基于深度残差LSTM的盾构姿态预测

Prediction of Shield Attitude Using Deep Residual Long Short-Term Memory Model

深度学习模型相比于常规机器学习模型能够更准确地预测盾构姿态,但在增加网络层数以提升性能时,常遇到网络退化问题。为解决此问题,提出基于深度残差LSTM的盾构姿态预测方法。该方法将残差连接融入长短期记忆(LSTM)神经网络,提升深层网络训练的可行性,并可以有效学习盾构时序数据中的长期依赖关系,同时利用贝叶斯优化算法进行超参数调优。依托浙江某盾构工程数据集对所提方法进行验证,以盾尾水平偏移预测为例,深度残差LSTM模型预测的决定系数(R^(2))达到了0.90,平均绝对误差(MAE)为0.76 mm,相较于LSTM模型(R^(2)为0.64,MAE为1.08 mm)和人工神经网络模型(R^(2)为0.68,MAE为1.93 mm),深度残差LSTM模型可以更准确地预测盾构姿态。此外,与LSTM模型相比,深度残差LSTM模型能有效利用更多的网络层(从5层增加到8层),证明了残差连接在防止网络退化、加强盾构数据特征学习能力方面的显著作用。

Traditional machine learning models often fall short in accurately predicting shield tunneling posture because of performance limitations when increasing network depth.To address this issue,a shield posture prediction method that leverages a deep residual long short-term memory(LSTM)model is proposed.This method integrates residual connections into the LSTM framework to address network degradation and enhance the model's ability to learn longterm dependencies in shield tunneling time-series data.Additionally,a Bayesian optimization algorithm is employed to fine-tune hyperparameters and optimize the shield posture prediction model.Validation conducted in a real-world shield tunneling project in Zhejiang demonstrates that the deep residual LSTM model outperforms conventional LSTM and artificial neural network models.Taking the shield tail horizontal deviation prediction as an example,the deep residual LSTM model has a determination coefficient(R^(2))of 0.90 and a mean absolute error(MAE)of 0.76 mm.In comparison,the LSTM model yields an R^(2) of 0.64 and MAE of 1.08 mm,whereas the artificial neural network model shows an R^(2) of 0.68 and MAE of 1.93 mm.Furthermore,compared to the LSTM model,the deep residual LSTM model can effectively utilize more network layers(from 5 to 8 layers),demonstrating the significant role of residual connections in preventing network degradation and improving feature learning.