基于长短期记忆网络的TBM掘进预测模型及围岩等级对预测精度的影响

A LSTM-based model for TBM performance prediction and the effect of rock mass grade on prediction accuracy

全断面隧道掘进机(TBM)施工过程可以采集大量丰富数据,这使得建立数据驱动的TBM掘进预测模型成为可能。为探究围岩等级信息对模型预测精度的影响,以TBM掘进循环上升段数据为主要输入,基于深度学习中的长短期记忆(LSTM)网络,建立两种考虑围岩等级信息的、一种不考虑围岩等级信息的TBM稳定段推力、扭矩预测模型。保持其他条件相同,对三种模型进行训练、测试。结果表明,在由于围岩等级分布不均匀而导致的训练样本数较少的情况下,可以通过精确的掘进循环参数提取与数据增强,使三种模型预测推力、扭矩时均能达到较高的精度(误差分别在8%、14%以内)。两种引入围岩等级信息作为输入参数的方法并不能明显提高模型的预测精度。三种模型中,以围岩等级信息作为LSTM网络初始状态的模型,其预测精度以不到1%的优势领先其他模型的预测精度。该优势主要由于该模型能够较好地预测V级围岩中的掘进循环稳定段参数。

The massive data collected by a working Tunnel Boring Machine(TBM)make data-driven TBM performance prediction models possible.To explore the effect of surrounding rock mass grade on model prediction accuracy,given the rising stage data of a TBM advancing cycle,three long short-term memory(LSTM)models were established to predict the thrust and torque in the stable stage.Two models included surrounding rock mass grade as auxiliary inputs,and one did not.The prediction performances of the three models were compared.The results show that the three models can achieve high accuracy when predicting thrust and torque(the error is within 8%,14%)through delicate cycle parameter extraction and data augmentation,even though the number of training samples is limited due to the imbalanced distribution of rock mass grades.The two methods of introducing surrounding rock mass grade as auxiliary inputs can not significantly improve the prediction accuracy of the models.Among the three models,the prediction accuracy of the LSTM model initialized by rock mass grade is less than 1%higher than that of other models.This superiority is mainly because the model can predict the parameters in the stable stage more accurately in class V surrounding rock.