← Back to Sessions

Fully Automated DAS Signal Denoising Using Weakly Supervised Machine Learning and Spliced Optical Fibers

Description: 

We present a weakly supervised machine learning method for suppressing strong random noise in distributed acoustic sensing (DAS) data. The method aims to map random noise processes to a chosen summary statistic, such as the distribution mean, median or mode, whilst retaining the true underlying signal. This is achieved by splicing (joining together) two fibers hosted within a single optical cable, recording two noisy copies of the same underlying signal corrupted by different realizations of random observational noise. A deep learning model can then be trained using only these two noisy copies of the data to produce a completely denoised copy. We use a dataset from a DAS array deployed on the surface of the Rutford Ice Stream in Antarctica. Despite being a very low anthropogenic noise environment, strong random noise processes heavily dominate the signal from microseismic icequake events. We demonstrate that the proposed method greatly suppresses incoherent noise and enhances the signal-to-noise ratios (SNR) of these microseismic events, enhancing the performance of subsequent processing steps, such as event detection. We further demonstrate that, following training, this approach is more efficient and effective than standard frequency filter routines and a comparable self-supervised learning method, known as jDAS. Our preferred model for this task is extremely lightweight (three hidden layers, 47,330 model parameters), processing 30 secs data recorded at a sample frequency of 1000 Hz over 985 channels (~ 1 km of fiber) in < 1 sec. The model is trained in a ‘weakly supervised’ manner, such that it requires no manually-produced labels (i.e., pre-determined examples of clean event signals or sections of noise) for training, and doesn’t require prior assumptions on the distribution of the noise or event signals, other than that they are independent. Due to the inherently high noise levels in DAS recordings, efficient data-driven denoising methods, such as the one presented, will prove essential to time-critical DAS detection and early warning processing workflows, particularly in the case of microseismic monitoring.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 05:30 PM (local time)

Presenting Author: Maximilian J. Werner

Student Presenter: No

Invited Presentation: 

Authors