基于改进VGG16网络的TBM围岩识别方法研究

Research on TBM Surrounding Rock Identification Method Based on Improved VGG16 Network

根据隧道掘进机(TBM)工程对前方地质条件的超前预警需求,开展围岩识别算法研究。针对原始VGG16网络结构待定参数过多以及知识蒸馏训练模式准确率不足的问题,提出一种基于改进VGG16网络的围岩识别方法。基于原始VGG16网络结构,优化其分类层结构,并减少网络复杂度,大幅降低网络待定参数量。基于传统知识蒸馏训练模式,优化其训练逻辑,提升网络对目标任务的特征提取能力。采用某隧道掘进工程的岩渣图像数据集,对上述方法进行验证。试验结果表明,该方法可在小幅提升准确率的同时,大幅减少网络的待确定权值参数。综上所述,该方法创新性地同时改进模型结构和训练模式,更适用于硬岩掘进条件下的围岩识别任务。

According to the demand for advance warning of forward geological conditions in tunnel boring machine(TBM) engineering, research on the surrounding rock identification algorithm is carried out. To address the problems of too many pending parameters in the original VGG16 network structure and insufficient accuracy of the knowledge distillation training mode, an improved VGG16 network-based surrounding rock identification method is proposed. Based on the original VGG16 network structure, the classification layer structure is optimized, and the complexity of the network is reduced to significantly reduce the number of pending parameters. Based on the traditional knowledge distillation training mode, the training logic is optimized to improve the feature extraction ability of the network for the target task. A rock slag image dataset from a tunnel boring project is used to validate the above method. The experimental results show that the method can significantly reduce the parameters of the network to be determined weights while slightly improving the accuracy rate. In summary, the method innovatively improves both the model structure and training mode and is more suitable for the task of surrounding rock identification under hard rock excavation conditions.