Autor(es): Lu, Faming ; Liu, Yi ; Lin, Zedong ; Han, Xiangqi ; Liu, Cong
Data: 2026
Identificador Persistente: http://hdl.handle.net/10362/191498
Origem: Repositório Institucional da UNL
Assunto(s): Fully mechanized mining; Coal and rock identification; Vibration signals; Wavelet transform; Siamese network; Software
Lu, F., Liu, Y., Lin, Z., Han, X., & Liu, C. (2026). SiamWT-CRNet: A Siamese Wavelet Network with Cross-Domain Feature Fusion for Dynamic Coal-Rock Recognition in Top-Coal Caving Systems. Applied Soft Computing, 186, Part A, Article 113984. https://doi.org/10.1016/j.asoc.2025.113984
To address the limitations of traditional deep learning methods due to strong data dependency and insufficient interpretability when recognizing “under-releasing" and “over-releasing" phenomena during top-coal caving in longwall mining, this study proposes an intelligent recognition framework, SiamWT-CRNet, based on joint time-frequency domain analysis of vibration signals. It leverages wavelet-domain Siamese network architecture combined with a cross-wavelet feature enhancement mechanism to achieve high-precision dynamic identification of the coal-rock interface. It introduces a cross-scale feature fusion strategy based on multi-family wavelet bases, constructing a physically interpretable enhanced feature space through heterogeneous wavelet decomposition. A lightweight Siamese wavelet convolution module, ECWT, is designed to integrate recursive wavelet decomposition with an improved attention mechanism, enabling focused extraction of critical frequency-band features while reducing parameter complexity. Furthermore, a cross-wavelet contrastive learning paradigm is adopted, where a dual-branch network is employed to mine the intrinsic differential features of coal and rock vibration signals. This is coupled with a hard-voting classifier to achieve efficient decision-making. Experimental results demonstrate that the proposed method significantly outperforms traditional models in terms of recognition robustness under strong noise interference. Moreover, the decision-making mechanism has been validated through frequency-domain interpretability analysis, aligning well with engineering expertise.
