Research On The Method For Lithology Identification Based On Deep Belief Network

Well logging is a kind of engineering technology used to detect oil and gas,coal and other mineral resources(such as uranium,potash,and hydrate)in wells by measuring the resistivity,ultrasonic,radioactivity,and nuclear magnetic resonance information of the formation.The logging data can be used to evaluate the formation parameters such as lithology,electrical properties,porosity,saturation and permeability.Particularly,the lithology identification results play a very important role in the oil and gas exploration.However,the logging data is influenced by many factors.The lithology type and the geology structure are very complicated,and the stratigraphic lithology differs greatly in different regions.Therefore,the existing methods have the problem of low accuracy when determining the lithology and have to establish different models for different regions,which is time-consuming and laborious.Accordingly,how to accurately identify the lithology becomes a key problem in the logging processing and interpretation.Based on the deep research on the application of traditional machine learning algorithm in lithology identification,this thesis discusses the models of Deep Belief Network(DBN)in detail.The concrete work is listed as follows.Firstly,the pre-processing of the original logging data not only improves the quality of the logging curve,eliminates the influence of the non-stratigraphic factors on the logging curve,but also reduces the number of log curves involved in the data analysis.This provides a good data base for next step model training.Secondly,a lithology identification method based on the DBN is designed.The method adopts the pre-training and tuning method to obtain a good initial state of the DBN model by unsupervised pre-training of lithology data,and then tune the parameters for the DBN model through the back propagation algorithm,which enable the network to have a strong feature learning ability.Again,the influence of different DBN parameters on the lithology identification results is studied.By adjusting the parameters of the DBN and fixing the other parameters,the experimental results are compared and the optimal network parameters are obtained.This provides a solid foundation for the application of the DBN model.Finally,a complete set of interactive " Lithology Identification " system is developed,with several common machine learning algorithms integrated,which realizes the full function of data preprocessing,model training and model application.The system has been applied to practical lithology identification and achieved good results.