Feature extraction and visualization for damage detection on adhesive joints, utilizing lamb waves and supervised machine learning algorithms

Abstract

In recent years, the aeronautical industry has grown with the adoption of new materials for structural applications, and adhesives emerged as an alternative to standard methods to join parts. This has sparked the use of nondestructive testing and structural health monitoring methods, namely, those based on Lamb waves to assess the integrity of structures and detect damage. Methods using time-series sensor data and machine learning algorithms have shown great promise in classifying the extent of damage in plate-like structures. Despite this, robust methods are still missing for choosing relevant features of Lamb waves that optimize the learning process to classify damage. In this paper, a powerful time-series specialized feature extraction method is implemented to detect and classify weak adhesion, meaning a zero-volume defect denoting any intermediate level of adhesion between complete adhesion and kissing bond. Initially, over 75 different types of features, with varying internal parameters, are extracted from raw data. Then, using the Benjamini-Hochberg procedure, some of these features are selected as relevant for the damage classification problem. After the initial selection, the features are handled with machine learning techniques, namely the Naïve Bayes and random forest classifiers, which not only lead to high classification metrics using all features, but also reveal and isolate those features that yield the best differentiation between damage categories. The selection methodology accounts for robustness by utilizing different layers of selection and classification, validating the feature relevance in relation to the appropriate set of classes. As such, different damage types and ranges can be adopted in the proposed multi-class classification pipeline.