<p>The identification of crack propagation processes and microcrack types in surrounding rocks is of great importance for ensuring the safety and stability of underground engineering during subterranean resource extraction and underground space utilization. In this study, a deep learning framework integrating moment tensor (MT) theory and convolutional neural networks (CNN), termed MT-CNN, is proposed for automatic microcrack classification. Acoustic emission (AE) waveform-based representations are used as model inputs, while the corresponding crack types determined by MT inversion serve as classification labels. A series of uniaxial compression experiments was conducted to establish the AE database. Subsequently, three types of AE signal representations and four CNN architectures, including MobileNetV2, InceptionV3, ResNet50, and VGG16, were employed to construct and evaluate the MT-CNN models in order to identify the optimal framework. By analyzing the learning curves and classification performance, the results indicate that the VGG16 model combined with continuous wavelet transform (CWT) time–frequency representations achieves the best performance, with a test accuracy of 84.02%. Furthermore, a multi-metric evaluation framework incorporating accuracy, ROC curve, F1-score, Cohen’s kappa coefficient, and Hamming loss was introduced for comprehensive model assessment. The results demonstrate that the CWT-VGG16 model achieves the highest overall score of 0.82, confirming its superiority among all tested configurations.</p>