Lithology Identification Based On Convolutional Neural Network

Lithology is an important index in oil and gas exploration,and its identification is a crutial part for reservoir evaluation.There is a complex nonlinear relationship between lithology and its related factors.With the development of surveying technology,lithology data present massive characteristics.Therefore,how to establish complex nonlinear relationship from massive data is difficul in lithology identification.Convolutional neural network(CNN)has demonstrated a strong feature extraction capability and as a result,this paper conducts lithology identification based on this method.First of all,due to the high cost of logging,the collected data are generally not enough to accurately describe the lithology.In addition,velocity of shear wave can reflect the features of lithology to some extent.Therefore,this paper introduces the relation for the lithology with different depths of sample points and its different scale information to establish the related image dataset based on the data of well logging and velocity of shear wave from Shengliluo69# well.Specially,to avoid the interference that a lithologic sample contains many types of lithology but uses only single category label,sequence category labels are adopted.Secondly,the identification method based on 2D CNN is proposed.Additionally,to emphasis the characteristics with large contribution,algorithm of attention machine is added to the model.Afterwards,the simplified Network in Network(NiN)block is proposed since it can combine the convolutional layer and multilayer perceptro,through which the nonlinear operations can be added in the convolutional layer and thus,the cross-channel interaction and information integration can be achieved.Lastly,to deal with the imbalance of the data categories,the simplified NiN block is incorporated with the augmentation algorithm.The main contents are as follows:(1)Lithology dataset is established.Firstly,the effective the logging feature and the velocity of seismic shear wave were selected to establish the sample point data set based on tne analysis of the logging data.Then,the lithology image was integrated from the low frequency information obtained by multi-resolution wavelet filtering.Especially,the sequence category label was used for the images to avoid the error classification.Thus,the the lithology image dataset with sequence category label is established.The experimental results show that the features are more characterized and the accuracy of the identification is imporved by adopting the single label.(2)A lithology recognition method based on 2D CNN is proposed,and on this basis,the attention mechanism is taken into consideration.As generally few lithology categories and features are contained in the dataset of one area,this paper builds a 6-layer simple 2D CNN network for lithology recognition,and the network structure of 2D CNN is determined through experiments.In addition,since humans tend to pay more attention to the area of interest when observing the outside world,this parper introduces the attention mechanism into the established 2D CNN network,and assigns more weight coefficients to the features with larger contributions.This is achieved by adding Squeeze-and-Excitation(SE)module at the second,fourth,and sixth convolutional layers.The experimental results show that the proposed lithology recognition method based on the established lithology image data set with sequence tags and 2D CNN is effective,and the improved lithology recognition method incorporating with the attention mechanism is even more effective.(3)A lithology identification method based on simplified NiN block and a lithology identification method based on simplified NiN block and data augmentation are proposed.Firstly,the selection among different basic blocks is conducted and the NiN block is finally chosen.Accordingly,the network framework is built.Then,through the experiments,the network structure(including the number of NiN blocks,the number of channels of the convolutional layer,and the size of the convolution kernel),initialization method and optimization algorithm are determined.The augmentation algorithm is developed based on this network model to solve the imbalance problem in the dataset.This is mainly for small dataset.The sample point data set is first augmented by linear interpolation,and then the augmented sample point sequence is converted into a lithological image.In order to increase the robustness of the method,Gaussian noise is also added to further augmente the data.Finally,the augmented lithology image data set is used to train a lithology recognition network based on simplified NiN blocks.The experimental results show that the lithology identification method based on the simplified NiN block has relatively good performance,and the lithology identification method based on the simplified NiN block and data augmentation is more effective.