42 G. P. Tsialiamanis et al. Fig. 5.1 Traditional (a) and transfer (b) learning schemes (following [2]) problems and attempting transfer of information between the two sub-problems in a manner motivated by transfer learning [2]. Transfer learning is the procedure of taking knowledge from a source domain and task and applying it to a different domain and task to help improve performance on the second task [2]. Transfer learning is useful because a model trained on a dataset can not naturally be applied on another due to difference in data distribution, but can be further tuned to also apply on the second dataset. An accurate representation of the difference between traditional and transfer learning schemes can be seen in Fig. 5.1. The SHM problem herein will be addressed using neural networks [3], for which transfer learning has been proven quite efficient (although usually in deeper learning architectures [7, 8]). Due to the layered structure of the networks, after having created a model for a task, transferring a part of it (e.g. some subset of the layers) is easy. The method is used in many disciplines, such as computer vision [4, 5]. The most commonly-used learners are Convolutional Neural Networks (CNNs), which can be very slow to train and may need a lot of data, which in many cases can be hard to obtain (e.g. labelled images). These problems can be dealt with by using the fixed initial layers of pre-trained models to extract features of images, and then train only the last layers to classify in the new context. In this way, both the number of trainable parameters and the need for huge datasets and computation time are reduced. Another topic that transfer learning has been used in is natural language processing (NLP) [6], where the same issues of lack of labelled data and large amounts of training time are dealt with by transferring of pre-trained models into new tasks. Further examples of the benefits of transfer learning can be found in web document classification [7, 8]; in these cases, in newly-created web sites, lack of labelled data occurs. To address this problem, even though the new web sites belong to a different domain than the training domain of the existing sites, the same models can be used to help classify documents in the new websites. In the context of the current work, transfer learning is considered in transferring knowledge from one sub-problem to the other by introducing pre-trained layers into new classifiers. The classification problem that will be presented is related to damage class/location. A model trained to predict a subset of the damage classes (source task) with data corresponding of that subset (source domain), will be used to boost performance of a second classifier trained to identify a different subset of damage states. 5.2 Problem Description Similar to the aforementioned applications, in SHM machine learning is also used for classification and regression. In data driven SHM one tries to identify features that will reveal whether a structure is damaged or what type of damage is present and so, labelled data are necessity. Therefore, in SHM applications lack of labelled data about damage location or severity is a drawback. SHM problems can be categorised in many ways but are often broken down according to the hierarchical structure proposed by Rytter [9]: 1. Is there damage in the system (existence)? 2. Where is the damage in the system (location)?
RkJQdWJsaXNoZXIy MTMzNzEzMQ==