An efficiently designed CNN-Transformer fusion network for automatic and real-time microseismic signal classification

Abstract

The automatic, rapid, and accurate identification of microseismic (MS) signals is paramount for real-time rockburst hazard early warning. However, the identification robustness and accuracy of current MS signal classification algorithms face significant challenges due to severe noise interference and limited deployment resources in practical engineering applications. In this study, a lightweight and robust CNN-Transformer fusion network, termed MS-LRFormer, is proposed for more precise and real-time MS signal classification. The MS-LRFormer features a hierarchical pyramid structure, enabling multi-scale feature representations to capture structural and semantic information across varying levels of the input data, thereby enhancing its classification robustness and accuracy. Specifically, a vertical stacking strategy is adopted, in which the lightweight CNN modules are employed for early-stage local information extraction and the efficient Transformers are used for capturing long-range dependencies in later stages. The extracted high-level semantic information is ultimately fed into the attention-based SeqPool module for feature compression and signal classification. To comprehensively evaluate its performance, the proposed MS-LRFormer is assessed across five distinct evaluation methods, in which results from the held-out test set demonstrate an impressive classification accuracy of 98.2%. Compared to seven industry-leading deep learning models, MS-LRFormer exhibits superior feature extraction capabilities, a lightweight design, and enhanced robustness. Moreover, the practical on-site application further validates the lightweight and robust performance of the MS-LRFormer, classifying 13,918 field waveform samples in 165 s with an accuracy of 98.9%, confirming its suitability and potential for widespread engineering applications.