Stratigraphic lithology,as the basic research object of petroleum geological characteristics,can accurately reflect the distribution of oil and gas reservoir productivity,indicate the direction of oil and gas exploration,provide reference for similar basin exploration,and is necessary for field logging.Rock classification is the preliminary preparatory work for subsequent reservoir simulation management.Typical reservoir rock type analysis can help to understand the flow law and interaction mechanism of fluid in rock mass.How to characterize and understand underground lithology distribution has always been an important issue in geology.In oil and gas exploration,it is a common method to identify the lithology of underground formation by cuttings logging.Geologists collect cuttings in time according to certain depth intervals and lateness time,carry out systematic observation,identification and description,and establish stratigraphic profiles to understand the lithological changes of underground strata.Accurate identification of lithology is the premise of accurate determination of porosity and oil saturation,and also the basis of reservoir characteristics,reserves calculation and geological modeling.The most traditional method for identification and identification of cuttings is to collect the collected cuttings manually,extract effective information from the photographs and analyze them.Manual identification under a polarizing microscope requires professional personnel to interpret,and it is labor intensive and subject to subjective factors,especially controlled by human visual resolution,so that the information extracted and the conclusions of the analysis are affected.There are errors,which affect the observation and identification results of the reservoir microstructure.And with the gradual popularization of the PDC bit in the oilfield logging process,the collected rock fragments are finely divided and the number is small,which greatly affects the working efficiency and correct rate of the cuttings description.Only the traditional cuttings description method can not meet the current demand for cuttings logging.In view of the current problem of identification of cuttings,this paper conducts an in-depth study on the feature description of cuttings.Considering the great results in the field of image processing and analysis,the deep convolutional network using the advantages of image classification and recognition is proposed.Identify the cuttings image.The main work of this thesis is as follows:Firstly,the traditional digital image feature recognition method of cuttings image is studied.The experimental demonstration research on the commonly used image recognition feature color and texture extraction methods is carried out.It has been found that image recognition is difficult to achieve better results in the case of unstable features and low feature dimensions.After analyzing the difficult problem of cutting debris image and the status quo of image recognition research,it is proposed to use the deep convolutional neural network combined with the metric learning Triplet structure to identify the image to avoid the influence of illumination and background to improve the accuracy of recognition.Triplet metric learning can map the features extracted by the convolutional neural network to the feature metric space,and train the network model parameters to improve the network model by measuring the feature distance to increase the distance between the samples and the feature distance between the similarities.Recognition accuracy.In order to verify the effectiveness of the proposed method,a comparative experiment was designed for the identification methods mentioned in the paper.In the traditional digital image color and texture feature classification and recognition,and convolutional neural network image classification and the proposed convolutional neural network combined with the metric learning Triplet Loss contrast experiment,the proposed method is better for the identification of cuttings images.Several other methods have improved significantly,and the recognition accuracy has increased by 9.8% and 3.8%,respectively.At the same time,several data amplification techniques are studied and analyzed in this paper.Several data amplification schemes are proposed to improve the accuracy of model identification. |