Hyperparameter Tuning of an Artificial Neural Network (ANN) for Laboratory Acoustic Emission Source Location

This is the authors' abstract. We don't add key points for this paper.

Acoustic emission (AE) testing, a technique that has been extensively applied in experimental rock mechanics, traditionally requires source location analysis with both pre-processing and post-processing steps. Performing real-time decision-making that relies on in-situ rock fracture information is hindered by this primarily post-testing analysis technique. This paper examines the potential for using an artificial neural network to replicate the algorithm and output near instantaneous raw source locations; through hyperparameter tuning, this raw output can be improved to approach the accuracy of the standard algorithm. On-disk waveform features are extracted by a trained convolutional neural network (CNN) signal processing model and passed to the ANN through an interlayer. This architecture not only provides the potential for parallel processing but also the ability to directly "read out" raw source locations from the on-disk data. The results of this model provide promising source location predictions despite being only a conceptual example of an ANN for this problem. A further developed model may contain a trained CNN for extracting important waveform features; however, this topic exceeds the scope of this paper and will be the theme of future work. 1. INTRODUCTION Acoustic emission (AE) testing has been extensively used for delineating the rock fracture process in a variety of scientific and engineering applications [1-6]. This technique uses piezoelectric AE sensors to detect energy releases from rock fractures. Arrival times of AE signals and sensor coordinates can be passed to an optimization function, based on the time delay of wave propagation, for locating the source of an energy release. Procedures within laboratory AE source location analysis, in general, include steps of pre-processing, source location, and post-processing as depicted in Fig. 1. Such procedures have been well-established and practiced for decades. During pre-processing, signals are recorded onto the disk and the signal processing codes are applied to the recorded signals for extracting the arrival times and other features. The results of pre-processing (i.e. the arrival times of AE signals) are then fed into a source location algorithm. This algorithm first produces raw results of source locations and then refines them by filtering out disqualified locations based on the physical principle [7] or mathematical (i.e., error or residual) feature [8]. After completing the previous two steps, source locations can be further used for visualizing the rock fracture propagation path, source type classification [6, 9-11], clustering/declustering analysis [12, 13], and other advanced analysis [14-17].