In this paper, we propose a technique for classifying different types of damages in honeycomb composite sandwich structures (HCSS) using guided wave-based structural health monitoring (GW-SHM) systems that can work well even when data loss is lost for chunks of time. Although damage classification is important for deciding the course of action for usage, repair, and replacement, we show that there is overlap in the amplitude characteristics between different types of damages. This problem can be further exasperated when data is lost in chunks due to attenuation, physical damage, electromagnetic interference, and hardware faults. Reliable signal models are not always available for interpolation-based data recovery in such cases. First, we simulated the loss of 10% to 50% samples in experimentally collected data and recovered the signals using orthogonal matching pursuit with an error consistently below 0.1%. Next, we extracted ten features from the recovered signals in both time and frequency domains. We eliminated four features based on correlation analysis to improve the classification performance. We tried multiple classifiers to distinguish between healthy structure and four types of damages: lost film adhesive, teflon release film, core crush, and high density core in HCSS, and obtained perfect classification accuracy with random forest classifier and an optimal feature set based on feature importance. Due to smaller size and computational efficiency, these models are best suited for edge implementation and in situ monitoring. Also, models using these statistical features are portable to other structures as they are independent of material properties.
|