A research team from Sichuan University has proposed a lightweight and robust entropy-regularized unsupervised domain adaptation framework (LRE-UDAF) for cross-domain microseismic (MS) signal classification, addressing key challenges in the practical deployment of MS signal identification for deep underground engineering disaster early warning. Published in Engineering, the study presents a solution tailored to resource-constrained, noise-prone and data-scarce underground environments, with the framework integrating a lightweight feature extractor and a novel unsupervised domain adaptation (UDA) module to enable accurate and efficient MS signal classification.
MS monitoring is a critical real-time 3D technique for early warnings of deep underground disasters, yet automatic signal classification faces three core hurdles: limited deployment resources, severe on-site noise interference and insufficient labeled data in early monitoring stages. Conventional deep learning methods often lack lightweight design for harsh underground conditions and struggle with domain shift caused by varying geological conditions and monitoring equipment. To tackle these issues, the LRE-UDAF is designed with two core components: a lightweight and robust feature extractor, and a UDA module based on bi-classifier adversarial learning with entropy regularization.
The feature extractor combines an improved ShuffleNet unit (ISNU) and a dual attention adaptive residual shrinkage block (DAARSB). The ISNU integrates squeeze-and-excitation blocks to enhance computational efficiency and feature representational capacity, while the DAARSB adopts an improved threshold function and dual-attention mechanism for adaptive noise suppression, effectively removing redundant noise-related information from MS signals. Source-domain experiments using 30 000 labeled single-channel acceleration waveforms from a southwest China tunnel project show the feature extractor achieves a classification accuracy of up to 97.7% for blast, MS and noise signals, outperforming mainstream models such as EfficientNet and MobileNet-V2 in both lightweight performance—with only 0.155 million parameters—and noise robustness, reaching 85% accuracy for MS signals with signal-to-noise ratio in the [10,15) interval.
The UDA module leverages a bi-classifier disparity metric named classifier determinacy disparity (CDD) and entropy regularization, conducting three-stage adversarial learning to align source and target domain feature distributions, enabling knowledge transfer from labeled source data to unlabeled target data. Cross-domain experiments across two practical engineering projects with distinct data distributions validate the framework's effectiveness: transferring from the SINOSEISM-monitored tunnel dataset to the ESG-monitored same-tunnel dataset raises classification accuracy from 77.7% to 94.3%, while transfer to an ESG-monitored hydropower underground powerhouse dataset boosts accuracy from 87.6% to 97.3%. Additional cross-domain transfer scenarios all see notable accuracy improvements, confirming the framework's strong generalization ability.
Ablation studies and comparative experiments further verify the indispensable role of ISNU and DAARSB in the feature extractor, as well as the superiority of CDD and entropy regularization in the UDA module over other mainstream UDA methods. The research also notes potential future optimizations, including dynamic hyperparameter tuning, extension to multi-channel signal processing and enhanced interpretability of domain alignment, to further improve the framework's adaptability for practical engineering applications.
The paper "Lightweight and Robust Cross-Domain Microseismic Signal Classification Framework with Bi-Classifier Adversarial Learning," is authored by Dingran Song, Feng Dai, Yi Liu, Hao Tan, Mingdong Wei. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.10.023
