A Retrospective Analysis of Deep Learning for Tunnel Asset Management: Balancing Efficiency, Ethics, Sustainability, and Security in the MIRET Framework

Efficient maintenance of transportation tunnels (TT) is paramount for infrastructure safety and longevity. Predictive maintenance approaches offer significant advantages, enabling proactive interventions and extended asset lifespan. However, traditional tunnel diagnostic methods can be time-consuming and expensive. This paper presents a retrospective analysis of deep learning-based methods for defect segmentation in TT images, specifically focusing on ethical, sustainable, and secure implementations within the MIRET (Management and Identification of the Risk for Existing Tunnels) framework, also considering the future constraints of the EU's AI Act. Traditional tunnel diagnostics are slow. Deep learning, particularly convolutional neural networks (CNNs), offers a promising solution for automated defect segmentation. The MIRET project aims to develop a robust risk management framework for tunnels using ARCHITA, a multi-dimensional mobile mapping system. Integrating deep learning-based defect segmentation with ARCHITA's data transforms tunnel inspection and risk assessment. Applying deep learning in MIRET raises regulatory and ethical concerns under the EU's AI Act, which emphasizes fairness, transparency, and accountability. Potential biases in training data could compromise safety. The AI Act identifies certain categories as high-risk. Among these, AI systems used for managing critical infrastructure stand out. Major construction works, including tunnels and bridges, are crucial elements within the infrastructure. For tunnels, the inspection phase and subsequent vulnerability assessment are critical. MIRET-Tunnel AI, software implementation from MIRET, based on convolutional neural networks, plays a crucial role in detecting defects in tunnel lining structures, contributing to efficient asset management.