Abstract:
Accurate identification of microseismic events is the cornerstone of coal mine disaster early warning and geological safety protection. However, the blasting vibrations and strong noise interference in the complex underground environment pose severe challenges to effectively identifying microseismic signals. To address this issue, a dual-branch network M2-Mamba based on multimodal feature extraction is proposed, aiming to achieve high-accuracy recognition of microseismic signals. This method constructs one-dimensional and two-dimensional time-frequency spectrum maps of signals through fast Fourier transform and continuous wavelet transform respectively, and co-learns the essential characteristics of signals in the frequency and time-frequency domains through two different branches. The one-dimensional spectrum branch uses a convolutional neural network (CNN) to extract frequency features, while the two-dimensional time-frequency branch uses Mamba to capture the dynamic time-frequency evolution patterns of signals. In the two-dimensional branch, a hybrid vision Mamba block (LF-Mamba) is designed, and a progressive frequency decoupling strategy is used to model the low-frequency global structure and high-frequency local details of the time-frequency spectrogram respectively: only the low-frequency components of the time-frequency spectrum are input into the Mamba, allowing it to focus on integrating the overall contour of the signal and time-varying modes; High-frequency components are processed by the cumulative heavy parameterized convolution module DenRepDW-3, enhancing the model's ability to extract time-frequency textures and local details. At the same time, to address the energy distribution differences in time and frequency dimensions between micro-vibration and burst signals, a partitional cross-selective scanning mechanism is proposed to enhance Mamba's directional modeling capabilities for key time-frequency features of the two types of signals. Finally, a multi-loss function joint optimization strategy is introduced to strengthen feature learning capabilities for different modal branches under the coordinated constraints of fusion and auxiliary losses. Experimental results on real coal mine datasets show that the classification accuracy of M2-Mamba reaches 97.56%, significantly outperforming the comparative method and verifying its effectiveness in coal mine microseismic signal recognition tasks.