Ultrasonic Guided Waves for Metal Plate Structures

This joint project with Prof. Nico Declercq aims at developing a smart embeddable sensor for the detection of defects (mainly corrosion and cracks) in metal-plate structures using guided waves, opto-acoustics and machine learning. By the end of the project, the resulting system will be expected to be integrated onto a mobile tetherless robotic system, hence allowing structural inspection of ship hulls and storage tanks with unprecedented resolution. To this day, the individual technologies have been demonstrated on large-scale systems that cannot be deployed on an autonomous mobile system because of both their weight, volume and energy requirements, and the complexity of the signal interpretation. Bringing them together into a small-sized tetherless smart system is precisely the purpose of this research. This raises significant challenges requiring an integrated view of multiple fields of research.

Ultrasonic probes for non-destructive inspection of materials have now reached a sufficient maturity level to allow their regular deployment in industrial applications, a process known as Structural Health Inspection (SHM). In general, to detect defects on metal plates, one can simply measure the time of traversal of the ultrasonic signal between the two sides of the plate, or excite a through-thickness standing wave, to get a precise estimate of the metal thickness. This information can then be used to detect thickness-affecting alterations such as corrosion or damages due to impacts. A promising and more sophisticated technique for such envisioned inspection is the use of Ultrasonic Guided Waves (UGW), in particular shear horizontal and Lamb waves. These waves propagate through the entire material thickness while following a direction parallel to the material surface. In the appropriate conditions, and based on the physical effect of acoustic mode conversion, the properties of the propagated spectrum can be used to infer the material state and in particular to detect cracks, corrosion or holes. This requires measuring the transmitted spectrum at multiple points precisely spaced along the wave propagation direction in order to build a space-time spectrum and analyze its variations. The complexity of the analysis of the transmitted signal in practical cases makes the detection and characterization of defects a challenging problem for signal processing and a promising application domain for machine learning techniques such as deep convolutional neural networks (CNN).

One of the main challenges when dealing with CNNs is to build a large enough dataset to train them. Our current work is focusing on using simulation to generate a dataset to train a model, and then refine this model with a moderate number of real-world examples.

Experimental platform

© 2019 – 2021 DREAM Lab – Georgia Tech Lorraine