Domain adaptation for acoustic inspection of concrete structures

本课题旨在实现混凝土结构锤击检测的自动化。监督学习方法的问题是，当系统的具体结构在训练和部署之间不同时，性能会下降（领域差距）。今年，人们的努力转向了基于深度学习的异常值检测方法，即自动编码器类型的方法，这在该领域受到了越来越多的关注。目前正在研究将这些方法与薄弱的监管框架结合起来的可能性。

This research aims at the automation of the hammering test for the inspection of concrete structures. Supervised learning approaches have theissue that when the concrete structure differs between training and deployment of the system, the performance is degraded (domain gap).This year, efforts have been shifted towards approaches based on Deep Learning for outlier detection, namely Autoencoder-type approaches, which have been gaining attention in the field. The potential of combining such methods with the weak supervision framework are being currently investigated.