为了解决通过室内试验评价渣土改良效果存在耗时长、成本高，无法满足目前隧道掘进中的预测需求等问题，基于深度学习，引入遗传算法GA（genetic algorithm）对传统BP（back propagation）神经网络重新设计和优化，创建GA-BP模型；选择均方根误差GMSE、平均绝对误差MAE和决定系数R2作为评价模型预测效果的研究指标；
利用机器学习中支持向量机算法与随机森林算法的预测结果与GA-BP模型的预测结果进行对比。试验结果表明： 1）基于深度学习的传统BP模型和本文创建的GA-BP模型的各项评价指标皆高于机器学习算法； 2）相较于传统BP网络的预测结果，GA-BP模型对内摩擦角、渗透系数和坍落度预测的最高相对误差分别降低了7.18%、5.02%、1.17%，且GA-BP模型的RMSE、MAE和R2值都优于传统BP模型和机器学习算法的预测结果。由此可得，基于深度学习的神经网络比机器学习算法更能提取到数据之间的关联性，且经过遗传算法优化后得到的GA-BP模型提高了传统BP模型的预测准确度。
The existing indoor tests on soil conditioning effect
have many disadvantages such as long time and high cost, as well as cannot meet
prediction needs in tunnel excavation. In this paper, based on deep learning,
genetic algorithm (GA) is introduced to redesign and optimize the traditional
back propagation(BP) neural network to create a GABP
model. The root mean square error(RMSE), the mean absolute error(MAE), and the
determinable coefficients 〖WT〗R2 〖WT５《TNR》〗are selected as the research indicators to evaluate the prediction
effect of the model. Finally, the prediction results of the support vector
machine algorithm and the random forest algorithm in machine learning are
compared with those of the GABP model. The
experimental results show the following: (1) Both the traditional BP model
based on deep learning and the GABP model created have
higher evaluation indexes than the machine learning algorithm. (2) Compared
with the prediction results of the traditional BP network, the highest relative
errors of the GABP model for the prediction of
internal friction angle, permeability coefficient, and slump are reduced by
7.18%, 5.02%, and 1.17% respectively. (3) The RMSE, MAE, and values of
the GABP model are better than the prediction results
of the traditional BP model and the machine learning algorithm. It can be
concluded that deep learningbased neural networks are
better at extracting correlations between data than machine learning
algorithms, and the GABP model obtained after
optimization by genetic algorithms improves the prediction accuracy of
traditional BP models.
