Abstract: At present, mine microseismic signal recognition still faces problems such as low signal-to-noise ratio, complex background interference and insufficient recognition accuracy in practical applications, which makes it difficult to meet the needs of real-time monitoring and early warning of mine disasters. Existing deep learning-based microseismic recognition algorithms are prone to inadequate feature extraction and poor model generalization when facing limited sample sizes and significant differences in signal characteristics. To address the above issues, a mine microseismic recognition algorithm (MMI-Net) based on collaborative learning and feature optimization is proposed. The algorithm is designed with feature optimization and cross-layer collaborative learning as the core: it adopts multi-scale convolution blocks to extract phase details and local energy distributions under different time scales, and introduces a hybrid module combining convolution and attention fusion to realize the complementarity of local features and global semantic information. Meanwhile, it achieves knowledge sharing and semantic feedback through cross-layer collaborative learning of multi-layer sub-networks, enhancing the model's robustness under complex backgrounds. Experimental results show that: ① Compared with mainstream deep learning models (ResNet, EfficientNet, DeiT, CaiT, etc.), MMI-Net achieves 98.14%, 98.18%, 98.06% and 98.16% in accuracy, precision, recall and F1-score respectively, while the computational complexity is only 16.24 GFLOPs, maintaining high computational efficiency. ② Ablation experiments indicate that multi-scale convolution blocks, channel adaptive scaling layers and cross-layer collaborative learning all make significant contributions to the algorithm's performance; the accuracy drops to 96.34%～97.85% after removing any module. ③ Under typical working conditions such as low signal-to-noise ratio, strong noise interference, sparse signals and feature overlap, MMI-Net still maintains stable recognition performance, with excellent robustness and generalization ability, providing reliable technical support for the intelligent recognition of mine microseismic events and disaster early warning.