Research On Lithology Recognition Based On Deep Learning

Lithology recognition is an important part of reservoir prediction.The traditional machine learning algorithm lacks the process of automatic feature extraction.Moreover,it cannot effectively utilize the local features of seismic data when the rock formation is recognized.In order to resolve the problem,deep learning models are designed and implemented,which can effectively extracting local features for lithology recognition.CNN-BRNNs are designed and implemented for lithology recognition based on Convolution neural networks(CNNs)and bi-directional recurrent neural networks(BRNNs).Lithology profiles recognized by CNN-BRNN are more continuous in the transverse direction and higher in the vertical direction than BP,CNN and BRNN.The persistent contrast divergence algorithm(PCD)is improved for learning Gaussian-Bernoulli Restricted Boltzmann Machines(GBRBMs)and BernoulliBernoulli Restricted Boltzmann Machines(BBRBMs).It solves the problem that the actual seismic data does not exactly conform to the model hypothesis distribution.The input of the model is extended from single point sampling to multi-point sampling.The local features of the formation is utilized more effectively.Moreover,the fault tolerance of the model is enhanced and lithological recognition effect is heightened.The experiment is conducted with the seismic data and lithological data.The results show that deep learning can effectively extract the local features of the seismic data and improve the recognition accuracy.