Experimental And Theoretical Study On The Robustification Of Acoustic Emission Inspection With Recurrent Neural Networks

This paper presents a summary of an experimental and theoretical study conducted to improve the performance of Acoustic Emission (AE) based structural health monitoring. Two broadband AE sensors are attached to a test specimen and a series of artificial test signals were generated using multiple pencil-lead breaks. The geometry of the test specimen is such that; one of the AE sensors always receives a direct wave and the second sensor receives reflection (or echo) of the same direct wave received at the first sensor. Time series data of the direct and reflected waves, recorded at a rate of 5 MHz, are used to train a multi-layer recurrent neural network (RNN) in a supervised setting. The inherent patterns learned from the training data are used to classify received acoustic emission waveforms into reflection and direct waves. Preliminary results show that the proposed technique, at least in a similar setting and material, improves the performance of AE based health monitoring.