Research On Method Of Petrophysical Facies Identification Based On Deep Learning

The current oil and gas exploration is focused on tight sandstone reservoir,but tight sandstone reservoir has complex pore relationship,poor physical property,heterogeneity and other phenomena,which makes the accuracy of conventional reservoir identification methods such as empirical statistical method in identifying tight sandstone reservoir is not ideal.The results show that the petrophysical facies is the coupling result of sedimentary,diagenetic and tectonic processes.It is used to understand the reservoir based on the formation principle of the reservoir,and is an important factor controlling the distribution of oil,gas and water.Therefore,the study of petrophysical facies identification is helpful to identify high quality tight sandstone reservoirs.In this thesis,the recurrent neural network and convolutional neural network are studied for petrophysical facies identification,and the following work is carried out:Firstly,based on the index selection method of random forest and principal component analysis,the index set of petrophysical facies identification is constructed.After that,it introduces the pretreatment method of well logging data and the background of mine data.Finally,it collects the actual mine data to build the mine data sample and preprocess it.Secondly,considering the timing information of logging data,a petrophysical facies identification model based on recurrent neural network is constructed.The differences of common recurrent neural network,long and short term memory network and bidirectional long and short term memory network models in rock physics were compared.Thirdly,in order to make up for the limitation of time step of long and short term memory network,the hidden information in logging data is further mined,and the petrophysical facies identification model is constructed based on onedimensional convolution and two-dimensional convolution respectively.In addition,an ablation experiment was conducted on the configuration of the one-dimensional convolution model to verify the superiority of the global mean pooling layer combined with onedimensional extended convolution and batch normalization.Fourthly,the one-dimensional convolutional model and the long and short term memory network model are combined into a parallel dual-pathway structure,and the petrophysical facies identification model of the onedimensional convolutional neural network is further optimized.On the one hand,the number of convolutional layers is reduced in the one-dimensional convolutional neural network pathway,and the squeezing-excitation module is tried to be added,in order to increase the feature connections between different channels and enhance the identification accuracy of the model.On the other hand,dimension transpose of the input sample layer tensor in the long and short term memory network pathway is carried out to improve the efficiency of model identification.Then,the attention mechanism is introduced into the pathway to detect the regions that contribute to the petrophysical facies in the logging data,and the classification activation mapping function of the global average pooling layer in the one-dimensional convolutional pathway is combined to realize the classification visualization of the model.Finally,the petrophysical facies identification of blind Wells is realized based on convolutional recurrent parallel network.In order to solve the problem of petrophysical facies identification,this thesis aims to find a high accuracy and high efficiency deep learning method,so as to liberate manpower to a greater extent and improve the exploration and development efficiency of tight sandstone highquality reservoirs.Through experimental analysis,the convolutional recurrent parallel network proposed in this thesis is suitable for the identification of petrophysical facies.The model has high accuracy and low complexity,and can provide classification basis for the relevant workers.It has a good performance in the actual mine data.