Damage Detection for Structural Health Monitoring Using Reinforcement and Imitation Learning

The deterioration and degradation of civil infrastructures such as bridges are common issues among most industrialized countries. They result in structural damages that are responsible for important expenses associated with maintaining the safety and serviceability of infrastructures. In the context of structural health monitoring (SHM), a common approach is to use sensors in order to monitor the condition of a structure, e.g., normal or abnormal, and identify the presence of structural damages. Detecting damages using a SHM system is a difficult task because they often develop over the span of years, and they typically have an effect on structural responses which is orders of magnitudes smaller than external effects such as the temperature. In addition to detecting anomalies, we are interested in quantifying the capacity to rightfully detect these damages, i.e., quantifying the structural damage detectability, in order to evaluate the viability and usefulness of a SHM system. When a structural damage occurs, structural responses depart from a normal condition to an abnormal one. Here, an anomaly is referred to such a departure, and is quantified with respect to its characteristics consisting in an anomaly’s magnitude and its duration. These characteristics represent the severity of a structural damage and its time to develop, respectively. Therefore, the core objectives for a SHM system is to detect changes in structural responses, i.e., anomalies, and further, to associate the magnitudes of these changes to the severity of structural damages. Existing anomaly detection methodologies are typically not able to associate the occurrence of anomalies to structural damages; These methods lack a mechanism to quantify the probability of detecting anomalies as a function of an anomaly’s magnitude, its duration, and the time required to detect an anomaly after its starting time, i.e., quantifying the anomaly detectability. The capacity of existing data interpretation methods used for detecting anomalies are limited: These methods are not able to reliably discriminate between a change due to a departure from the normal condition, i.e., an anomaly, and a change due to external effect variations, e.g., the seasonal temperature fluctuations. As a result, existing methods often yield a large number of false alarms that undermine the viability of a monitoring system.