Improved Method for Acoustic Emission Source Location in Rocks Without Prior Information

Source location is a fundamental problem in acoustic emission (AE) monitoring. It usually requires configuring two pieces of prior information: wave velocity structure in the medium and arrival-time picking. The primary objective of this study was to propose an approach that would not require prior information to locate the AE source in rocks in which prior information is difficult to obtain. Pencil-lead break (PLB) tests were implemented on a granite cube to collect AE full waveforms and corresponding PLB coordinates. A deep learning model, an fully-connected network (FCN)-based asymmetric auto-encoder, was proposed to map the relationship between the raw AE full waveform and the PLB location. This location was represented by a spatial Gaussian probability distribution. The trained deep learning model was employed to predict the PLB locations using AE waveforms that were not involved in model training. The proposed model achieved the prediction accuracy for AE source of 0.65, whose location errors are much less than the conventional location approach with prior information configuration. Moreover, we implemented an occlusion experiment for demonstrating that the deep learning model automatically recognized the valuable waveform segment from the raw AE signal. Even though the proposed model’s performance had room for improvement in terms of the quantity and diversity of training data, the results had verified that the model is successful in locating the AE source, especially regarding those mediums such as rocks in which it is difficult to obtain prior information.