Robust structural damage detection with deep multiple instance learning for sensor fault tolerance

Abstract

Many structural health monitoring systems rely on signals collected from sensors to localise and quantify damage on a structure. In the last decade, many machine learning models have been proposed to detect structural damage. These models in general are trained by data generated from finite element analyses and are used for structural damage detection based on the data measured at the same degrees of freedom of the structure as those used to train the model. Sensor failure – where one or more sensors does not produce a usable signal – is a common and significant problem, especially under extreme conditions such as severe impact or natural disasters like cyclones and earthquakes, leading to the trained model not applicable for damage detection because of unavailability of data at some degrees of freedom. Despite this, few methods have been developed to address such a challenge. This paper proposes a deep learning approach which views structural damage identification as a case of multiple instance learning to address sensor failure. The new method is trained and evaluated on numerical simulations, followed by validation on an experimental case. The results of the studies show strong performance in accurately predicting structural damage with data from less number of sensors compared to those used in initial training of the model, even when more than half of the original sensors fail.