Improving Long-Term Guided Wave Damage Detection With Measurement Resampling

Abstract

This paper studies strategies for improving long-term guided wave damage detection in an outdoor, uncontrolled environment. We leverage the reconstruction difference between a short-term and long-term principal component analysis to distinguish damage from other variations. While principal component analysis damage detection methods are not new, few are capable of detecting long-term damage in a dynamic environment and in a computationally feasible manner. In this paper, we study the factors that reduce damage detection performance, including the time window duration, the damage duration, and the number of principal components. We then propose a monotonic decreasing sampling strategy that reduces these issues and improves damage detection. Results show that the best receiver operating curve area under the curve score, using an 80-day window of guided wave data and 20-day damage duration, increases from 0.88 to 0.92 while the running time and required memory reduces to less than 1/4 of the original cost.