Luigi Portinale

In fastener production, ensuring consistent product quality is essential to protect the safety and performance of components used in demanding applications. Although mechanical sorting equipment and early generations of machine-vision technology have successfully handled dimensional and structural checks, the reliable identification of surface flaws remains difficult. Recent progress in inspection technologies, especially the combination of Artificial Intelligence (AI), Deep Learning (DL), and advanced vision systems, has opened the door to highly accurate, cost-effective surface-defect detection that can outperform human inspectors. Despite these advances, most AI/DL models are designed for general use and still require substantial application-specific training and tuning, which slows large-scale deployment and leads to high costs. The present work investigates the possibility of adopting open source DL models, specifically Transformer-based architectures and Convolutional Neural Networks (CNNs), to classify fastener component images into defective or non-defective, and to segment defective ones, in order to enhance detection localization and to support continuous, high-volume quality-control processes in a cheaper, flexible and fully modifiable way. Our findings indicate that the built-in inductive biases of CNNs continue to offer a significant edge. As a result, CNNs, being generally simpler to train, are well-suited for practical industrial applications where computational resources and latency are key concerns. At the same time, Transformer-based models can serve as complementary approaches, with their effectiveness varying based on the dataset’s properties and the type of defects involved.

Dimensionality Reduction (DR) is essential for filtering noise in high-dimensional data and enabling visualization, yet traditional non-linear methods often lack parametric mappings or distort global geometry. We propose a preliminary analysis of Contrastive Learning (CL) as a tool for explicit DR in supervised classification. By utilizing a siamese architecture with a SigLIP loss, we reconfigure latent spaces by attracting misclassified instances toward their correct class manifold while repelling them from the incorrect one. Preliminary experiments on X-ray images, sentiment analysis on textual customer reviews, and on a synthetic dataset for classification demonstrate that CL-based modification and CL-based reduction to just two dimensions can maintain or even improve classification accuracy compared to original high-dimensional spaces, while providing highly regular data structure improving data visualization as well.

The digitization of tourism has made numerous platforms available, but there remains a significant shortage of tools capable of promoting local events and activities. This study hypothesizes that a decentralized digital interface can mitigate over-tourism. We conducted an experiment by deploying a digital platform to assess the synergy between local providers and visitors through the Tourism Open-ended Experimentation Platform (TOEP), a multi-interface solution designed to directly connect tourism activity providers with residents and visitors. The platform integrates a web portal for providers and a mobile app for users, supported by a recommendation system based on individual profiles and preferences. TOEP stands out for its focus on local and sustainable tourism, facilitating the promotion of smaller events and helping to reduce the concentration of tourist flows in already saturated destinations. Initial validation, conducted with a panel of industry experts, highlighted the ease of use and good organization of the interfaces, with scores above average. Preliminary results confirm the relevance of TOEP as a tool for the sustainable and digitized promotion of local tourism, opening prospects for development towards a smart, participatory tourism ecosystem that can be replicated in different territorial contexts.

The growing decentralization of electro-energetic infrastructures and the consequent need for remote monitoring of bidirectional power flows have expanded the attack surface of modern Cyber Physical Power Systems (CPPSs), exposing them to more sophisticated cyber threats. This work proposes a cyberattack modeling and detection framework based on Dynamic Bayesian Networks (DBNs) to model and analyze causal dependencies between attack steps and detection analytics in power systems. We extend classical attack graph models by incorporating MITRE ATT&CK techniques and we present how this formalism can be converted in the corresponding DBN, enabling both predictive and diagnostic inference. To achieve real-time applicability, we analyze the Boyen–Koller (BK) approximate inference algorithm under multiple clustering strategies, including a heuristic configuration (CL) and a fully factorized baseline (FF), comparing them to exact inference (EX). Our evaluation, performed across simulated attack scenarios, with and without monitoring evidence, examines tradeoffs between inference accuracy and computational efficiency using Kullback–Leibler divergence, computation time and memory utilization metrics. Experimental results show that the CL configuration consistently achieves the lowest approximation error (KL divergence below 1.25 × 10⁻⁴ of the exact solution), while the FF clustering maintains comparable accuracy (within 2×10⁻² divergence) when performing inference on the target variable (‘‘UnstablePS’’) in scenarios without evidence. In terms of resource consumption, both CL and FF configurations reduce average computation time by an order of magnitude (0.03 s per slice vs. 0.22 s for exact inference) and memory usage by more than 95% (tens of MB vs. several GB), making them both a practical and scalable option for real-time cybersecurity inference in critical energy infrastructures.