岩溶区灰岩溶蚀程度卷积神经网络识别及敏感性分析  

Recognition and Sensitivity Analysis of Karstic Limestone Dissolution Degree Using Convolutional Neural Network

作　　者：张研 郭道静[2] 张树光 苏国韶 刘锋涛[2] ZHANG Yan;GUO Daojing;ZHANG Shuguang;SU Guoshao;LIU Fengtao(Guangxi Key Laboratory of Geotechnical Mechanics and Engineering,Guilin,Guangxi 541004,China;School of Civil and Architectural Engineering,Guilin University of Technology,Guilin,Guangxi 541004,China;School of Civil and Architectural Engineering,Guangxi University,Nanning,Guangxi 530004,China)

机构地区：[1]广西岩土力学与工程重点实验室,广西桂林541004 [2]桂林理工大学土木与建筑工程学院,广西桂林541004 [3]广西大学土木建筑工程学院,广西南宁530004

出　　处：《应用基础与工程科学学报》2023年第4期961-976,共16页Journal of Basic Science and Engineering

基　　金：国家自然科学基金项目(52068016,42067041);广西自然科学基金项目(2020GXNSFAA159118)。

摘　　要：为了解决岩溶区不同溶蚀程度灰岩合理、高效识别问题,以桂林七星区灰岩为研究对象,开展不同pH、不同循环次数的酸性干湿循环试验,构建不同溶蚀程度灰岩识别的卷积神经网络模型(CNN),分析不同pH值、不同循环次数对模型识别效果的影响,探讨样本数量、网络参数设置对模型影响的敏感性.研究表明,伴随酸液pH值的降低、干湿循环次数的增加,岩样表面溶蚀纹路及溶蚀产生的孔隙越明显,模型分类准确率越高;学习样本、预测样本数量较小时,准确率随着样本数量增加而增高,当学习样本、预测样本数量接近4∶1时,模型预测效果最佳,随后准确率随着样本数量增加而降低;模型对不同网络参数敏感性不同,学习率为0.1,迭代次数与样本更新数为50时,准确率最高.CNN模型预测准确率最高为97.6%,为岩溶区灰岩溶蚀程度有效识别提供一条新途径.To reasonably and efficiently recognize dissolution degrees of karstic limestones,a corresponding convolution neural network model(CNN)model was established based on experiments on a limestone quarried from Qixing District,Guilin.In the experiments,the specimens were treated with dry-wet cycles in acidic environments considering different pH values and different numbers of the dry-wet cycles.The effects of pH values and dry-wet cycle numbers on the recognition accuracy of the model were analyzed.How the data sample number and the network-related parameter settings affect the model was also discussed.The results show that the dissolution-induced patterns and pores on specimen surfaces become more obvious and the recognition accuracy becomes higher with a decrease in the pH value and an increase in the dry-wet cycle number.When the number of learning and prediction data samples is low,the recognition accuracy positively correlates with the sample number.When the ratio of the learning sample number to the prediction sample number is close to 4∶1,the model presents the highest recognition accuracy.For a relatively large number of data samples,the accuracy decreases with the increase in the sample number.The recognition accuracy shows different sensitivities to the network parameters.When the learning rate is 0.1 and the number of iterations and sample updates is 50,the accuracy reaches the peak.The proposed CNN model provides a new way to effectively recognize dissolution degrees of karstic limestones,and the recognition accuracy may reach 97.6%.

关 键 词：岩溶区灰岩 酸性干湿循环 溶蚀程度 卷积神经网络 图像识别 敏感性分析 

分 类 号：TU45[建筑科学—岩土工程]