DESIGN AND APPLICATION OF A STATISTICAL LEARNING METHODOLOGY TO REMOVE TEMPERATURE EFFECT ON STATIC SIGNALS FOR BRIDGE STRUCTURAL HEALTH MONITORING

Infrastructures are essential for the development and flourishing of countries; in this context, bridges play an irreplaceable role as links for goods and people. Nowadays, their structural integrity is threatened by ageing and increased traffic loads: according to the American Society of Civil Engineering, 42% of US bridges are at least 50 years old. The extension of the problem requires the application of techniques for resource prioritization and effective interventions. Structural health monitoring (SHM) has emerged as a promising quantitative tool for ensuring infrastructural network resilience and optimizing maintenance. In particular, damage detection is crucial in preventing catastrophic failures and minimizing repair costs. However, the influence of exogenous factors on sensor signals often complicates the identification of damage, since they might hide the insurgence of structural anomalies. The problem is especially critical when dealing with early-stage damage, characterized by a small extension. Among such external variables, temperature, in particular, significantly disturbs SHM both on seasonal and daily levels. The latter is further exacerbated by the thermal inertia of bridges, which renders the structural behavior highly non-linear with respect to temperature and solar radiation. This study proposes a novel approach to enhance the identification capability of structural health monitoring systems by removing the temperature effect on static measures thanks to a statistical learning algorithm. In particular, signals from inclinometers and LVDTs collected for a year are considered for temperature influence removal. The method works according to a two-step process described in Figure 1. First, the seasonal dependence is removed; then, the daily influence of temperature and solar radiation is treated by exploiting a machine learning algorithm able to estimate the structural response delay caused by thermal inertia. The methodology herein presented is applied and tested on two railway bridges: a Warren truss steel bridge and a CACP bridge with simply-supported spans. In both case studies, the variability of the measured signals associated with temperature reduced substantially. Indeed, the correlation coefficients dropped close to zero, and the standard deviation of signals diminished by 80% on average. Moreover, the different structural characteristics of the two test cases proved the methodology effective and versatile, suitable for different typologies of bridges. The algorithm described in this work enhances the damage detection power of structural health monitoring systems and fits with the idea of SHM deployment on a larger scale, given its successful application to diverse structural archetypes.