A Training Experiment on the Cost and Benefit of Training U-Net for Laboratory Signal Processing

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

U-Net is used by seismologists for picking phases of seismic waveforms. Applying U-net to process laboratory acoustic emission (AE) signals is useful for lab-scale experiments. In this paper, we document experiments on training two different U-Nets with a dataset containing 6532 manual pickings. This is to prepare for optimizing (or challenging) the maximum phase picking accuracy in the next step. Two label options (i.e., manual picking versus AIC auto-picking), two mask options, and two loss function options are considered in the training experiments. We observe the accuracy of the U-Nets, the shape of the output probability string, and the training and validation loss curves. The training time for U-Net-Two is shorter than that of U-Net-One. The binary-cross-entropy loss function is, as expected, proper for the initial stage of training. Applying the alternative option of the loss function may be useful for fine-tuning the model. We find that the length of the non-zero probability points in the label should be adjusted according to the frequency of the waveform. Shifting the signals and labels simultaneously will not affect the training. In the future this "shifting treatment" will be used for data augmentation. 1. INTRODUCTION While originally developed for biomedical imaging segmentation [1], U-Net is useful for many other tasks that require classification at the data-point-scale. In recent years, the application of U-Net for picking different phases of seismic waveforms has received broad attention [2]. Such a technique for processing seismic waveforms has immediate application to laboratory AE signal processing. However, training a U-Net for processing laboratory AE signals faces many challenges. First, environmental noise in laboratory AE signals can be catalog based, so the usefulness of conducting transfer learning using a well-trained U-Net is limited. Second, the training of a U-Net demands a large amount of manually labeled P-wave arrivals. Previously successful training of a U-Net for seismic waveform processing utilized an extensive data pool consisting of over 700,000 manually labelled P and S arrival times picked from over 30 years of earthquake recordings provided by the Northern California Earthquake Data Center [2].