岩石破裂类型识别是实现煤矿冲击地压灾害预测和预警的重要前提。微震是岩石破裂监测的有效手段之一，但常规的微震信号机器学习方法存在特征提取能力有限，以及受噪声影响导致的准确率不高且泛化性较差等问题。针对上述问题，提出了一种基于微震信号深度特征学习的岩石破裂类型识别方法。首先，通过巴西圆盘劈裂试验和直剪试验分别获取张拉型破裂微震信号和剪切型破裂微震信号，并将其时频谱图、Log−Mel频谱图和梅尔频率倒谱系数合并构造微震信号聚合（MSA）声谱图；然后，通过加入多特征并行密集连接块（MP−DenseBlock）和压缩与激发过渡层（SE−TransLayer）的改进DenseNet（SE−MPDenseNet）对MSA声谱图进行深度特征提取；最后，将提取的特征向量输入至添加Hinge Loss损失函数的改进LightGBM（HBL−LightGBM）进行分类，识别岩石破裂类型。通过真三轴加载试验模拟接近地下工程实际环境中的冲击地压灾害，结果表明，所提方法对于岩石破裂类型识别的准确率达92.12%，且具有较强的特征提取能力和泛化能力。

Accurate identification of rock fracture types is crucial for the prediction and early warning of coal mine rockburst hazards. Microseismic monitoring has been widely used for detecting rock fractures. However, conventional machine learning methods for microseismic signal analysis exhibited limited feature extraction capabilities and were highly susceptible to noise, leading to reduced classification accuracy and poor generalization performance. To address these limitations, this study proposed a novel rock fracture type recognition method based on deep feature learning of microseismic signals. In this study, microseismic signals corresponding to tensile and shear fractures were collected through Brazilian disc splitting and direct shear tests, respectively. These signals were then processed to construct a microseismic signal aggregation (MSA) spectrogram, which integrated time-frequency spectrograms, log-Mel spectrograms, and Mel-frequency cepstral coefficients. To enhance feature extraction efficiency, an improved DenseNet model (SE-MPDenseNet) was developed by incorporating multi-feature parallel dense blocks (MP-DenseBlock) and squeeze-and-excitation transition layers (SE-TransLayer). The extracted deep feature vectors were subsequently fed into an optimized LightGBM classifier (HBL-LightGBM), which was modified with a Hinge Loss function to improve classification performance. To evaluate the effectiveness of the proposed method, a true triaxial loading test was conducted to simulate rockburst hazards under realistic underground engineering conditions. Experimental results demonstrated that the proposed approach achieved a rock fracture type recognition accuracy of 92.12%, significantly outperforming conventional methods in both feature extraction capability and generalization ability. The findings indicate that the proposed method provides a robust and effective framework for microseismic-based rock fracture classification. It offers valuable insights for rockburst hazard monitoring and mitigation in mining and geotechnical engineering.