Localization of acoustic emission sources in a four-point bending test on a reinforced concrete beam using artificial intelligence

Abstract

In the field of Acoustic Emission analysis, the use of artificial intelligence is becoming increasingly important. Especially for localizing Acoustic Emission events via pattern recognition of time signals, Neural Networks hold great potential. As part of this research, this potential was tested by examining the locatability of Acoustic Emission sources on a reinforced concrete beam as a result of a four-point bending test. Specifically, the experiment involved a loading that increased until fracture of the specimen, while Acoustic Emissions were measured and recorded over 16 Acoustic Emission sensors. A total of 10,343 Acoustic Emission events were captured, of which 4,102 events could be localized three-dimensionally by triangulation. The remaining 6,241 events could not be assigned locations due to either insufficient determinability of the arrival times or too large measurement deviations. In this work, the combination of neural methods from supervised and unsupervised learning was explored. The acquired events were first divided into localized and non-localized signals. The localized events were subsequently grouped into 5, 10, 15, and 20 clusters using the K-means algorithm (unsupervised learning). Furthermore, four convolutional neural networks were trained by using the clusters as classes and the assigned events as samples per class. This approach achieved assignability with an accuracy of 71–94% and allowed the previously non-localizable Acoustic Emission events to be classified and localized by these Convolutional Neural Networks.