Open Access

ARTICLE

A Lightweight Manifold-Aware State Space Model for Efficient Seismic Signal Classification at the Edge

Pingan Peng^1,2, Qi Zhang^1,*, Ya Liu¹, Yue Han¹, Zhida Jiang¹, Linli Chen¹, Xuefeng Huo¹

1 School of Resources and Safety Engineering, Central South University, Changsha, China
2 State Key Laboratory for Fine Exploration and Intelligent Development of Coal Resources, China University of Mining and Technology, Xuzhou, China

* Corresponding Author: Qi Zhang. Email:

(This article belongs to the Special Issue: Advances in Artificial Intelligence for Geotechnical Engineering)

Computer Modeling in Engineering & Sciences 2026, 148(2), 24 https://doi.org/10.32604/cmes.2026.083624

Received 07 April 2026; Accepted 30 July 2026; Issue published 28 August 2026

Abstract

Unfilled goafs located beneath urban areas pose a significant threat to surface safety, and microseismic monitoring is an important tool for capturing rock-mass microfractures and supporting early warning. However, under field conditions, existing sequential models face challenges such as high computational complexity and deep feature degradation when deployed on edge devices, primarily due to the non-stationary characteristics of microseismic signals and the presence of complex ambient background noise. To address these issues, this paper proposes the Hierarchical Network with Manifold Hybrid Connection (H-NET-mHC), a lightweight manifold-aware state space model designed for edge computing. The model introduces a content-aware dynamic chunking strategy to adaptively identify informative local segments, effectively suppressing noise and extracting key phases while compressing long sequences. A manifold hybrid connection (mHC) module concurrently constrains feature-flow propagation using doubly stochastic matrices, thereby mitigating nonlinear degradation of high-dimensional features and stabilizing network training dynamics. Multiclass classification and ablation experiments using real-world data from a gypsum mine in Changde demonstrate that, with only 0.35 million parameters, H-NET-mHC achieves an accuracy of 97.77% and a microseismic-event recall of 96.11%. Under severe class imbalance, the two proposed mechanisms produce a complementary recall-oriented effect rather than an across-the-board improvement in all aggregate metrics. The resulting architecture balances computational cost and classification performance, supporting the engineering requirements for intelligent perception and early warning of mine microseismic signals.

---

Keywords

---