On March 25th, the Luxembourg Institute of Science and Technology (LIST) hosted the inaugural BEPROACT AI/ML Taskforce Workshop at the Belval Campus. The event brought together partners from across Europe to explore the transformative potential of artificial intelligence and machine learning (AI/ML) in infrastructure asset management. Partners explored top challenges, discussed AI prediction models, and dived deep into real-world applications. Legal and ethical considerations, including the AI Act, were also addressed. The day concluded with a collaborative problem-solving session and planning for future taskforce initiatives. As the workshop fostered dynamic exchanges among partners, it aligned with a key aspect of BEPROACT: sharing knowledge and experiences through technology that connects partners and domains of individual asset types.
Main objective of the workshop
The workshop aimed to strengthen collaboration across BEPROACT partners by identifying shared challenges, surfacing best practices, and accelerating the adoption of AI/ML solutions. Key goals included showcasing successful applications, facilitating technical knowledge exchange, and laying the groundwork for scalable innovations across different infrastructure domains. Highlights included lightning talks from each cluster and expert presentations from Université de Lille on predictive modelling and image classification. The day ended with a dedicated session by experts on the practical implications of the upcoming EU AI Act, offering insights into compliance and responsible innovation.
Addressed topics
Session 1 and 2: Opening Remarks, Framing, and Roundtable introductions
To start off, the first session addressed fragmentation in research, highlighting challenges like data ownership, privacy, and model generalization. It was noted that there was an opportunity to strengthen mutual understanding among partners and improve the secure sharing of data. The session concluded with a call for a summary document to better map participants' projects and tools. In the second session, participants introduced themselves, sharing their affiliations and research focuses to foster better collaboration and understanding.
Session 3: Use Case / Cluster Lightning Talks
This session aimed to identify reusable components in AI/ML work and share key technical and data-related challenges across domains. Boguslaw Jablkowski (BASt) presented challenges with geophones, including noise interference and waveform interpretation issues. While taking audience suggestions into account, the topic of the session led to an extended debate which overlapped with the following break due to enthusiasm of the audience.
Session 4: YOLOv8 for Image Analysis
The next session featured a presentation by the University of Lille on YOLOv8, a computer vision model for image analysis, focusing on crack detection and precise localization using deep convolutional networks. The model was applied in real-time classification and segmentation for video streams, with use cases in drone-based infrastructure inspections, helmet-mounted AI for tunnel workers, and integration with web/mobile platforms. The discussion covered data storage, potential for vehicle detection in traffic surveillance, and emergency detection applications. The model's ability to incrementally learn without overwriting prior data was highlighted, and a user interface demo showed instant crack detection from photos.
Session 5: AI for Water Quality Prediction
The fifth session covered AI for water quality prediction, presented by Abbass Chreim from the University of Lille. The focus was on predicting drops in dissolved oxygen using data from Flemish rivers. Various models were tested, with LSTM (Long Short-Term Memory) performing best. This model aims to predict dissolved oxygen drops, algae blooms, and fish kills to enable proactive decision-making. Challenges included improving accuracy for edge cases and ensuring model transferability across stations, with solutions like adaptive generalization and selective retraining proposed.
Session 6: Legal and Ethical Challenges in AI
The last session focussed on legal and ethical challenges in AI, covering data management across jurisdictions, compliance, copyright issues, and Text and Data Mining exceptions. It also addressed regulations for high-risk AI systems in sectors like health and transportation. The session highlighted the need for transparency in dataset sourcing, bias identification, and performance tracking, as well as the importance of ethical guidelines. Key takeaways included integrating governance from the start and maintaining ethical documentation alongside models and code.
Partners
A diverse group of partners was present during this BEPROACT workshop, including representatives from LIST, Université de Lille, Rijkswaterstaat (RWS), Dublin City Council, BaST, Digital Flanders, RTC4Water, Luxembourg National Data Service and more. The diversity of perspectives—from technical to legal—enriched the discussions and encouraged cross-cluster knowledge transfer.
What’s next for BEPROACT
Looking ahead, the AI/ML Taskforce will continue its momentum with upcoming workshops throughout 2025 and 2026. BEPROACT partners are encouraged to actively contribute, whether by sharing knowledge, showcasing tools, tackling shared obstacles, or co-developing scalable solutions. This collaborative energy is what will allow BEPROACT to drive meaningful impact in smart asset management.
Stay tuned as BEPROACT continues to lead the way in applying intelligent technologies to build smarter, more resilient infrastructure!