为达到掘进过程中掘进机可实时预测前方地质状况的目的，以实际工程案例的岩渣图片数据为基础，分析传统神经网络模型的结构特征，自主设计轻量化网络模型，并对岩渣图片进行训练，使用岩渣图像数据调整Mobilenet轻量化网络模型参数。结果表明： 1）传统的神经网络算法参数量多，占用内存量大，无法在现场部署算力有限的岩渣识别边缘终端，而自主设计的轻量化网络可以满足现场岩渣图像分类准确性的要求； 2）Mobilenet轻量化网络模型比传统网络模型减少了80%~90%的计算量，但岩渣图像分类准确性达到97%。

In order to predict the front geological conditions in real time during boring, the structural characteristics of traditional neural network model are analyzed based on the slag pictures of practical projects and an autonomous lightweight network model is designed. After training with slag pictures, the parameters of Mobilenet lightweight network model is adjusted with the data. The results show that: (1) With excess parameters and memory, the traditional neural network algorithm cannot deploy the edge terminal of slag recognition onsite, while the selfdesigned lightweight network can meet the accuracy requirements of onsite slag image classification. (2) The classification accuracy of Mobilenet lightweight network model reaches 97%, but its calculation amount is only 80%~90% of the amount of traditional neural network model.