A multi-view learning approach for lithology identification with incomplete logging datasets

Abstract Accurate lithology identification is often hindered by the limited use of multi-source information and the poor generalization capability of conventional methods, especially in geologically complex areas. Typically, the logging data in an area are multi-source and may be incomplete. In this study, a multi-view learning approach, which is named PDS-MVKNN and can effectively leverage multi-source logging data and accommodate incomplete datasets, is proposed for lithology identification. This approach incorporates the partial distance strategy, a method for calculating the distance between two feature vectors with missing values, into the multi-view K-nearest neighbours classifier and optimizes its view fusion strategy. In lithology identification on a dataset of complex igneous rocks, PDS-MVKNN achieves reliable performance even with a 10.43% data missing rate, reaching an identification accuracy of 93.80%. Compared with existing view fusion strategies, the model with the optimized strategy achieves a higher accuracy in lithology identification. By simulating different data missing rates in the dataset and various missing situations of logging curves in wells, this study also validates that PDS-MVKNN exhibits strong robustness to data missing.