为了定量研究复杂裂隙岩石变形破坏规律和裂隙扩展特征，利用裂隙网络模型 3D 打印技术制作了含20条随机节理的类岩石试件，利用数字图像相关方法(DIC)研究了试件破坏过程中应变场演化过程，并分析了每一条裂隙的扩展过程，探讨了动态裂隙对试件整体强度的影响。在此基础上，基于YOLOv5深度学习网络模型，结合DIC云图，提出了一种智能精准识别动态裂隙的算法。研究表明：含复杂裂隙试件破坏过程中往往伴随多条裂隙同时扩展和贯通，试件整体强度与动态裂隙扩展具有重要关系，统计动态裂隙的扩展情况可以半定量的判定试件整体强度。在每一条原生裂隙起裂前，总是首先出现应变集中区域，并且应变集中区域具有前兆性，预示着新裂纹的萌生。原生裂隙的动态演化基本可以分为原生裂隙、应变集中区、新生裂纹和交叉裂隙四种类型，其中，新生裂纹和交叉裂隙对试件整体强度影响最大。提出的智能精准识别动态裂隙算法的精确率、召回率和mAP都在80%以上，且mAP最高达到了91%，GIoU损失参数迭代训练后达到0.01，四种类型裂隙相对应的F1值分别为83%、89%、87%和85%，四种类型裂隙的总体识别精度可达86%。说明该方法在复杂裂隙岩体裂纹识别、定位分类是快速精确有效的。试件在受载时，智能识别算法重点识别并统计新生裂纹和交叉裂隙数量，当新生裂隙和交叉裂隙数量较多时，试件即将破坏，可提前进行预警。本方法可为复杂裂隙岩体破坏的智能监测和破坏预警提供新思路、新方法。

In order to quantitatively study the deformation and failure laws of complex fractured rocks and the charac⁃teristics of fracture expansion,the fracture network model 3D printing technology was used to produce rock⁃like speci⁃mens with 20 random joints,and the digital image correlation method(DIC)was used to study the strain field dur⁃ing the failure process of the specimens. The evolution process is analyzed,and the expansion process of each crack isanalyzed,and the influence of dynamic cracks on the overall strength of the specimen is discussed. Based on theYOLOv5 deep learning network model,combined with the DIC cloud image,an algorithm for intelligent and accurateidentification of dynamic cracks is proposed. Studies have shown that the failure process of specimens with com⁃plex cracks is often accompanied by the expansion and penetration of multiple cracks. The overall strength of the spec⁃imens has an important relationship with the expansion of dynamic cracks. Statistics on the expansion ofdynamic cracks can determine the overall strength of the specimens semi⁃quantitatively. Before each original crackstarts to crack,the strain⁃concentrated area always appears first,and the strain⁃concentrated area has precursory prop⁃erties,which indicates the initiation of new cracks. The dynamic evolution of primary cracks can be basically dividedinto four types: primary cracks, strain concentration zone, new cracks and cross cracks. Among them, crosscracks have the greatest impact on the overall strength of the specimen. The accuracy,recall,and PmA of the proposedintelligent and accurate identification of dynamic fissure algorithms are all above 80%,and the maximum average accu⁃racy mean ( PmA ) is 91%. The GIoU loss parameter reaches 0. 01 after iterative training. The F1 valuescorresponding to the four types of fissures are respectively 83%,89%,87% and 85%,the overall recognition accuracyof the four types of cracks can reach 86%. It shows that this method is fast,accurate and effective in the identification,location and classification of cracks in complex fractured rock masses.