A machine learning approach for the automatic long-term structural health monitoring

Measuring the response of a structure to the ambient and service loads is a source of information that can be used to estimate some important engineering parameters or, to a certain extent, to characterize the structural behavior as a whole. By repeating the data acquisition over a period of time, it is possible to check for variations in the structure’s response, which may be correlated to the appearance or growth of a damage (e.g. following some exceptional event as the earthquake, or as a consequence of materials and components aging). The complexity of some existing structures and their environment very often requires the execution of a monitoring plan in order to support analyses and decisions through the evidence of measured data. If the monitoring is implemented through a sensor network continuously acquiring over time, then the evolution of the structural behavior may be tracked continuously as well. Such approach has become a viable option for practical applications since the last decade, as a consequence of the progress in the data acquisition and storage systems. However, proper methods and algorithms are needed for managing the large amount of data and the extraction of valuable knowledge from it. This article presents a methodology aimed at making automatic the process of structural monitoring in case it is carried continuously over time. It relies on some existing methods from the machine learning and data mining fields, which are casted into a process targeted to delimit the need of the human being intervention to the training phase and the engineering judgment of the results. The methodology has been successfully applied to the real-world case of an ancient masonry bell tower, the Ghirlandina Tower (Modena, Italy), where a network made of 12 accelerometers and 3 thermocouples has been acquiring continuously since August 2012. The structural characterization is performed by identifying the first modes of vibration, whose evolution over time has been tracked.