RESEARCH PROJECTS

Innumerable engineering applications exist where quantities of interest in a system can only be measured indirectly in space, time, and/or data domain due to the existence and availability of relevant sensing technologies, e.g. monitoring rotating machinery vibration to detect ball bearing wear or analyzing video from a traffic camera to monitor road usage. In all such cases, raw measurements are processed to extract features highly correlated to the desired system measure while insensitive to all other influences. Often feature extraction is performed by projecting onto a basis set such as complex exponentials in Fourier Analysis. In this work, we seek to apply data-based and model-based approaches for identifying optimal minimum-basis sets to problems relevant to structural health monitoring. Our current work follows two branches for developing new damage detection methodologies, image-based correlation from full-field measurements and the adaptation of symbolic dynamics anomaly detection to SHM.

As experimental mechanics techniques continue to improve, dense measurements of strains, stresses, displacements, and other physical quantities become increasingly more feasible and economical. The continuing decrease in electromechanical sensor cost and size as well as the development of optical full-field measurement systems allows for dense measurement sets better represented as intensity images in two or even three dimensions than as a set of scalar point measurements. Image-based representations match the manner in which dense measurements are traditionally visualized in mechanical and civil applications but not analyzed. Many image analysis techniques from computer science, medicine, astronomy, and other fields are available to address problems where a new measurement set is to be compared to a baseline structural condition or model simulation to detect discrepancies indicating structural damage. Moment-based image analysis techniques such as complex moments and Shapelets allow entire image frames to be compared directly and provide significant data compression.

Symbolic dynamics provides a largely unexplored framework for extracting features applicable to multiple damage types and structural platforms. Dynamic systems theory dictates that dynamical systems (continuous-valued in continuous time) have a fully isomorphic (same structure) representation in symbolic dynamics (discrete-time, discrete-alphabet digital streams). Discretization occurs through a process called partitioning, and practical algorithms for partitioning noisy data have been recently developed. The symbolic transformation greatly reduces data-handling requirements while maintaining full data fidelity, and, most importantly, allows the use of sophisticated algorithms from information theory and communications engineering. Thus, utilization of symbolic data and analysis techniques in SHM systems can greatly reduce memory requirements and necessary computational power while admitting the use of low-resolution sensors and potentially even exploit compressive sensing. Symbolic dynamics based analysis is ideally suited for use in SHM network nodes employing wireless data transfer and energy harvesting technologies.

Whether using a 2D orthogonal basis set for image analysis or a discrete valued symbolic alphabet for time series analysis, proper encoding is critical to the success of the feature extraction process. Data driven and statistic model driven approaches allow for the creation of optimal damage sensitive features with robustness to noise and other contaminating factors. Our work aims to integrate and improve on proven techniques while incorporating cutting edge sensing technologies to develop reliable, practical SHM systems for a variety of structural platforms.