基于优化的多输出神经网络预测软弱土压缩和回弹指数研究  

Prediction of the compression index and swell index of soft soils via an optimized multiple-output neural network

作　　者：陈凯 林军 聂利青 段伟 CHEN Kai;LIN Jun;NIE Liqing;DUAN Wei(Anhui Provincial Key Laboratory of Intelligent Underground Detection,Anhui Jianzhu University,Hefei 230601,China;School of Architecture and Civil Engineering,Jiangsu Open University,Nanjing 210036,China;College of Civil Engineering,Taiyuan University of Technology,Taiyuan 030024,China)

机构地区：[1]安徽建筑大学智能地下探测安徽省重点实验室,合肥230601 [2]江苏开放大学建筑工程学院,南京210036 [3]太原理工大学土木工程学院,太原030024

出　　处：《地质科技通报》2025年第2期130-145,共16页Bulletin of Geological Science and Technology

基　　金：国家自然科学基金项目(52308355,51908250);2023江苏高校“青蓝工程”项目;安徽省智能地下探测技术研究院开放课题(2022B1)。

摘　　要：软弱土的压缩指数C_(c)和回弹指数C_(s)是计算土体沉降和回弹的重要参数,采用机器学习算法可高效预测C_(c)和C_(s),减少室内试验周期和费用。引入孔压静力触探(CPTU)原位测试数据,利用土类指数I_(c)量化土层信息,融合室内试验和原位测试数据,改进遗传算法优化的BP神经网络(GA-BPNN),实现多输出功能,同时预测C_(c)和C_(s)。通过相关性分析,确定多输出GA-BPNN模型输入参数,利用TC 304标准场地数据库,将预测结果与多输出BPNN模型、单输出GA-BPNN模型比较,进而验证多输出GA-BPNN模型能效,并获得预训练模型参数。在南京有限场地数据条件下,进一步讨论多输出GA-BPNN模型的优越性,分析预训练、原位测试数据对模型效果的影响,最后进行敏感性分析。结果表明,利用标准场地数据获得预训练多输出GA-BPNN模型,在有限数据条件下,可有效预测C_(c)和C_(s);加入原位测试数据的的GA-BPNN模型预测C_(c)(R^(2)=0.96)和C_(s)(R^(2)=0.97)精确度较高,预测结果更加接近实测值,预测结果相关性与已有研究保持一致。通过预训练的多输出GA-BPNN模型,可在有限场地数据条件下,快速准确预测软弱土的C_(c)和C_(s),对工程实践中的多元参数预测具有良好的应用前景。[Objective]The compression index C_(c)and swell index C_(s)of soil are critical parameters for calculating soil settlement and swelling.Utilizing machine learning algorithms to predict these indices quickly and efficiently can significantly reduce testing duration and costs.[Methods]In this study,we introduce Piezocone Penetration Test(CPTU)in-situ data and quantify soil layer information using the Soil Behaviour Type(SBT)index Ic.We then combine laboratory data with CPTU data to develop a multi-output genetic algorithm-optimized backpropagation neural network(GA-BPNN)model.The input parameters for the multi-output GA-BPNN model were determined through correlation analysis.Using the TC304 standard site database,the prediction results from the multi-output GA-BPNN model were compared with those from the multi-output BPNN model and the single-output GA-BPNN model,verifying the effectiveness of the multi-output GA-BPNN model and obtaining pre-trained model parameters.For sites with limited data in Nanjing,the superiority of the multi-output BPNN model was further evaluated by analyzing the impact of pre-training and in-situ test data on model performance.A sensitivity analysis was also conducted to assess the robustness of the model.[Results]The results demonstrate that the pre-trained multi-output GA-BPNN model,derived from standard site data,can effectively predict the compression and swell indices under limited data conditions.When combined with in-situ test data,the multi-output GA-BPNN model exhibits high prediction accuracy for these indices,with predicted values closely matching measured data.The consistency of the predicted results aligns well with existing studies.[Conclusion]The pre-trained multi-output GA-BPNN model can efficiently predict the compression and swell indices of soft soil under limited data conditions.The proposed method shows significant potential for multi-parameter prediction in engineering practice,enhancing the efficiency and reliability of geotechnical engineering assessments.

关 键 词：压缩指数 回弹指数 多输出 优化神经网络 GA-BPNN模型 软弱土 

分 类 号：P642[天文地球—工程地质学]