Structural health monitoring via a knowledge-integration framework: prospective spatial early-warning analysis of the NPL footbridge

Introduction Condition monitoring of ageing bridges still relies mostly on periodic inspection, which identifies damage only after it has developed. This study seeks to develop prospective monitoring, in which routine sensor data give warning that the structural condition of bridges is changing while there is still time to act. The physical idea we draw on is critical slowing, where a structure approaching a change of state recovers more slowly from small disturbances and becomes more variable. The challenge is to turn this idea into a practical detector that can work on ordinary bridge instrumentation, separate real structural change from seasonal effects, and be evaluated without any information flowing backwards from the evaluation period into the method’s training. Methods We study the National Physical Laboratory (NPL) Footbridge, an instrumented reinforced-concrete cantilever, using its tilt and temperature records for May to October 2012. The entire pipeline is prospective: the environmental correction is a deliberately simple regression fitted on the stable baseline only, chosen after a fuller 17-term model proved too ill-conditioned for reliable extrapolation; the detectors are fitted, calibrated, and evaluated on strictly separated time windows. Results The spatial pattern is clear under this pipeline: all three sensors in the damage zone show a significant rise in variance through the autumn, while all three reference sensors show no significant trend, and the result does not depend on which reference sensor is chosen. A recursive filter provides a complementary early-warning signal through the growing autocorrelation of its innovations, which is robust across sampling rates from 5 minutes to 1 hour. Detectors which carry these physical indicators respond 3 weeks earlier than purely statistical detectors and cover most of the predefined evaluation interval, at the cost of alarms in mid-August whose status—early warning or false alarm—cannot be settled without independent structural evidence; we present both readings. We also openly report that the classical autocorrelation and dominant-eigenvalue indicators are not robust on these data and that no failure time can be extrapolated. Discussion The study is a single-structure proof of concept on a post-damage baseline; the complete code and processed data are provided with the study.