Abstract: Mine microseismic signals usually have characteristics such as unstable waveform variations and weak events being easily masked by noise. Existing deep learning methods lack information interaction between local and global features, and it is difficult to distinguish real microseismic events from noise by relying only on features at a single scale or a single level, resulting in low identification accuracy, which makes it difficult to meet the requirements of rock burst early warning. To address the above problems, a mine microseismic identification network based on feature optimization and collaborative learning was proposed. The network performed feature optimization on microseismic waveform images through a multi-scale convolution module and a hybrid module to extract fine-grained semantic features. Information interaction between subnetworks at different levels was established through collaborative learning, so that the local features extracted by shallow subnetworks and the global features obtained by deep subnetworks were deeply integrated, thereby enhancing the network's ability to identify microseismic signals. The experimental results showed that: ① The multi-scale convolution module and collaborative learning of multi-layer subnetworks had a positive effect on improving network performance. ② Compared with networks such as ResNet-18, EfficientNet, BeiT, CaiT, and DeiT, the proposed network achieved the highest accuracy, precision, and F₁ score. ③ The proposed network showed more concentrated attention to key feature regions of microseismic waveform images and achieved significant identification performance.