Machine Learning Based Sensing and Data Processing Strategies for Next Generation Structural Health Monitoring

Abstract Prioritization of infrastructure repairs suggests a need to collect data from structures, which contain condition information over an extended period of time. As capable sensing devices become more abundant and advanced, they provide an opportunity to collect massive datasets and integrate traditional means with emerging data types obtained by smartphones or image-based sensing systems. Nevertheless, the information in structural datasets is relatively low, has unpredictable content due to anomalies and uncertainty exist in data. Therefore, traditional structural health monitoring (SHM) techniques create scalability and efficiency problems and turn out to be inadequate while working with large structural datasets. This dissertation provides data-centric techniques specialized for next-generation structural health monitoring which considers the challenges of the structures and monitoring systems. In this respect, integrated frameworks based on machine learning and deep learning, as well as innovative sensing and data collection techniques are proposed to perform a sustainable, cost-effective, automated and reliable condition assessment of structural systems. This dissertation first introduces a real-time, multi-task deep learning-based scheme for damage diagnosis to provide efficient learning, shorter training time and lower computation cost while exploiting the large data volumes. The sensitivity of the algorithm to data uncertainty is evaluated on the numerical models and the series of tests using multiple Digital Image Correlation (DIC) systems and different damage setups. Another way to deal with big data in SHM is reducing the transported, processed and stored data. This dissertation introduces a scalable and robust compressed sensing approach which takes advantage of data sparsity} The proposed method classifies the state of the structure by processing a subset of the dataset. Another challenge with SHM is that the data is generated rather expensively. This dissertation addresses it by proposing a deep learning-based data collection strategy which accurately predicts the expensive and laboriously impractical sensor data from inexpensive data sources. A novel training strategy is also presented to overcome the challenges in training of long sequences generated to assess fatigue of the structures. Finally, a finite-difference inspired convolutional neural network framework is introduced to learn inherent data characteristics}from the given data and iteratively estimate the future dynamical behavior by using only a few trainable parameters.