time computation of fractional dynamics within the computational constraints of industrial hardware. By embedding both the fractional derivative evaluation and the solver within the PLC, we eliminate the need for external computational resources and demonstrate the feasibility of deploying advanced identification techniques in real industrial environments. Experimental results confirm that accurate parameter estimation can be achieved while meeting strict PLC cycle-time requirements, highlighting the potential of our approach to bring fractional-order modeling and intelligent optimization closer to widespread industrial application.
Cardiology is one of the most intricate specialities in the area of medicine. In order to detect potential diseases, echocardiography is widely used as a non-invasive imaging technique in medical practice. This medical test obtains complete images that cardiologists can evaluate for detecting possible anomalies. In this work, a deep learning-based Content-Based Image Retrieval (CBIR) system is developed to assist specialists in heart disease diagnosis. Several experiments have been carried out to determine the optimal parameters for the CBIR system. Our findings suggest that the consistency and balance of the information input into the CBIR system are more important than the amount of data used.
Fractional-order models provide a powerful framework for representing industrial processes exhibiting memory effects, anomalous diffusion, and distributed-parameter dynamics. This paper presents a hybrid time–frequency identification approach for fractional-order process models, optimized using an Enhanced Grey Wolf Optimizer (E-GWO) with adaptive leadership weighting. The proposed method integrates time-domain step response fitting and frequency-domain analysis into a unified cost function. Three fractional model structures are considered: FFOPDT, FSOPDT, and FHOPDT. Model performance is assessed in terms of parameter convergence, time-domain validation, frequency-domain accuracy, and estimation error. Simulation results indicate that the FHOPDT model outperforms lower-order alternatives on the selected benchmark, achieving significant improved fitness values and parameter estimation errors below 1% . The proposed framework offers a promising tool for the high-fidelity modeling of fractional-order industrial processes.
The aim of this study is to analyze the evolution of the number of registered programs in higher education. The research is quantitative, longitudinal, and descriptive-explanatory, with a time-series analysis design. The data were obtained from the Statistics Department of the Ministry of Education and Sciences on programs registered from 2012 to 2025. The results show that Medicine, followed by Law, are the most frequently registered programs throughout the period analyzed.
This work demonstrates accurate non-contact measurement of polymer thickness from a single short-wave infrared (1350 nm) image. We captured 1 211 ABS samples of ten nominal thicknesses (0.5 to 5.2 mm) cropped to 224 × 224 images for supervised regression. Four model families were benchmarked: a five region if interest mean values as features to feed FFNN, a full-image FFNN, a compact CNN, and five YOLOv11 backbones modified with a linear head. After 5 to 10 independent trainings per model, the tiny YOLOv11-n achieved the best median of mean square errors of 0,0046 (0.08 mm RMSE), outperforming CNN (0.015) and both dense baselines (greater than 0.07). The results confirm that SWIR imaging combined with modern detection backbones enables fast polymer metrology suitable for thickness estimation.
Landslides are a ubiquitous natural hazard that can cause severe economic losses and human casualties, particularly when they occur near populated areas and critical infrastructure. The impacts can be devastating, with the potential to bury entire towns, rupture pipelines, block waterways, and render roads unusable. In this study, we propose a comprehensive, multi-pronged approach to landslide detection and prediction. This approach integrates the analysis of Interferometric Synthetic Aperture Radar (InSAR) data from the Sentinel-1 satellite with an improved landslide inventory, a detailed examination of failure mechanisms, and the identification of key causal factors. We further enhance the methodology by employing deep learning techniques to enable the early detection of landslides. Specifically, we leverage a Long Short-Term Memory (LSTM) model to predict the movement of unstable slopes, as LSTM models have demonstrated significant promise in this field.
The rapid expansion of domestic Internet of Things (IoT) technologies has introduced significant cybersecurity challenges, particularly among digitally connected populations such as university students. However, high levels of device adoption do not necessarily imply adequate security awareness or protective practices. This study assesses IoT adoption, perceived cybersecurity knowledge, exposure to incidents, and proactive security behaviors among university students in Spain and Mexico. A descriptive cross-sectional survey design was employed to conduct a multidimensional comparison across country, educational level, and gender. The results highlight significant differences and a persistent structural gap between technological adoption and effective security practices. This study provides a transnational comparative analysis and integrates a citizen science approach within an educational intervention framework. The findings underscore the need to incorporate targeted cybersecurity literacy strategies into higher education curricula to bridge the adoption–security gap in domestic IoT environments.
This work presents a comparative analysis of deep learning architectures for hyperspectral image (hyperimage) classification, focusing on the discrimination of water-hydrocarbon mixtures. Two main architectures were evaluated: 3D Convolutional Neural Networks (3DCNNs) and hybrid 2DCNN–LSTM models, using different convolutional kernel sizes and dataset volumes. The study aimed to assess the performance, robustness, and computational efficiency of these models under varying data availability. Experimental results reveal that both architectures achieve high classification accuracy with large datasets, often surpassing 0.99 in overall accuracy (OA). However, hybrid models demonstrated superior robustness in data-scarce scenarios, outperforming 3DCNNs in stability and class-wise recall, particularly when classifying complex hydrocarbon mixtures. The hybrid models also proved computationally more efficient: they achieved comparable or better accuracy than 3DCNNs with significantly reduced training times, up to seven times faster under certain configurations. These findings highlight the potential of lightweight 2DCNN–LSTM hybrids and scalable solutions for real-time hyperspectral classification tasks, particularly when training data is limited or rapid retraining is required. The results further suggest that careful tuning of kernel size and recurrent unit complexity is critical to balancing performance and computational cost in practical applications.
Quantum Machine Learning has shown significant promise but suffers from catastrophic forgetting in continual learning scenarios. In this work, a hybrid classical-quantum approach to mitigate catastrophic forgetting is proposed, based on a rigorous evaluation of how the topology of Variational Quantum Circuits interacts with Dark Experience Replay. By constructing a controlled Noisy Intermediate-Scale Quantum simulation environment based on the Fashion-MNIST dataset, the forgetting dynamics are analyzed across three distinct quantum circuit ansatzes: Highly Expressive, Hardware-Efficient and Tree Tensor Networks. Extensive evaluation demonstrates that, when relying on naive fine-tuning, dense quantum entanglement structures exhibit more than 30
In an increasingly digitized society, facial recognition has evolved from a convenience to a critical accessibility tool, particularly for vulnerable populations facing cognitive or motor impairments. By eliminating the cognitive load of traditional passwords, these biometric systems facilitate essential digital inclusion. However, this reliance creates a heightened imperative for security; as these users often depend on passive authentication methods, they are disproportionately susceptible to sophisticated adversarial attacks. Consequently, protecting these interfaces against manipulated or synthetically generated identities is not merely a technical challenge, but a fundamental safeguard against the financial and social exploitation of at-risk demographics. This paper aims to develop a novel GAN-based facial identification architecture capable of realistic image synthesis and adversarial detection, thereby ensuring robustness against targeted security breaches. With great emphasis on real-world applications, some experiments under diverse settings have been carried out. The obtained results are very promising contributing toward safe and secure authentication technologies and open new research lines to be explored.
The transition from preventive to predictive maintenance has become a strategic priority in modern manufacturing, where unplanned downtime, quality losses, and inefficient maintenance scheduling directly affect productivity, cost, and delivery performance. This study presents an enterprise case study on machine failure prediction using neural networks and evaluates the operational impact associated with deploying a trained model in a production environment. Operational data were obtained from a metal-processing enterprise and covered critical assets including CNC machining centers, hydraulic presses, industrial compressors, and conveyor modules. The dataset consisted of multisensor time-series variables representing vibration, bearing temperature, motor current, pressure, power consumption, cycle time, alarm frequency, and operator interventions. A Long Short-Term Memory (LSTM) neural network was developed to predict the probability of machine failure within a 72-h forecasting horizon. The model was trained on historical condition-monitoring and production data after preprocessing, temporal windowing, normalization, and class-imbalance handling via class weighting. Its performance was benchmarked against logistic regression, random forest, gradient boosting, and multilayer perceptron models. In addition to conventional classification metrics, the study evaluated business-level outcomes after integrating the model into maintenance planning. The proposed LSTM architecture achieved the strongest overall predictive performance on the test set, with an accuracy of 0.990, precision of 0.781, recall of 0.872, F1-score of 0.824, and ROC-AUC of 0.941. After deployment, the enterprise observed a 37.8
The objective of this study is to evaluate and compare the effectiveness of different image feature descriptors—Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Gray-Level Co-occurrence Matrix (GLCM)—combined with various classification algorithms for the automated assessment of pavement surface conditions. A custom dataset was built from 80 high-resolution images of concrete slabs, evenly split between cracked and intact surfaces. Each image was subdivided into smaller sections to increase the data volume, resulting in 180 samples per class. The study investigates the classification performance of Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), Naïve Bayes, Decision Trees, and Artificial Neural Networks (ANNs) using these features. Through repeated 5-fold cross-validation, the models were evaluated based on accuracy, precision, recall, specificity, and F1-score. The goal is to identify the most effective combinations of descriptors and classifiers for reliable pavement defect detection, especially in resource-constrained environments or when dealing with texture-rich surfaces. The results highlight that SVM performs best with HOG, while ANNs yield superior results with LBP and GLCM features. This work contributes to the development of scalable, data-driven approaches for urban infrastructure monitoring and supports future improvements through hybrid descriptors and advanced neural architectures.
This paper presents AI2EPD, a platform oriented to healthcare professionals for monitoring therapeutic adherence in assisted living environments. The system combines home sensors with inference models to estimate adherence to activities defined in an individualised therapeutic contract. The focus is on the design and evaluation of different forms of visual representation of adherence levels, with the aim of facilitating clinical interpretation and decision making. Through multiple interfaces developed, different visualisation strategies are explored and tested in a study with healthcare professionals. The results suggest that the way data is presented has a significant impact on the perceived usefulness of the system, highlighting the need for representations adapted to different clinical profiles. This approach reinforces the role of soft technologies and interactive systems in supporting patient-centred clinical decision making.
Immersive Virtual Reality (VR) has emerged as a promising educational medium, yet its potential for supporting auditory learning remains underexplored. This study investigates the use of immersive VR environments to enhance auditory musical retention in higher education. Thirty-four university participants completed a pre–post experimental procedure involving blind auditory recognition of ten unfamiliar classical music excerpts. Between tests, participants underwent an immersive VR training experience in which each musical fragment was embedded within a dedicated three-dimensional scenario designed to support contextual and semantic association. Learning outcomes were assessed through an ordering-based auditory memory task. Results show a statistically significant improvement in post-training performance compared to baseline, indicating enhanced auditory retention following immersive exposure. Exploratory analyses suggest that prior experience with VR may moderate learning gains, highlighting the role of technological familiarity in immersive learning. These findings provide empirical evidence that carefully designed immersive VR environments can support auditory memory and contribute to the development of pedagogically grounded applications of VR in music-related education.
This paper presents a comparative study of fractional-order system identification using Particle Swarm Optimization (PSO) and the Grey Wolf Optimizer (GWO). A challenging benchmark problem involving a high-order fractional system with repeated poles and significant memory effects is considered. Model parameters are estimated by minimizing the error between the true system response and that of the identified model. The performance of both optimization algorithms is evaluated through time- and frequency-domain analyses, as well as convergence behavior and statistical metrics, including mean squared error (MSE), root mean square error (RMSE), and integral absolute error (IAE). Results indicate that GWO consistently outperforms PSO across all evaluation criteria, achieving approximately an 80% reduction in MSE and over a 60% improvement in IAE. Furthermore, GWO demonstrates faster and more stable convergence, avoids premature stagnation, and accurately identifies all system parameters, including the correct system order ( n=5 ), whereas PSO converges to an incorrect model structure. These findings highlight the robustness and effectiveness of GWO for the identification of complex fractional-order systems.
This paper presents the design and development of X-RAPT (eXtended Reality Adaptive Programming Training), a simulation framework aimed at enhancing robotics education through immersive extended reality (XR) technologies. X-RAPT is a multi-user platform that enables students to learn and practice industrial robotic programming in safe, virtual environments. The system combines real-time VR interaction, hand-tracking, voice recognition, and a web-based task editor to provide an adaptive and scalable learning experience. Using realistic 3D environments and virtual replicas of industrial robots such as the UR3 and ABB IRB120, X-RAPT allows learners to engage in training scenarios that simulate tasks like object assembly, labeling, and packaging. The platform supports collaborative sessions with role-based functionality and cross-device synchronization, enabling intuitive interaction and immediate feedback. Preliminary results indicate that X-RAPT is an effective tool for lowering the entry barrier to robotics programming and fostering hands-on learning in virtual training environments.
The integration of autonomous agents within Industry 4.0 environments relies heavily on high-precision indoor positioning, commonly enabled by Ultra-Wideband (UWB) technology. In such scenarios, motion-induced bias becomes a critical factor limiting localization accuracy, especially under high-dynamic operating conditions. In this work, we present a novel and rigorous characterization of motion-induced localization errors in the widely adopted asynchronous Symmetric Double-Sided (SDS-TWR) technique. This analysis is complemented by a comparative study with the synchronous Time-Difference-of-Arrival (TDoa) approach under varying kinematic conditions. The comparison is conducted in a realistic three-dimensional industrial scenario, where the positioning performance of both methods is systematically evaluated. To ensure a fair assessment, a Genetic Algorithm (GA) is employed to optimize the sensor deployment for each localization scheme, in accordance with their respective analytical error models. Results show a high sensitivity of the asynchronous SDS-TWR method to target motion, particularly at high speeds, which recommends the pursuit of the unaffected synchronous models for high-dynamic industrial applications.
The adjustment of input parameters in plastic injection molding remains one of the primary challenges within this manufacturing process. To determine the optimal adjustment parameters previous to the injection, it is essential to predict the critical quality indicators, such as the weight of the injected parts. Thus, in this paper, we propose the use of two different Machine Learning (ML) ensembling strategies for a robust prediction of its actual value. To build the necessary dataset, 197 parts were injected while systematically recording both the injection parameters and the resulting part weights, capturing the relationship between process parameters and part quality for the subsequent metamodel development. In this sense, this paper evaluates the bagging and stacking ensemble strategies to accurately predict part quality in the complex and environmentally sensitive injection molding process. The achieved results present a superior performance of the stacking ensemble method, offering a 99.96
Simulation is a key pillar in the development of robotic systems, especially in aerial robotics, where multi-agent coordination and autonomous decision making present significant challenges. This work introduces a system for the automatic generation of simulation environments for multi-UAV formations using ROS 2 and Gazebo. The tool enables the creation of geometric or customized configurations, generating simulation packages ready for direct deployment in ROS 2 without manual setup. This automation accelerates experimentation, reduces errors, and improves reproducibility. The generator supports user defined parameters or internal position calculations, adapting to different mission profiles and cooperative strategies. The results demonstrate the system’s capability to generate high fidelity scenarios useful for evaluating formation control and swarm behavior under realistic constraints. Aligned with the Sim2Real paradigm, this framework facilitates transfer to the real world. Future work envisions the integration of Artificial Intelligence (AI) techniques for adaptive scenarios.
Grading programming exams is challenging when runtime errors prevent automated tests from executing, leading to scores that may not reflect students’ partial understanding. This paper introduces an LLM-assisted grading pipeline for first-year Python exams that combines traditional execution checks with rubric-based inspection to award fair partial credit. The pipeline leverages in-context learning to integrate the grading workflow, from code simulation to evidence-based scoring, into a unified prompt architecture. We evaluate three prompting strategies to identify a suitable balance between processing efficiency and grading precision. Our results show a strong correlation with official human-assigned grades and indicate that the pipeline behaves as a reliable, conservative estimator. Rather than replacing the instructor, the proposed approach is intended to complement conventional execution-based assessment by improving transparency, supporting auditability, and helping teachers review partial credit decisions under an explicit rubric.
