Research On Fault Recognition Method Based On Image Semantic Segmentation

In underground exploration,faults can directly reflect the geological and reservoir conditions such as reservoir description,trap evaluation and well placement,which greatly affects the efficiency and benefit of reservoir development.In the process of seismic data interpretation,the conventional fault interpretation method is that professionals manually identify and interpret fault data,and then evaluate the underground geological structure and oil and gas storage.However,manual fault identification takes a long time,is highly subjective,requires high professional ability,and the process is cumbersome.With the introduction of machine learning and deep learning,many new directions and ideas are provided for fault recognition.In this paper,fault recognition is classified as a binary classification problem of fault and non-fault.Image semantic segmentation algorithm can learn the context information of seismic data,and integrate fault details with fault semantic information to improve the learning ability of the model,so as to achieve better fault recognition.1.Construct U-Net network to identify faults.In this paper,both residual block and attention mechanism are introduced into U-Net network,which can curb the negative influence of poor feature fusion on fault recognition and achieve better feature fusion.At the same time,the weighted cross entropy loss function is introduced to solve the problem of data imbalance between fault samples and non-fault samples,so as to effectively prevent the adverse impact of data imbalance on model training,and improve the accuracy of model identification of fault details.2.Construct UNet++ network to identify faults.U-Net network directly fuses the fault feature information with multi-scale difference,which is not good for fault recognition.In view of this,this paper designs a fault recognition method based on the UNet++ model,which directly introduces more dense convolution blocks and skip connections,can extract more scale fault feature information,and perform feature fusion on similar fault feature information of adjacent two layers layer by layer to improve the accuracy of the model in fault recognition.At the same time,the structure of the model is determined by comparing the number of different layers and the number of convolution nuclei.In addition,the introduction of ECA attention mechanism in the model can prevent the negative impact of redundant fault feature information on fault recognition in the dense jump connection to a certain extent,and better learn favorable feature information to improve the noise resistance of the model.By comparing the number of convolutional nuclei,the number of network layers and the attention mechanism,the validity of the proposed network structure is verified.The fault prediction results of the seismic section in the actual working area show that the proposed UNet++ network with ECA attention mechanism and dense jump connection can clearly identify the fault details in the seismic data,improve the accuracy and F1 fraction value of the model,and improve the noise resistance of the model.