A Robust Deep Learning Model for Monitoring Subsurface Using DAS Data Focusing on Noise, Interpretability, and Uncertainty

Abstract

Distributed acoustic sensing (DAS) is characterised by high-dimensional, temporal characteristics and high-frequency, continuous data, achieved by converting a fibre-optic cable into a nodal array of sensors. However, DAS data are often affected by noise issues and missing values caused by sensor failure, hardware failure, or environmental disturbances. Many researchers have applied the conventional long short-term memory (LSTM) method due to its ability to process long sequences of data; however, it has struggled with complex, high-spatial data, limited interpretation, and the inability to quantify uncertainty in predictions. Additionally, LSTMs are vulnerable to noisy inputs, which compromises their robustness and trustworthiness in real-time monitoring applications, such as early warning detection systems. To address these limitations, this paper examines the application of bidirectional long short-term memory (Bi-LSTM) networks with an attention mechanism for seismic event detection using DAS data. Due to the high dimensionality and complexity of DAS signals, deep learning models are prone to overfitting, which compromises their performance on unseen data. This study proposes an advanced deep learning framework incorporating regularisation techniques, including dropout, L2 regularisation, and the early stopping method. Dropout mitigates over-reliance on specific features by randomly deactivating neurons during training, while L2 regularisation penalises large weights to promote simpler, more generalisable models, and early stopping prevents overfitting by monitoring validation performance. Furthermore, Bayesian inference is integrated with Monte Carlo dropout to enable uncertainty estimation, allowing the model to provide confidence intervals for detected seismic activity. By combining strong regularisation strategies with uncertainty quantification, the proposed Bi-LSTM framework achieves a balance between accuracy and generalisation, demonstrating its potential as a reliable tool for real-time seismic event detection using complex DAS data.