Simulation-Based Classification for Structural Health Monitoring; A Parametrized Component Model-Order-Reduction Approach
Abstract
We present a Simulation-Based Classification approach for Structural Health Monitoring which takes advantage of recent advances in parameterized Model Order Reduction (pMOR) to inexpensively form quasi-exhaustive synthetic classifier training datasets. We also provide a framework for error analysis which can be exploited either to assess experimental risk - a measure of misclassification - or to decrease experimental risk. We present the approach through a cradle-to-grave example: a physical microtruss with nominal (undamaged) and damaged states, a shaker-actuated experiment with strobo-scope/camera displacement data acquisition, a Helmholtz elastodynamics best-knowledge" (bk) model, physically motivated features with good discrimination properties, Reduced Basis pMOR for rapid prediction of bk model features, offline construction of an extensive synthetic training dataset, and application of several state-of-the-art machine learning algorithms - at least one of which performs very well on both synthetic and real (experimental) data. We then identify several important tasks for further enhancement of the approach: a component-based pMOR technique which can treat very high-dimensional parametrizations of damage as required forlarge-scale systems; techniques for automatic feature extraction; exploration of time-domain features and impulsive excitation as well as ambient (operational) loading for passive classification; enhancement of classifier training with offline experimental data; and approaches to anomaly detection.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712077
Entities
People
- Anthony Patera
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy