洛宁抽水蓄能电站斜井工程作为国内首个大直径大倾角斜井TBM施工项目，采用溜碴槽自溜出碴工艺替代传统皮带运输，可显著降低设备投入与能源消耗，但也存在溜碴状态人工监测调控效率低、水资源浪费和人工成本高等问题。针对上述问题，依托洛宁抽水蓄能电站斜井工程，对大直径大倾角斜井的TBM溜碴状态展开研究，借助高清相机对溜碴槽内的溜碴状态进行实时拍摄，依据拍摄图像内水与碴料的比例建立6个溜碴状态的分类标准（分别为停机溜碴状态、欠水溜碴状态、半欠水溜碴状态、最佳溜碴状态、半富水溜碴状态、富水溜碴状态），借助4种不同架构的神经网络模型（Densenet121、Resnet152、Wide ResNet50_2_L、EfficientNetV2）分别对溜碴状态进行智能分类识别。4种模型采用相同的超参数优化方法，模型比选结果显示： EfficientNetV2表现最佳，测试集的S_Weighted F1和准确率A均为0.883，测试集单张图像预测时间为0.109 s，所提溜碴状态智能识别方法能够满足实际工程对溜碴状态智能识别的实时性和准确性要求。

As the pioneering project in China utilizing TBM construction for such shafts, the Luoning pumped storage power station employs a gravity-driven muck discharge process via chute, replacing traditional belt conveyance. While significantly reducing equipment investment and energy consumption, this method poses challenges including inefficient manual monitoring and control of muck flow status, water wastage, and high labor consumption. To address these issues, high-definition cameras are used to continuously image the state of the muck flow in the chute in real time. Based on the observed water-to-muck ratio within these images, a six-category classification system for muck flow status was established: idle status, water-deficient status, semi-water-deficient status, optimal status, semi-water-rich status, and water-rich status. Four distinct convolutional neural network architectures (DenseNet121, ResNet152, Wide_ResNet50_2_L, and EfficientNetV2) are subsequently employed for intelligent muck flow status classification, all utilizing identical hyperparameter optimization. Model comparison demonstrates that EfficientNetV2 achieves superior performance, attaining a weighted F₁ score and A of 0.883 on the test set, with a per-image prediction time of 0.109 s. This approach effectively fulfills the stringent real-time and accuracy requirements for intelligent muck flow status recognition in practical engineering applications, promoting safer and more efficient automated construction.