Heterogeneous Transfer Learning in Structural Health Monitoring for High Rise Structures

Abstract

Structural Health Monitoring aims to utilise sensor data to assess the integrity of structures. Machine learning is opening up the possibility for more accurate and informative metrics to be determined by leveraging the large volumes of data available in modern times. An unfortunate limitation to these advancements is the fact that these models typically only use data from the structure being modeled, and these data sets are typically limited, which in turn limits the predictive power of the models built on these datasets. Transfer learning is a subfield of machine learning that aims to use data from other sources to inform a model on a target task. Current research has been focused on employing this methodology to real-world structures by using simulated structures for source information. This paper analyzes the feasibility of deploying this framework across multiple real-world structures. Data from two experimental scale models were evaluated in a multiclass damage detection problem. Damage in the structures was simulated through the removal of structural components. The dataset consists of the response from accelerometers equipped to the structures while the structures were under the influence of an external force. A convolution neural network (CNN) was used as the target-only model, and a Generative adversarial network (GAN) based CNN network was evaluated as the transfer learning model. The results show that transfer learning improves results in cases where limited data on the damaged target structure is available, however transfer learning is much less effective than traditional methods when there is a sufficient amount of data available.