基于深度学习的滚刀磨损预测模型对比

Comparison of Deep Learning-Based Models for Predicting Cutter Wear

摘要:

盾构开挖过程中的滚刀磨损影响因素极其复杂，常规的预测公式中系数给定的参考值实用性较差，预测精度较低，考虑采用机器学习方法预测滚刀磨损量。依托广州地铁某线路区间硬岩地层情况，基于滚刀磨损量的实测数据，采用文献调研分析滚刀磨损的影响因素，优选15种特征作为输入参数，预测参数为滚刀磨损量，经过数据处理共得到4 514条数据序列作为输入，采用深度学习的方法，包括卷积神经网络（CNN）、长短期记忆网络（LSTM）和转换器（Transformer）模型，与传统的多层感知机（MLP）模型进行了对比。结果表明，深度学习模型能够很好地预测滚刀的磨损趋势，其中，CNN模型在测试集每环磨损量上的相关系数为0.987 2，均方根误差为0.013 8，表现优于其他的模型。CNN模型在累计开仓区间上的相关系数为0.996 3，均方根误差为0.210 8，平均误差百分比只有1个开仓段超出+20 %区间，效果优于传统的基于神经网络滚刀磨损量预测的方法。所提出的基于CNN滚刀磨损预测模型显著提高了预测精度，能够准确预测滚刀磨损量，为开仓换刀提供了参考。

Abstract:

Cutter wear during shield tunneling excavation is influenced by a wide range of complex factors， making accurate prediction challenging. Traditional coefficient-based formulas often lack practical applicability， resulting in low prediction accuracy. To address this issue， this paper explores the use of machine learning methods to predict cutter wear， focusing on hard rock layers encountered in a specific section of the Guangzhou Metro. Based on measured field data， 15 influential features were selected as input parameters， with cutter wear amount as the target output. After data preprocessing， a total of 4，514 data sequences were obtained for model training and evaluation. Several deep learning models—including convolutional neural networks （CNN）， long short-term memory networks （LSTM）， and Transformer models—were compared with a traditional multi-layer perceptron （MLP） model. The results indicate that deep learning models effectively captured the wear trend of the cutter. In particular， the CNN model achieved the best performance， with a correlation coefficient of 0.987 2 and a root mean square error （RMSE） of 0.013 8 for wear per ring on the test set. For cumulative cutter wear over the opening range， the CNN model attained a correlation coefficient of 0.996 3 and an RMSE of 0.210 8. Only one segment exceeded a +20 % average error， demonstrating the model’s robustness. Overall， the CNN-based prediction model significantly outperformed traditional neural network approaches， offering enhanced accuracy and reliability. This method provides a valuable reference for timely cutter replacement decisions during shield tunneling， contributing to more efficient and safer excavation operations.