Active learning for Structural Health Monitoring

The main aim of Structural Health Monitoring (SHM) is the development of diagnostic systems capable of detecting structural degradation in real time, including: damage detection, localisation and quantification, prognosis (remaining life) and detection of performance-related anomalies.
Machine learning techniques have become increasingly important to the field of SHM. Advanced pattern recognition algorithms can now learn complicated relationships from system data, whilst providing a variety of approaches to novelty detection online and in real-time. In the data-based approach to SHM, the absence of data from damaged structures in many cases forces a dependence on novelty detection as a means of diagnosis. Unfortunately, this means that benign variations in the operating or environmental conditions of the structure must be handled very carefully, lest they lead to false alarms.
Unfortunately, complex systems prove costly to test in the laboratory, whilst also having a vast range of potential damage modes. Therefore, retrieving training data to aid wholly supervised methods becomes infeasible; consequently, a high demand for adaptive, semi-supervised learning methods exists within the context of SHM.
Active learning is a form of semi-supervised, or 'query' learning; its main premise is that an algorithm can achieve greater accuracy from fewer training labels if it can choose the data from which it learns. This eliminates the time-consuming process of labelling large volumes of trivial data, as well as significantly reducing the need for extensive sets of training data. Following unsupervised clustering techniques, an active learner then poses queries relating only to unlabelled instances for which it is uncertain - these data points are then identified by an oracle.
With the intention of utilising active learning as an adaptive intelligence in an online environment, the planned work throughout the duration of this PhD aims to provide significant advances in machine learning techniques currently used by the SHM community.
Through collaborations between the University of Sheffield and Vatenfall, the dynamics research group has recently gained access to additional diagnostic data - gathered over a one year period - from the Lillgrund offshore wind farm in Denmark. This large dataset provides an abundance of potential research routes, as well as the further opportunities to test and validate any methods with real data. Analysis will be carried out on a variety of data types and sources, and as a result, the techniques developed throughout this PhD can be applied to a range of real-world problems in the Structural Health Monitoring domain.