Study On Uranium Logging Lithology Identification Based On ADASYN-SSA-XGBoost

Efficient and accurate lithology identification can help characterise the geology and improve the efficiency of uranium exploration and development.Machine learning techniques have strong data mining capabilities and show a greater advantage in identification studies.In view of this,this thesis takes a sandstone uranium mining area in the southern part of the Songliao Basin as the research object,takes conventional logging data in the selected research area as the data set,and adopts an improved machine learning method to perform lithology identification on conventional logging data.To improve the performance of lithology identification and provide reference value for uranium exploration and development.(1)To complete logging curves containing missing values,this thesis proposes a logging curve completion method based on a long short-term memory(LSTM)neural network,considering the time-series nature of logging data and the complex mapping relationships between different logging parameters.Firstly,the logging data is pre-processed using the depth correction method and the maximum-minimum normalisation method.Secondly,the LSTM neural network is used to build a time-series prediction model to predict the missing values of the curve by multiple variables and fill the predicted values to the missing values to complete the logging curve.The experiments demonstrate that the LSTM neural network algorithm can accurately predict the curve data and is effective in completing the missing logging data,laying the data foundation for subsequent lithology identification studies.(2)To improve the hyperparametric search efficiency of conventional integrated machine learning algorithms,the Sparrow Search Algorithm(SSA)was used to optimise the Extreme Gradient Boosting(XGBoost)algorithm to construct the SSA-XGBoost lithology identification model.The number of trees,the maximum depth of trees and the learning rate were found to be 14,19 and 0.78 respectively.The recognition accuracy of the overall lithology was 95%,96% and 97%,respectively,which were 19%,20% and 20% better than the unoptimized XGBoost algorithm.(3)For improve the effect of category imbalance on the recognition effect of minority lithology and further enhance the overall lithology recognition effect,ADASYN was used to balance the logging data.The algorithm is able to automatically generate corresponding samples according to the distribution of samples around the minority category,and to keep the number of minority category samples consistent with the number of majority category samples by increasing the number of samples at the boundary.A comparison experiment with SMOTE and MAHAKIL data balancing algorithms was also conducted to verify the advantages of the ADASYN algorithm in balancing uranium logging data.Combining the ADASYN algorithm and the SSA-XGBoost lithology identification model,the final ADASYN-SSA-XGBoost lithology identification model was constructed.The model achieved above 97% about F1 values for fine sandstone,mudstone,calcareous sandstone,calcareous mudstone and siltstone,and 99%,97% and 99% accuracy for the overall lithology on the three uranium logging data sets,respectively.The model proposed in this thesis is feasible in the field of lithology identification research.