Active gamma-ray well logging pattern localization with reinforcement learning
Yuan Zi, Lei Fan, Xuqing Wu, Jiefu Chen, Shirui Wang, Zhu Han · 2022
Gamma-ray well-log depth matching is one of the essential tasks in the well-logging data processing. Up until now, welllog curves analysis and pattern hand-picking matching require intensive human-involved labour. Removing this contamination is challenging due to the complex nature of the patterns that correspond with the unknown strata. Besides, signal corruption such as blur, shift, and noise also obstructs the development of the automated pattern localization method. We propose a data-driven offline reinforcement learning (RL) framework to overcome those challenges. This system can learn the human attention-inspired active signal pattern localizing strategy. This framework leverages logged historical human labelled data to learn the process of human vision focus for logging patterns. Following top-down reasoning, this agent learns to deform a bounding window using simple transformation actions to determine the target signal segments’ location The experiments on augmented field Gamma-ray welllog dataset show the promising localization capability of reinforcement learning.