Study On Probabilistic Inversion Method Of Logging-While-Drilling Azimuthal Resistivity Logging Data Based On Ensemble Learning

Resistivity logging is an important measurement method to characterize the electrical parameters of geological formations,which is widely used in fields such as qualitative evaluation of oil and gas reservoirs and stratigraphic section division.Inversion of formation parameter from resistivity logging data is an important aspect of geophysical inversion.Currently,inversion of resistivity logging data is mainly deterministic,where the inversion process treats the formation parameters as deterministic values and predefines the number of layers using algorithms such as machine learning to construct the inversion model.However,this type of inversion model has poor adaptability and makes it difficult to assess the reliability and uncertainty of the inversion results.In addition,there is limited research on the inversion of multi-layer formation logging data when the number of layers is unknown,as well as probabilistic inversion of resistivity logging data.To address the aforementioned issues,this paper introduces the NGBoost ensemble learning algorithm for probabilistic prediction and constructs a probabilistic inversion model that uses probabilistic prediction to quantify the uncertainty of the inversion results.Selecting suitable machine learning models as the base learners of the NGBoost algorithm and improving the probabilistic inversion model can enhance the model’s predictive capabilities;A lightweight boundary detection algorithm was designed to determine the layer boundary position and number of layers in real-time,meeting the requirements of geosteering;The boundary detection algorithm was combined with the probabilistic inversion model to propose a new probabilistic inversion method for multi-layer formation logging data.This approach enables real-time inversion of multi-layer formation logging data while providing an estimate of the inversion results’ uncertainty.The specific research contents are as follows:(1)The performance of the six machine learning models in the deterministic inversion of the logging data with drilling azimuth was compared,and the best machine learning model,XGBoost,was selected by combining the evaluation metrics.XGBoost is used as a base learner and combined with the NGBoost algorithm framework to construct a probabilistic inverse model N-XGBoost based on a hybrid integrated learning algorithm.The inversion model is applied to the probabilistic inversion of logging data from homogeneous formations with different layer numbers,and its application effects on the prediction of real formation logging curves are initially explored to verify the accuracy,reliability,real-time,robustness and practicality of the probabilistic inversion model.(2)A boundary detection algorithm based on second-order difference is designed based on the relationship between the azimuthal signal of the drilling azimuth electromagnetic wave logging tool and the formation boundaries.The detection algorithm is used to obtain the geometric parameters of the formation in real time from the azimuthal signal measured by the logging tool in the formation and to determine the type of the formation in the direction of the drill bit travel.The experimental results show that the detection algorithm can accurately and rapidly extract the geological parameter information from the azimuthal signals of multilayered homogeneous formations and noisy azimuthal signals.(3)The boundary detection algorithm is combined with the trained probabilistic inversion model to construct a probabilistic inversion method for multi-layer formation with drilling azimuthal logging data.Experiments are designed to verify that the inversion method can achieve probabilistic inversion of multi-layer formation logging data without predicting the number of layers,effectively evaluate the uncertainty of the inversion results and obtain reliable inversion results.In conclusion,the probabilistic inversion method of multi-layer formation logging data with fused boundary detection algorithm proposed in this paper can provide reliable log interpretation for geosteering work.