Acoustic multiplets detection based on DBSCAN and cross-correlation
Théotime de la Selle, Jérôme Weiss, Stéphanie Deschanel
Abstract
Non-destructive detection of fatigue crack propagation in industrial parts remains nowadays a key challenge in various engineering fields. Acoustic emission (AE) signals specific to incremental fatigue crack growth can be detected, cycle after cycle, as precursors to final fatigue rupture. These so-called acoustic multiplets are characterized by strongly similar waveforms, triggered at almost the same load during the fatigue cycle, and arising from the same source. Detecting such multiplets provides information about the crack growth process and for industrial parts in service, allows an early warning of potential failure. We developed a method based on a density-based data clustering algorithm (DBSCAN) working with a dissimilarity metric derived from the cross-correlation of AE waveforms to automatically classify acoustic multiplets in fatigue and other fields. Automatized processes described here allow to use the algorithm both on laboratory and industrial fatigue cases, and are designed to work in-operando. Our methodology is tested on AE signals recorded during different laboratory fatigue tests. This demonstrates the robustness of the algorithm to detect different multiplets for different materials, test conditions, specimen geometries, or acoustic sensors.