Construction site AI: Opportunities and limits of artificial intelligence in civil engineering

Citation
smartGeotech 2025
Contribution to Conference
smartGeotech 2025  
Sensor faults in structural health monitoring (SHM) systems create significant challenges for maintaining infrastructure safety and operational efficiency. Faults, e.g. caused by sensor aging and environmental conditions, may result in inaccurate monitoring data and unreliable diagnostics. Artificial neural network (ANN) models utilizing time-series data have been proven powerful in diagnosing sensor faults in SHM systems, leveraging the ability of ANN models to analyze complex patterns within large datasets. However, the lack of transparency in – generally – black-box ANN models limits trust and usability, creating barriers to the widespread adoption of ANN models for sensor fault diagnosis. In particular, typical ANN decision-making processes lack explainability and interpretability, thus preventing end- users from confidently relying on ANN-based solutions for SHM applications. To overcome the aforementioned limitation, this study introduces an explainable sensor fault diagnosis (ESFD) approach, drawing from the field of explainable artificial intelligence (XAI), whose application in the context of SHM has been scarce. The ESFD approach essentially consists of a time-series-based ANN concept, designed to enhance the explainability and interpretability of sensor fault diagnosis in SHM by incorporating XAI methods that provide insights into decision-making processes. The ESFD approach is validated using SHM data from a pedestrian bridge for identifying combined sensor faults. The performance of the ESFD approach is assessed using XAI evaluation metrics, including “sufficiency” as a faithfulness metric. The results from the validation tests demonstrate that the ESFD approach improves the explainability and interpretability of time-series-based ANN models for sensor fault diagnosis in SHM. This advancement is expected to support the development of explainable and autonomous SHM systems. Future research will focus on decentralizing the ESFD approach for explainable, real-time decision-making to enhance scalability and reduce computational overhead.
DDC Class
600: Technology