Abstract: Phase identification is the most fundamental and important task in seismology and is crucial to in-depth analyses of regional seismic activity and the precise detection of underground structures. The considerable growth in earthquake observation data and the rapid advances in machine learning technology have meant that various machine learning-based seismic phase recognition methods have emerged, such as Eqtransformer and Phasenet, which have significantly promoted the development of this field. However, these models often face challenges when processing short-period dense array data, such as high missed detection rates for small microseismic events and the incomplete waveform recognition of medium-to-strong earthquakes. To address these issues, this study developed a microseismic phase recognition model that uses channel and multi-head attention mechanisms. It effectively integrates coarse- and fine-grained features through residual connections, optimizes feature representation using a channel attention mechanism, and enhances the ability to focus on key information through a multihead attention mechanism. Training data are derived from multiple sources and have high confidence-to-noise ratios; therefore this study adopted a strategic data augmentation method to construct noisy (including Gaussian random white noise and STEAD dataset noise) and denoised datasets in a 6:4 ratio to improve the generalization ability of the model in complex field noise environments. The experimental results show that compared to existing models, such as Eqtransformer and Phasenet, the proposed model significantly improves the recognition rate for small microseismic events and optimizes the recognition accuracy of P-wave and S-wave arrival times. In addition, the model has a fast recognition speed and can complete the comprehensive recognition of one month of data from a single station within ten minutes, which demonstrates that it can efficiently process big data. This study successfully constructed an efficient and accurate microseismic phase recognition model by introducing advanced attention mechanisms and deep-learning architectures. This model considerably improves the ability to identify small and micro earthquakes, enhances model generalization, accelerates the recognition process, and provides a powerful tool for micro earthquake detection.