DB-MS-CAN: A robust deep learning framework for microseismic velocity inversion via Dual-Branch Attention and Multi-Scale Feature Aggregation

Abstract

Accurate real-time assessment of underground stress fields is critical for mine safety, yet conventional inversion methods struggle to balance precision with computational efficiency. To address these challenges, we propose the Dual-Branch Multi-Scale Convolutional Attention Network (DB-MS-CAN), which employs a decoupling-and-aggregation strategy. The framework integrates a Dual-Branch Fusion Module (BFM) to isolate sparse transient signals from persistent mining noise and a Multi-Scale Dense Feature Aggregation (MDFA) module to reconstruct complex geological structures across varying spatial scales. To ensure rigorous validation, the model was evaluated on both geologically constrained real-world datasets and an independent synthetic benchmark governed by wave propagation physics. Results demonstrate that DB-MS-CAN achieves a structural similarity index (SSIM) of 0.925 and an root mean square error (RMSE) reduction of approximately 40% compared to traditional ray-based tomography. The proposed model achieved significant performance improvements compared to advanced baselines, including CNN-LSTM and Vision Transformers. Notably, the model maintains high fidelity (correlation coefficient > 0.90) under extreme 0 dB SNR conditions and achieves an inference speed of ~7 seconds per event on a single GPU, outperforming iterative solvers by orders of magnitude. This framework provides a robust and efficient solution for dynamic hazard early warning in deep mining environments.

Keywords

Microseismic inversion / Deep learning / Dual-Branch Attention / Multi-Scale Feature Aggregation

References

[1] Feng F, Chen S, Zhao X, Li D, Wang X, Cui J. Effects of external dynamic disturbances and structural plane on rock fracturing around deep underground cavern. Int J Coal Sci Technol. 2022; 9(1): 15. doi: 10.1007/s40789—022—00487—z

[2] Tang Z, Liu X, Xu Q, Li C, Qin P. Stability evaluation of deep—buried TBM construction tunnel based on microseismic monitoring technology. Tunnelling and Underground Space Technology. 2018; 81: 512-524. doi: 10.1016/j.tust.2018.08.02

[3] Liu JP, Si YT, Wei DC, Shi HX, Wang R. Developments and prospects of microseismic monitoring technology in underground metal mines in China. J Cent South Univ. 2021; 28(10): 3074-3098. doi: 10.1007/s11771—021—4839—y

[4] Dai R, Wang Y, Zhang D, Ji H. Automatic P—phase—onset—time—picking method of microseismic monitoring signal of underground mine based on noise reduction and multiple detection indexes. Entropy. 2023; 25(10): 1451. doi: 10.3390/e25101451

[5] Aghamiry HS, Gholami A, Operto S, Malcolm A. ADMM—based full—waveform inversion for microseismic imaging. Geophys J Int . 2022; 228(1): 259-274. doi: 10.1093/gji/ggab332

[6] Zhang B, Xia D, Tan J, et al. Acoustic wave equation full waveform inversion based on optimal transport and L2 norm hybrid weighting. J Ocean Univ China. Published online March 18, 2025. doi: 10.1007/s11802—026—6167—1

[7] Zhang S, Li M, Cui G, Mao G. Jiyu xishu zhengzehua FWI he mOWL—QN de weidizhen zhenyuan weizhi yu zhenyuan zibo tongshi fanyan fangfa [A simultaneous inversion method for microseismic source location and source waveform based on sparse regularization FWI and mOWL—QN]. Oil Geophysical Prospecting. 2025; 60(5): 1146-1159. [In Chinese]. doi: 10.13810/j.cnki.issn.1000—7210.20240309

[8] Sun P, Zheng J, Zhao J, Yang Y, Wang Y. Fully automatic differentiation with coupling deep neural networks for full—waveform inversion. J Seismic Explor. 2025; 34(6): 16-28. doi: 10.36922/JSE025410085

[9] Zhang J, Song G, Zhu F, Hu B, Zhang H. Elastic wave CT with multistage inversion process based on improved SIRT. Phys Lett A . 2024; 522: 129739. doi: 10.1016/j.physleta.2024.129739

[10] Cai S. Zhu beidong shuangyuan weizhen CT fanyan suanfa yanjiu ji yingyong [Research and application of the active—passive dual—source microseismic CT inversion algorithm] [master’s thesis]. Xuzhou, China: China University of Mining and Technology; 2024. [In Chinese]. doi: 10.27623/d.cnki.gzkyu.2024.002106

[11] Tian Z, Pei S, Yu Q. Inversion imaging research of cross hole seismic CT in complex Karst areas: Guiyang Metro Line 3 under construction presented as an example. CT Theory and Applications. 2024; 33(5): 561-567. [In Chinese]. doi: 10.15953/j.ctta.2023.132

[12] Yuan C, Zhang J. Vertical slowness—constrained joint anisotropic parameters and event—location inversion for downhole microseismic monitoring. Remote Sensing . 2025; 17(3): 529. doi: 10.3390/rs17030529

[13] Olya BAM, Mohebian R, Moradzadeh A. A new approach for seismic inversion with GAN algorithm. J Seismic Explor. 2024; 33(3): 1-36.

[14] Zhou Y, Han L, Zhang P, Zeng J, Shang X, Huang W. Microseismic data—direct velocity modeling method based on a modified attention U—Net architecture. Applied Sciences . 2023; 13(20): 11166. doi: 10.3390/app132011166

[15] Li L, Zeng X, Pan X, Peng L, Tan Y, Liu J. Microseismic velocity inversion based on deep learning and data augmentation. Applied Sciences. 2024; 14(5): 2194. doi: 10.3390/app14052194

[16] Chang A, Gross L, Hörning S. Full waveform inversion constrained to well log velocity data and seismic data using random mixing. J Appl Geophys. 2025; 243: 105968. doi: 10.1016/j.jappgeo.2025.105968

[17] Gou R, Song D, He X, et al. Microseismic response characteristics and stress anomaly zoning in deep outburst—prone coal seams with multi—concealed structures: A case study. J Appl Geophys. 2025; 242: 105932. doi: 10.1016/j.jappgeo.2025.105932

[18] Ma XP. Dadi dianci he yuanzhen zoushi shuju sanwei lianhe fanyan yanjiu [Research on three—dimensional joint inversion of magnetotelluric and teleseismic arrival time data] [PhD dissertation]. Changchun, China: Jilin University; 2025. [In Chinese]. doi: 10.27162/d.cnki.gjlin.2025.000803

[19] Wamriew D, Charara M, Pissarenko D. Joint event location and velocity model update in real—time for downhole microseismic monitoring: A deep learning approach. Comput Geosci. 2022; 158: 104965. doi: 10.1016/j.cageo.2021.104965

[20] Cao D, Su Y, Cui R. Multi—parameter pre—stack seismic inversion based on deep learning with sparse reflection coefficient constraints. J Pet Sci Eng. 2022; 209: 109836. doi: 10.1016/j.petrol.2021.109836

[21] Song C, Lu M, Li Y, et al. Spatial pattern learning: dip structure constraint multi—view convolutional neural network for pre—stacked seismic inversion. IEEE Trans Geosci Remote Sens. 2023; 61: 1-16. doi: 10.1109/TGRS.2023.3307897

[22] Zou L, Jia J, Ye Y, Wu B. STA—Fault3D: a lightweight 3D seismic fault detection network based on spatial—temporal asymmetric convolution set. Applied Sciences. 2025; 15(22): 12153. doi: 10.3390/app152212153

[23] Ou L, Huang L, Zhao Y, Wang Z, Shen H, Li X. Intelligent phase picking of microseismic signals based on ResUNet in underground engineering. J Cent South Univ. 2025; 32(9): 3314-3335. [In Chinese]. doi: 10.1007/s11771—025—6077—1

[24] Ma B, Han L, Liu W, Wu Z, Li C. A deep learning velocity modeling method based on a novel attention mechanism networks. J Seismic Explor. 2024; 33(5): 1-14.

[25] Yang R. Convolutional neural network for image classification research — based on VGG16. In: Proceedings of the 2024 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML); November 22—24, 2024; Shenzhen, China. IEEE; 2024: 213-217. doi: 10.1109/ICICML63543.2024.10958042

[26] Li S, Xie B. Boosting UNet performance via VGG—based encoder for medical image segmentation. In: Proceedings of the 2025 6th International Conference on Computer Engineering and Application (ICCEA). April 25—27, 2025; Hangzhou, China. IEEE; 2025: 1964-1968. doi: 10.1109/ICCEA65460.2025.11102997

[27] Tang W, Sun J, Wang S, Zhang Y. Review of AlexNet for Medical Image Classification. EAI Endorsed Trans e—Learn. 2023; 9. doi: 10.4108/eetel.4389

[28] Wei K, Liu H, Cao H, Chen M. Deep learning inversion research on frequency—domain borehole—to—surface electromagnetic method based on CNN—LSTM network. Comput Tech Geophys Geochem Explor . 2025. [In Chinese]. doi: 10.12474/wthtjs.20250226—0001

[29] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2017). July 21—26, 2017; Honolulu, Hawaii, USA. IEEE; 2017: 2261-2269. doi: 10.1109/CVPR.2017.243

[30] Liu L, Song W, Zeng C, Yang X. Microseismic event detection and classification based on convolutional neural network. J Appl Geophys. 2021; 192: 104380. doi: 10.1016/j.jappgeo.2021.104380

[31] Libby SA. Modelling faults as fracture swarms to introduce multi—scale heterogeneity. In: Proceedings of the 58th U.S. Rock Mechanics/Geomechanics Symposium . June 23—26, 2024; Golden, Colorado, USA. American Rock Mechanics Association; 2024. doi: 10.56952/ARMA—2024—0934

[32] Chen F, Liang Z, Cao A. ConvLSTM for predicting short—term spatiotemporal distribution of seismic risk induced by large—scale coal mining. Nat Resour Res. 2023; 32(3): 1459-1479. doi: 10.1007/s11053—023—10193—5

[33] Luo H, Xu X, Pan Y, Yu J, Zhang Y, Zhang L. The CGAS deep learning algorithm for P—wave arrival time picking of mining microseismic events. IEEE Access. 2023; 11: 102961-102970. doi: 10.1109/ACCESS.2023.3317084

[34] He Z, Jia M, Wang L. UACNet: a universal automatic classification network for microseismic signals regardless of waveform size and sampling rate. Eng Appl Artif Intell. 2023; 126: 107088. doi: 10.1016/j.engappai.2023.107088

[35] Xu H, Li K, Shan P, et al. Automatic picking method for the first arrival time of microseismic signals based on fractal theory and feature fusion. Fractal and Fractional. 2025; 9(11): 679. doi: 10.3390/fractalfract9110679

[36] Yang Y, Wang H, Li Y, Birnie CE, Alkhalifah T. Automatic microseismic event location using deep neural networks in anisotropic media. In: Proceedings of the 83rd EAGE Annual Conference & Exhibition. June 6—9, 2022; Madrid, Spain. European Association of Geoscientists & Engineers; 2022: 1-5. doi: 10.3997/2214—4609.202210621

[37] Aguiar AC, Myers SC. Microseismic event relocation based on PageRank linkage at the Newberry Volcano geothermal site. Bull Seismol Soc Am. 2018; 108(6): 3656-3667. doi: 10.1785/0120180115

[38] Antolik M, Ekstrom G, Dziewonski AM, Boschi L, Kustowski B, Pan J. Multi—resolution global models for teleseismic and regional event location. In: Proceedings of the 23rd Seismic Research Review: Worldwide Monitoring of Nuclear Explosions. October 2—5, 2001; Jackson Hole, Wyoming, USA. National Nuclear Security Administration; 2001: 136-146.

[39] Yang L. Theoretical analysis of Adam optimizer in the presence of gradient skewness. Int J Appl Sci. 2024; 7(2): 27-42. doi: 10.30560/ijas.v7n2p27

[40] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM. 2017; 60(6): 84-90. doi: 10.1145/3065386

[41] Bazulin M, Sabitov D, Charara M. Determination of the elastic parameters of a VTI medium from sonic logging data using deep learning. Comput Geosci. 2021; 152: 104759. doi: 10.1016/j.cageo.2021.104759

[42] Li W, Liu H, Wu T, Huo S. A high—resolution velocity inversion method based on attention convolutional neural network. IEEE Trans Geosci Remote Sens . 2023; 61: 1-14. doi: 10.1109/TGRS.2023.3311788

[43] Cao W, Guo XB, Tian F, et al. Seismic velocity inversion based on CNN—LSTM fusion deep neural network. Appl Geophys. 2021; 18(4): 499-514. doi: 10.1007/s11770—021—0913—3

[44] Sun QH, Zong ZY, Li X. Probabilistic seismic inversion based on physics—guided deep mixture density network. Pet Sci. 2024; 21(3): 1611-1631. doi: 10.1016/j.petsci.2023.12.015

[45] Lu J, Wu C, Huang J, Li G, Yuan S. Multi—receptive field distillation network for seismic velocity model building. Eng Appl Artif Intell. 2024; 133: 108547. doi: 10.1016/j.engappai.2024.108547