Neural Network Detection of Fatigue Crack Growth in Riveted Joints Using Acoustic Emission

摘要

The purpose of this research was to demonstrate the capability of neural networks to discriminate between individual acoustic emission (AE) signals originating from crack growth and rivet rubbing (or fretting) in aluminum lap joints. AE waveforms were recorded during tensile fatigue cycling of six notched and riveted 7075-T6 specimens using a broadband piezoelectric transducer and a computer interfaced oscilloscope. The source of 1,311 signals was identified based on triggering logic, amplitude relationships, and time of arrival data collected from the broadband transducer and three additional 300 kHz resonant transducers bonded to the specimens. Normalized spectra were projected onto a two-dimensional feature space using a Kohonen self organizing map (SOM). Then 132 crack growth and 137 rivet rubbing spectra were used to train a back-propagation neural network to provide automatic pattern classification. Although there was some overlap between the clusters mapped in the Kohonen feature space, the trained back-propagation neural network was able to classify the remaining 463 crack growth signals with 94 percent accuracy and the 367 rivet rubbing sig-nals with 99 percent accuracy.