Automated Near-Optimal Feature Extraction Using Genetic Programming with Application to Structural Health Monitoring Problems

摘要

Structural health monitoring systems at their core measure structural response andinfer real-time damage and performance information. Conventional data processinginvolves extraction of low-dimensional features from time series measurements thatare then input to a classification or outlier detection algorithm. Desirable features arehighly sensitive to changes of interest in the structure while also robust in the presenceof noise and varying operational and environmental conditions. Traditional featuredesign requires experts with domain-specific knowledge resulting in an expensive andtime-consuming process. Recently, genetic programming, an evolutionarycomputation method, was adapted to provide automated, data-driven development offeature extraction processes with minimal user input in a supervised learningapproach. This study experimentally validates the genetic programming approach withcomparisons to common feature sets. Demonstrated applications include ultrasonicdamage detection, condition monitoring for rotating machinery, and vibration-basedstructural health monitoring.