Performance evaluation of compressive sensing based lost data recovery using OMP for damage index estimation in ultrasonic SHM

Abstract

Compressive sensing (CS) has been widely explored for data compression and signal recovery in presence of lossy transmission in structural health monitoring (SHM) applications. Discussions of lost data recovery using CS reported in literature are typically limited to acceleration signals obtained from vibration based SHM systems. Moreover these reports limit the study to performance analysis of recovery of signals in time domain, while feasibility of these algorithm on subsequent damage analysis using recovered signals remains unexplored. A systematic evaluation of performance of CS based signal recovery for algorithmic estimation of damage index (DI) in ultrasound SHM systems is important for determining their practicality for automated SHM applications. In this paper, we study the feasibility of DI estimation in ultrasonic guided wave testing of honeycomb composite sandwich structures (HCSS) using signals recovered from lossy sensor recordings. We emulate signal loss by masking the sensor recordings in an experimentally measured dataset comprising of an HCSS panel with two defects (disbond and high density (HD) core) instrumented with eight piezoelectric wafer and employ orthogonal matching pursuit (OMP) based signal recovery algorithm. Our analysis suggests that while OMP-based signal recovery algorithm is a robust and reliable signal recovery technique, producing signal reconstruction errors lesser than 8.4% for data loss as high as 50%, the magnitude error in DI estimation is significant and varies for different signal difference coefficient (SDC) algorithms. We propose alternate SDC definition, SDCPA, computed using peak amplitude of the Hilbert transform (HT), that shows consistently less error than the conventional cumulative-sum-based SDC definition for the HCSS case study. Further we study trends of error in recovery of lossy time domain signals as well as DI computation as a function of data loss parameters, for both random as well as continuous data loss. Our findings indicate that conventional DI computation algorithms for ultrasonic SHM need to be revisited when used in compressive sensing paradigm.