https://hdl.handle.net/20.500.11984/1148 2026-08-17T10:26:30Z 2026-08-17T10:26:30Z Ugarte Valdivielso, Jone Barrenetxea, Manex Torres, Asier Stewart, Brian G. Aizpurua Unanue, Jose Ignacio https://hdl.handle.net/20.500.11984/14653 2026-07-28T06:15:51Z 2025-01-01T00:00:00Z

Power Transformer Modelling for the Evaluation of a Reinforcement Learning-Based Inrush Current Minimization Strategy Ugarte Valdivielso, Jone; Barrenetxea, Manex; Torres, Asier; Stewart, Brian G.; Aizpurua Unanue, Jose Ignacio Inrush currents in power transformers can affect the transformer lifetime and the reliability of electrical grids. Various inrush current minimization techniques have been proposed to mitigate these effects. To validate any minimization strategy, an accurate transformer model is required. This study focuses on developing a transformer model for inrush current minimization. The transformer is modelled through duality-based transformation, recognized for its accurate core representation. In addition, the laboratory environment and its constituent components are also modelled. The developed model is validated against real inrush current measurements. Finally, the model is employed to test a Reinforcement Learning (RL) based inrush current minimization strategy. The performance of the proposed method is validated by contrasting its outcomes with inrush current data obtained from traditional uncontrolled Circuit Breaker (CB) operations. The findings indicate that the proposed inrush current minimization approach reduces the peak inrush current by 77% compared to traditional uncontrolled CB switching.

2025-01-01T00:00:00Z Illarramendi, Miren Agirre, Joseba Andoni Picatoste, Aitor Igartua, Juan Ignacio https://hdl.handle.net/20.500.11984/14652 2026-07-28T06:15:50Z 2025-01-01T00:00:00Z

Brains Without Brawn: Evaluating CPU Performance for Code Generation with Large Language Models Illarramendi, Miren; Agirre, Joseba Andoni; Picatoste, Aitor; Igartua, Juan Ignacio This research presents a comparative analysis of the performance of various Large Language Models (LLMs) for code generation tasks executed on Central Processing Units (CPUs) without the use of dedicated Graphics Processing Units (GPUs). The study evaluates key metrics including inference time, code generation accuracy, CPU and memory usage, and energy consumption. By conducting repeated experiments, we assess the impact of model size and optimization on efficiency in environments lacking GPU resources. Energy consumption is measured using tools like CodeCarbon, focusing on the environmental impact of running these models on CPU-based systems. The findings offer insights into the trade-offs between model precision, resource usage, and energy efficiency, providing valuable guidance for developers and researchers aiming to balance performance and sustainability in low-resource computing environments.

2025-01-01T00:00:00Z Arana-Arexolaleiba, Nestor Serrano, Antonio Chrysostomou, Dimitrios https://hdl.handle.net/20.500.11984/14650 2026-07-28T06:15:49Z 2025-01-01T00:00:00Z

Human-assisted reinforcement learning and dynamic force patterns in contact-rich manipulation for robotic disassembly Available to Purchase Arana-Arexolaleiba, Nestor; Serrano, Antonio; Chrysostomou, Dimitrios Purpose The escalating rates of material consumption and energy usage, along with resulting waste, pose significant environmental challenges. Circular economy principles, remanufacturing and disassembly strategies emerge as potential solutions to mitigate these impacts. However, disassembly processes face a critical hurdle due to the inherent variability of products. The purpose of this study is to explore the potential of human-assisted reinforcement learning approaches to effectively manage this variability and improve the adaptability and performance of disassembly operations. Design/methodology/approach This paper investigates the application of reinforcement learning employed with robot–human interactions to address this variability in disassembly, particularly in contact-rich tasks. In addition, this paper proposes using an impedance controller augmented with overlaid force oscillation to mitigate contact forces and jamming occurrences during occluded contact-rich manipulations. Findings The findings indicate that manipulation generalization capabilities are enhanced when policies receive human hints regarding where to focus their actions. Employing force overlay types – such as Lissajous curves and spiral shapes – under different parameters significantly reduces contact forces caused by friction during disassembly. These overlays lead to up to 10% reduction in mean force magnitude, a 55% decrease in the occurrence of contact forces exceeding the established threshold and a 28% reduction in task completion time compared to non-overlay execution. Originality/value The paper’s value lies in its innovative integration of human hints into reinforcement learning frameworks, which significantly enhances the adaptability and performance of disassembly tasks. This approach not only addresses the limitations of previous methods but also demonstrates substantial improvements in efficiency and flexibility, making it highly applicable to industrial settings.

2025-01-01T00:00:00Z Barredo Ferreira, Jorge Eceiza, Maialen Flores, Jose Luis Iturbe Urretxa, Mikel https://hdl.handle.net/20.500.11984/14625 2026-07-14T06:15:54Z 2026-01-01T00:00:00Z

Sow Smarter, Not Harder: Evaluating LLM-Generated Seeds for Fuzzing Critical Infrastructure Barredo Ferreira, Jorge; Eceiza, Maialen; Flores, Jose Luis; Iturbe Urretxa, Mikel Software vulnerabilities in critical infrastructure components can lead to severe disruptions. While fuzzing effectively identifies such weaknesses, the quality of initial seed inputs significantly impacts its effectiveness. This study evaluates how large language models (LLMs) can generate better fuzzing seeds for critical infrastructure software. We compared seven LLMs—ChatGPT-4-Turbo, Claude 3.0 Opus, Claude 3.7 Sonnet, DeepSeek-V3, Gemini 2.0 Flash, Grok 3, and Mistral 7B—with manual baselines across six programs, including industrial control libraries, routing components, and network firmware. Over 20 independent 24-h campaigns per model and program, LLM-generated seeds achieved 14.8% higher code coverage, detected 56.3% more unique crashes, and reached first crashes 373.9% faster than manual methods. Performance patterns emerged across different infrastructure protocols, with certain models excelling at complex SCADA data formats while others performed better for network security components. The 56.5% computational efficiency improvement benefits resource-constrained operational technology environments. These findings demonstrate that LLM-generated seeds can meaningfully enhance vulnerability detection in software underlying critical infrastructure systems, offering a practical approach to strengthening resilience against cyber threats .

2026-01-01T00:00:00Z