Open Access

ARTICLE

Dual-Stream Deep Learning for Health Monitoring of HDPE Geomembranes in Landfill Containment Systems

Yuhao Zhang^1,2,3, Peiqiang Zhao^1,2, Xing Chen^1,2, Shaoxuan Zhang⁴, Xinglin Zhang^1,2,*

1 Institute of Solid Waste and Hazardous Chemicals, Gansu Academy of Eco-environmental Sciences, Lanzhou, 730000, China
2 Key Laboratory of Recycling and Control of Industrial Waste in Gansu Province, Lanzhou, 730000, China
3 Zhengdong Space Industry Co., Ltd., 5 South Road of Agriculture Exhibition Hall, Beijing, 100020, China
4 School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou, 730070, China

* Corresponding Author: Xinglin Zhang. Email:

(This article belongs to the Special Issue: Advanced Detection Technologies and Interpretable Machine Learning Methods in Energy Infrastructure)

Structural Durability & Health Monitoring 2025, 19(5), 1343-1365. https://doi.org/10.32604/sdhm.2025.066558

Received 11 April 2025; Accepted 03 June 2025; Issue published 05 September 2025

Abstract

The structural integrity monitoring of high-density polyethylene (HDPE) geomembranes in landfill containment systems presents a critical engineering challenge due to the material’s vulnerability to mechanical degradation and the complex vibration propagation characteristics in large-scale installations. This study proposes a dual-stream deep learning framework that synergistically integrates raw vibration signal analysis with physics-guided feature extraction to achieve precise rupture detection and localization. The methodology employs a hierarchical neural architecture comprising two parallel branches: a 1D convolutional network processing raw accelerometer signals to capture multi-scale temporal patterns, and a physics-informed branch extracting material-specific resonance features through continuous wavelet transform (CWT) and energy ratio quantification. A novel gated attention mechanism dynamically fuses these heterogeneous modalities, adaptively weighting their contributions based on localized signal characteristics—prioritizing high-frequency transient features near damage zones while emphasizing physics-derived energy anomalies in intact regions. Spatial correlations among distributed sensors are modeled via graph convolutional networks (GCNs) that incorporate geometric topology and vibration transmission dynamics, enabling robust anomaly propagation analysis.

---

Keywords

---