MS-Mamba: A Lightweight State-Space Model for Microseismic Signal Identification

The reliable identification of microseismic events is an important prerequisite for microseismic monitoring. However, the complex underground mining environment and limited monitoring resources make it challenging for existing identification algorithms to balance accuracy and efficiency. Therefore, this paper proposes a lightweight microseismic signal classification model, MS-Mamba, based on time–frequency image analysis. This model efficiently integrates global time–frequency dependencies and local detail features through Lightweight Receptive Field Feature Interaction (LRFFI) and uses the embedded DB-Mamba to enhance the modeling of temporal and frequency-domain features. The inverted residual module is also introduced to complement local high-dimensional feature representations, thereby boosting classification performance without substantial computational overhead. To accommodate different computational budgets, three model variants of different scales are designed. Experimental results show that MS-Mamba achieves a favorable balance between accuracy and efficiency across different scales. Among them, MS-Mamba-B achieves the highest accuracy of 98.66%; MS-Mamba-T still reaches 97.94% with only 0.13 GFLOPs of computational cost. These results indicate that MS-Mamba achieves a favorable trade-off between classification accuracy and computational efficiency.