Exploring stratigraphic structure and geological lithology is helpful for oil explorers to accurately evaluate the oil-bearing status of oil and gas reservoirs and provide guidance for subsequent drilling and production.At present,in the field of petroleum exploration,lithology identification is still the main method,which is timeconsuming,laborious and easily affected by subjective factors.Therefore,looking for an efficient and accurate lithology identification method is crucial at this stage.Nowadays,deep learning is developing rapidly,but its application in this field is just beginning to emerge.It is of great significance to apply the deep learning method to this field.In this thesis,an intelligent lithology identification system is designed and implemented based on Efficient Net and a series of improvements on neural network.The main contents are as follows:(1)The recognition effects of Resnet50,Xception and Efficient Net-B3 depth neural network on rock data set are studied and compared,and the Efficient Net-B3 network with the best recognition effect is selected for key research.Based on Efficient Net-B3 deep learning network,an intelligent lithology recognition model based on deep learning is constructed;(2)Using transfer learning to leverage the weights of the original neural network to reduce the amount of data required by the model,speed up the convergence of the model and improve its accuracy.(3)Using group normalization instead of batch normalization to solve the problem of accuracy degradation when the batch size is too small due to hardware memory limitation.(4)Cosine annealing learning rate algorithm is used to solve the poor accuracy caused by gradient descent algorithm when training neural network;(5)Cost function is used to solve the low accuracy of small samples when the samples are not balanced;(6)Data enhancement is used to expand the data set to slow down the error,so the overfitting problem is solved,the recognition accuracy is improved,and the generalization of neural network is enhanced.Based on the above improvements,this study established a model suitable for core image classification.In order to objectively describe the performance of the improved model,it used the open data set to test,and took 12 kinds of open core images of olivine,Augite,Hornblende,Biotite,plagioclase,Andalusite,Cross Stone,Garnet,Actinolite,Oolite,Sandstone,Limestone,and so on from China science database as the samples of the evaluation model to do experiments,with the recognition accuracy up to 99.42%and recognition speed to 3236 pieces / min.In this study,we have used Efficient Net-B3 neural network to recognize the core image in the field of lithology recognition to see good results,which provides a new idea and new method to solve lithology recognition. |