Evaluating the Data Augmentation on Acoustic Emission (AE) Signals for Training a Deep Neural Network

Abstract

ABSTRACT: Previous manual picking experiments and U-Net training experiments have illustrated the limitations of training a U-Net for laboratory acoustic emission (AE) signal processing. The overall goal of this series of studies is to devise an approach for training a U-Net without excessive demand for manual picking. In this paper, we evaluate two data augmentation approaches: multiplier and shifting. The multiplier aims to mimics weak signals, while the shifting is deemed physically acceptable. Our evaluations indicate that both data augmentation approaches create sensitive augmented datasets for the current model. These findings suggest the necessity of model retraining, and further evaluation of the multiplier approach after model retraining. This practice also provides a more comprehensive understanding of the development circle of deep-learning models for the current AE signal processing task. Additionally, we suggest an approach of transfer learning by utilizing the deeper blocks of a well-trained U-Net for seismic waveform processing. 1. INTRODUCTION The application of U-Net in seismic waveform processing in recent years has received broad recognition [1]. Such U-Net models predict the seismic phase arrivals (i.e., P-wave and S-wave arrivals) by outputting strings of probabilities. The point of the highest probability above a threshold will be identified as the arrival time of the corresponding phase. For laboratory acoustic emission (AE) signal processing, deep-learning models with similar phase picking functions are demanded. However, many technical issues need to be tested before a successful application of the U-Net model in laboratory AE signal processing. First, the data structure and the environmental noise for field-scale seismic waveforms and laboratory AE signals are different. Seismic waveforms recorded by the accelerometer and seismometer contain three waveform strings for the three directions/components, whereas laboratory AE signals record only one waveform string. In most conditions, this AE waveform is for, i.e., primarily sensitive to, the sensor's normal direction. Furthermore, the environmental noise of the AE signals can be catalog-dependent, and would be significantly different from test to test. These facts set the limit for directly applying the well-trained U-Net of seismic waveform processing into laboratory AE signal processing via transfer learning. Second, the training of U-Net demands an extensive amount of manually labeled P-wave and S-wave arrivals, while acquiring such labels can be time-consuming and challenging in the laboratory AE signal processing. For instance, the successful training of the U-Net for seismic waveform processing has utilized over 700,000 analyst-labelled P and S arrival times [1].. Accumulating such data pool takes over 30 years. While for laboratory AE signal processing, by consuming the effort of two students for a full semester, only a few thousand manually labelled P-wave arrivals have been obtained [2]; and yet we still have difficulties in obtaining S-wave arrivals with good accuracy [3].