PSI - Issue 62

Fabio Severino et al. / Procedia Structural Integrity 62 (2024) 276–284 Severino et al. / Structural Integrity Procedia 00 (2019) 000–000

277

2

challenges in system transparency, safety, and compliance with regulations that have been designed for processes executed with more traditional, human-centric, tools. Consequently, there is a need to shift towards designing AI systems that are not only accurate but also inherently trustworthy (Braunschweig and Ghallab (2021)). Li et al. (2023) highlighted the rising interest in trustworthy AI from both technical and regulatory perspectives. On the regulatory front, various national and international organizations have endeavored to establish a framework of ethical guidelines. These guidelines aim to steer AI development towards trustworthiness (Jobin et al (2019)). A notable instance is the release of the “Ethics Guidelines for Trustworthy Artificial Intelligence” by the High-Level Expert Group on AI (2019) set up by the European Commission. These guidelines list seven key requirements for an AI system to be deemed trustworthy: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; accountability; diversity, non-discrimination and fairness; and societal and environmental well-being. In Canciani et al. (2024) a framework for trustworthy AI systems has been presented; it leverages blockchain technology to achieve compliance to five of these requirements. The framework uses immutable and tamper-proof data structures that meticulously record every step in the development and deployment of AI systems. Additionally, it guarantees the non-repudiability of all operations within these processes, thus establishing a robust mechanism for maintaining the integrity and trustworthiness of AI applications. The practical application of this technology was exemplified in the context of the healthcare sector in Canciani et al. (2024). The present paper discusses the application of the same technology to critical infrastructure maintenance, presenting three application scenarios: Structural Defects Recognition by ML systems, Risk Classification of Infrastructure by Explainable AI, and Trustworthy Structural Health Monitoring using Blockchain & IoT. Through these case studies, we examine how Trustworthy AI systems can enhance operational scenarios in highly regulated domains, with a special emphasis on the role of third-party audits. Our analysis will illustrate the improvements in transparency and accountability that this approach brings to the current practice. 2. Background In recent years, the application of artificial intelligence techniques have seen a steady increase in the field of structural health monitoring, with corresponding increases in predictive power of the models that help determine the state of infrastructure works (see among others Zinno et al. (2023)). AI has been used for classification tasks (e.g., determining the risk class of an artefact), for defects recognition in images (e.g. identifying cracks), for pattern detection is inertial signals (e.g., vibration propagation), for automatic extraction of information from historical documents (e.g., OCR and indexing), etc. Meanwhile, blockchain technologies have been used mostly to record – in an immutable manner – the data obtained by in-situ sensors, contributing to the reliability of the collected data. These two techniques, however, are seldom used in conjunction; in particular, while data from sensors are often fed to AI systems for time series analysis and anomaly detection, the AI models themselves and the corresponding outputs are usually not stored in the blockchain (if any) used for sensor readings. In contrast, in our approach the models with corresponding code, training data, parameters, input readings and output values, are all stored in a ledger with the goal of obtaining, beyond the immutability of the stored data, the proof of the relationships that hold between all the components. In the following we will briefly survey the two techniques (AI and distributed ledgers) with a special focus on how they can be combined to establish the trustworthiness of the results obtained by AI methods in the field of infrastructure monitoring. 2.1. Artificial Intelligence and Machine Learning After over 70 years of developments, and following significant theoretical breakthroughs in the last decade, Artificial intelligence (AI) in general, and machine learning (ML) in particular, are rapidly transforming our world. AI encompasses all techniques that enable machines to exhibit intelligent behavior. This includes a wide range of approaches, from traditional methods like expert systems and rule-based programming to more recent advancements in machine learning, natural language processing, and computer vision. ML (a subset of AI) focuses on enabling machines to learn from data without explicit programming. The learning can be based on presenting large sets