A multi-scale CNN-Transformer hybrid network for microseismic signal arrival picking: Model analysis and engineering application

Abstract

The automatic, rapid, and accurate picking of P- and S-arrivals is crucial for fully automated processing of microseismic (MS) data. However, complex engineering environment with limited deployment resources and strong noise interference imposes significant challenges on automatic picking methods. To this end, this study proposes a lightweight and robust multi-scale CNN-Transformer hybrid network (MCTH-Net) that utilizes CNNs to extract local information and employs Transformers to capture long-range dependencies. The MCTH-Net employs a hierarchical CNN-Transformer hybrid encoder and a lightweight multi-layer perceptron (MLP) decoder structure. Specifically, the hierarchical encoder alternately stacks Conv and Former blocks to generate multi-scale feature maps. Subsequently, the All-MLP decoder further integrates multi-scale features to produce a more robust segmentation mask for dense prediction tasks. A two-stage dataset synthesis approach and Gaussian label smoothing further enhance its generalization. The evaluation of MCTH-Net on the test set reveals impressive performance, achieving mean absolute errors (MAE) of 1.000 and 1.015 sample points for P- and S-arrivals, respectively. Compared to other industry-leading networks, MCTH-Net excels in both lightweight design and robustness performance. To further validate its practicality, MCTH-Net is applied to real-time P- and S-arrivals picking and MS event localization in practical engineering scenarios. Among the four methods, MCTH-Net demonstrates the most consistent MS source location estimations, with the lowest positioning deviation of 9.34 m. These results highlight MCTH-Net holds promising prospects for engineering applications.