Résumé : This paper presents the results of a long-term concrete monitoring campaign in an underground tunnel in Brussels. The system consists of several pairs of embedded ultrasonic piezoelectric transducers arranged in a pitch-catch configuration which have been placed in the concrete ceiling of the tunnel in areas where old concrete was demolished and then repaired. The monitoring system is fully automated and sends the recorded signals to a cloud-based system in our university where they are post-processed to extract indicators of structural changes in the monitored regions, and send automated email reports. A first period of six months is studied, during which the monitored areas have been repaired with skim mortar. The post-processing of the measured signals allows to identify clearly the time of repair in each zone and the evolution of the hardening process of the repair mortar. A second monitoring period of one year is then studied where it is found that despite the proposed improvement to the time stretching technique used to filter out the effects of changing environmental conditions, our indicators are still showing variations in periods when the temperature is very high in the tunnel. A method based on observed statistical correlations between the indicators computed in the different regions is then proposed and shown to be very efficient to remove the remaining variability and make the system very robust to environmental changes. Extreme value statistics is also presented as a tool to establish relevant thresholds for alarm-triggering with a very low level of potential false alarms. With all these developments, the monitoring system can automatically detect structural changes in the tunnel in real time while being robust to unavoidable changes in the environmental conditions in the tunnel.