摘要
传统的现场岩心识别与编录主要依赖技术人员的经验,不确定性因素较多,使用手机或相机拍摄自然图像是采集岩心信息最为便捷的方式。通过搜集大量的岩心样本,采用薄片鉴定的方式确定岩心的类型和名称,然后在不同的光照、尺度条件下拍摄岩心图像,形成图像和名称标记相对应的卷积神经网络的训练数据集。为解决数据增强和不同训练批次在不同测试数据集上产生的识别准确率的差异性问题,提出基于多训练模型的岩心联合识别方法,同时采用多个模型对图像进行识别,综合确定识别结果。选择8个数据集对模型进行测试,使用4个模型联合识别的准确率比单模型无数据增强时最大提升20.34%,平均提升9.13%;比单模型有数据增强时最大提升4.41%,平均提升2.75%,对每个测试集的识别准确率均有明显提升,总的识别准确率达91.56%,有效避免了使用单模型识别时对部分数据集识别效果好,而对部分数据集识别效果差的问题。为了在现场快捷使用岩心识别模型,采用TensorFlow Lite框架研发了岩心识别手机APP,通过手机拍摄图像,并进行识别。在河北省保定市博野县地热勘探中的测试结果表明,该APP的现场识别准确率可达85%,较实验室测试时有所降低,说明野外的拍摄环境与岩心状态比实验室测试时更加复杂,不过,其依然可以作为一个辅助工具为现场工作人员提供重要的参考。研究表明,通过选用更复杂的卷积神经网络、不断扩大岩心图像数据集、采用更有效的数据增强方法和策略、建立某个区域的专有岩心识别模型等手段,可以进一步提升岩心图像的识别准确率,为智能钻探的决策提供更有效的信息。
The traditional on-site core identification and recording mainly rely on the experience of technicians,and there are many uncertain factors.Limited by the site conditions,using mobile phones or cameras to capture the natural images is the most convenient way to collect the core information.Therefore,it is necessary to study the feature information extraction technology of core image and apply it to the identification and prediction of core type and other information.Specifically,a large number of core samples were collected,the thin-section identification method was employed to determine the core types and names,and then the core images were taken under different lighting and scale conditions to form the training data sets of convolutional neural networks corresponding to the image and name markers.In order to solve the problem of the different identification accuracy generated by data augmentation and different training batches on different test datasets,a joint core identification method based on multiple training models was proposed,and multiple models were used simultaneously to identify the images so as to comprehensively determine the final identification results.Besides,8 datasets were selected to test the model,and the accuracy of joint identification using multiple models(4 models were used herein)was improved by 20.34%at maximum(9.13%on average)compared to the single model without data augmentation,and 4.41%at maximum(2.75%on average)compared to the single model with data augmentation.In general,the identification accuracy for each test set is significantly improved,with a total identification accuracy of 91.56%.Hence,the proposed method effectively avoids the problem that a single model has the performance good for some datasets but poor for the others during the identification.In order to quickly use the core identification model in field drilling,a core identification mobile APP was developed using the TensorFlow Lite framework,so that the core images could be taken directly with the mobile phone at the scene for identification.The test results of geothermal exploration in Boye County,Baoding City,Hebei Province,show that APP has the field identification accuracy up to 85%,which is lower than that obtained in the laboratory test.This indicates that the shooting environment and core state in the field are more complex than that during laboratory testing,but it can still be used as an auxiliary tool to provide an important reference for field workers.The research shows that the identification accuracy of core images can be further improved by using the more complex convolutional neural networks,expanding the core image datasets,adopting the more effective data augmentation methods and strategies,and establishing a proprietary core identification model for a specific area,so as to provide more effective information for the decision-making of intelligent drilling.
作者
高辉
吴振坤
柯雨
谭松成
何思琪
段隆臣
GAO Hui;WU Zhenkun;KE Yu;TAN Songcheng;HE Siqi;DUAN Longchen(Faculty of Engineering,China University of Geosciences,Wuhan 430074,China;National Center for International Research on Deep Earth Drilling and Resource Development,China University of Geosciences,Wuhan 430074,China;School of Automation,China University of Geosciences,Wuhan 430074,China;School of Future Technology,China University of Geosciences,Wuhan 430074,China)
出处
《煤田地质与勘探》
EI
CAS
CSCD
北大核心
2023年第9期64-71,共8页
Coal Geology & Exploration
基金
国家自然科学基金重点项目(61733016)。
关键词
自然图像
岩心识别
深度学习
卷积神经网络
智能钻探
natural image
core identification
deep learning
convolutional neural network
intelligent drilling
