Back
Science

Sichuan University Researchers Propose Framework for Cross-Domain Microseismic Signal Classification

Generated on: Last updated: 1 sources
View source

New Framework Aims to Improve Early Warnings for Underground Disasters

A research team from Sichuan University has developed a new computational framework designed to classify microseismic signals across different engineering environments. The system is intended for use in early warning systems for deep underground disasters, aiming to address challenges related to limited computational resources, on-site noise, and a lack of labeled data in new monitoring locations.

The Challenge: Classifying Signals in Harsh Conditions

Microseismic monitoring is a real-time, three-dimensional technique used to provide early warnings for disasters such as rock bursts in deep underground engineering projects, including tunnels and hydropower stations. A core technical challenge is the automatic classification of signals—distinguishing between blast events, microseismic events, and noise.

The researchers identified three primary obstacles to deploying such classification systems:

  • Limited computational and storage resources at on-site monitoring stations.
  • Severe on-site noise interference that degrades signal quality.
  • Insufficient labeled data available during the early stages of monitoring a new project.

To address these issues, the team developed the Lightweight and Robust Entropy-regularized Unsupervised Domain Adaptation Framework (LRE-UDAF). Its primary function is to transfer knowledge from a labeled "source domain" (e.g., an existing tunnel project) to an unlabeled "target domain" (e.g., a new underground powerhouse), improving classification accuracy where training data is scarce.

How the Framework Works

The LRE-UDAF consists of two integrated core components:

1. A Lightweight Feature Extractor
This component processes raw signal data. It combines an Improved ShuffleNet Unit (ISNU), designed for computational efficiency, with a Dual Attention Adaptive Residual Shrinkage Block (DAARSB), which is intended to suppress noise in the signals.

2. An Unsupervised Domain Adaptation Module
This module facilitates knowledge transfer. It employs a method called bi-classifier adversarial learning, utilizing a metric termed classifier determinacy disparity (CDD) along with entropy regularization. This approach aims to align the feature distributions between the source and target domains so the model performs accurately on new, unlabeled data.

Key Experimental Results

The study's findings are based on experiments using single-channel acceleration waveform data.

Source-Domain Performance:

  • Using 30,000 labeled waveforms from a tunnel project in southwest China, the feature extractor component alone achieved a classification accuracy of up to 97.7% for blast, microseismic, and noise signals.
  • The feature extractor contains only 0.155 million parameters, making it lightweight.
  • For microseismic signals with a signal-to-noise ratio (SNR) in the range of [10, 15), the model maintained an accuracy of 85%.

Cross-Domain Transfer Performance:
The full LRE-UDAF framework was tested on transferring knowledge between datasets from different projects or monitoring systems:

  • Transferring from a tunnel dataset monitored with SINOSEISM systems to data from the same tunnel monitored with ESG systems increased classification accuracy from 77.7% to 94.3%.
  • Transferring from the tunnel dataset to an ESG-monitored hydropower underground powerhouse dataset increased accuracy from 87.6% to 97.3%.

Publication and Future Work

The paper, titled "Lightweight and Robust Cross-Domain Microseismic Signal Classification Framework with Bi-Classifier Adversarial Learning," was published in the journal Engineering. The authors are Dingran Song, Feng Dai, Yi Liu, Hao Tan, and Mingdong Wei.

The researchers noted that ablation studies confirmed the importance of the ISNU and DAARSB components in the feature extractor, as well as the effectiveness of the CDD and entropy regularization in the adaptation module.

The study lists potential directions for future optimization, including dynamic hyperparameter tuning, extension to multi-channel signal processing, and enhancing the interpretability of the domain alignment process.

The full text of the paper is available via open access at: https://doi.org/10.1016/j.eng.2025.10.023