This paper is copyrighted by ASNT. For reprints, please contact ASNT at http://www.asnt.org/ or at The American Society for Nondestructive Testing, Inc., 1711 Arlingate Lane, Columbus, OH 43228-0518, USA.

Harvey L. Bodine, Robert L. Dawes, Steven S. Henley

Martingale Research Corporation
mrcinfo@martingale-research.com

Eric v. K. Hill

Embry-Riddle Aeronautical University
600 S. Clyde Morris Blvd
Daytona Beach, FL 32114-3900
(904) 226 - 6760
hille@db.erau.edu

Abstract

Under the auspices of a National Science Foundation Phase I Small Business Innovation Research (SBIR) grant, an advanced neural network architecture and specialized signal processing techniques were successfully applied to the problem of automatically detecting and classifying acoustic emissions embedded in a noisy data stream. Our approach for the collection of known acoustic data was to construct a set of test articles suitable for the production of acoustic emissions. The pressure vessels were cylindrical, made of aluminum, and contained rivets as part of their structure in order to produce rivet fretting and rubbing noises that would clutter the crack growth data. We collected the acoustic emission events created during test-to-destruction of the vessels, and archived the data along with test-specific data for each acoustic event. The AE events were created by pressure cycling each vessel from 0-80 psi with a sinusoidal pressure variation, in order to produce a combination of crack growth, rivet fretting, and crack-face rubbing events. We then performed detection and classification processing using the Parametric Avalanche Stochastic Filter (PASF) neural network to assess the feasibility of this method for automatically detecting and classifying the AE events. Additionally, we obtained tensile specimen data from a previous unrelated research effort and used this data to provide much of the initial testing and evaluation of our classification algorithms.

The results of this project demonstrated that applying speech recognition techniques in conjunction with the PASF neural network provided effective detection and classification of acoustic events for both metal pressure vessels and metal tensile specimens. We achieved crack growth classification rates of up to 96% for tensile data and 93% for pressure vessel data. An important note is that these results were achieved by applying continuous input data without any external information about the acoustic emission event, such as the time of arrival or event location.

Back to NDT Page
Back to Publications