Abstract Thin-walled structures are ubiquitous in industries such as automotive, civil engineering, consumer electronics or medical devices; and many times these structures, or a significant part of them, can be approximated by a plate as in aerospace and shipbuilding. In order to prevent damages and increase safety, on-condition maintenance is increasingly being used due to the nowadays ability to continuously sensing and processing data in real-time. A key feature to assess the probability of damage is the strain caused by loads. In this paper, our goal is to estimate the strain in the whole structure based on measurements that only capture a 1.2% of its total surface. We show that the problem is equivalent to reconstructing an image with 98.8% missing pixels and present a novel procedure referred to as the Recurrent Inpainting Model (RIM). We use Finite Element Methods to simulate a thin-walled structure under different loads and create a large data set of instances. Then, we use RIM to carry out the reconstruction task along with tests of robustness against sensor failure, transferability to other sensor morphologies and generalization to 3D hollow structures. The results in all the tasks clearly outrank the next best deep learning architecture.
Cruz-Alonso et al. (Mon,) studied this question.