Conjunction of Active and Semi-Supervised Learning for Wireline Logs-Based Automatic Lithology Identification

Abstract

The sample labeling is the basis for wireline logs-based automatic lithology identification (WLALI). Due to the paucity of core data, the expert-annotation accounts for most labeling workload. In the traditional workflow of WLALI, the samples for labeling are passively selected. The labeled samples thus fail to reflect the overall distribution and feature of whole data set, leading to the difficulty in training the prediction model with robust generalization performance. To solve this problem, a novel workflow of hybrid active and semi-supervised learning-based WLALI has been proposed. First, we alter the conventional labeling process to active selection mode by employing the approach of active learning. In such way the labeled samples are representative of whole data features. Second, we propose a novel active learning algorithm based on density difference of Gaussian probability (DDGP) to obtain a better performance of sample query. Third, the semi-supervised learning method is introduced to combine the active learning algorithm, attempting to minimize the involvement of hand-crafted labeling under the target performance. In the evaluation, the proposed workflow not only has achieved favorable result in the uppercase image dataset, but also the practical application shows great promise. When compared with the workflow of random sampling + semi-supervised learning that is similar to the traditional mode, the average F1-score of proposed workflow of DDGP + semi-supervised learning can increase by approximately 3.47%.