Automatic Well-Log Depth Shift with Multilevel Wavelet Decomposition Network and Dynamic Time Warping

Well log measurements serve as continuous records providing indirect assessments of formation properties, constituting the primary input for the development of both static and dynamic reservoir models. In fact, not all well logging data of a single well are measured at once, high-precision depth matching of well logging data is crucial for later rock physics interpretation and machine learning correlation extraction. To address this imperative, A novel approach based on the Multilevel Wavelet Decomposition Network integrated with a Gated Recurrent Unit network, complemented by Dynamic Time Warping was constructed for automated well-log depth correction. The Multilevel Wavelet Decomposition Network is adept at extracting frequency information from well logging data, thereby enabling a nuanced understanding of the intricate geological features. Simultaneously, the Gated Recurrent Unit networks efficiently capture depth sequences information, enriching the contextual understanding of the subsurface formations. The incorporation of Dynamic Time Warping ensures accurate depth matching and correction, optimizing the alignment of well logs.This methodology was rigorously evaluated in the 2023 SPWLA PDDA Machine Learning Competition, yielding noteworthy results. The misaligned well logs have been successfully rectified, encompassing crucial parameters such as RHOB, NPHI, and RD, aligning them seamlessly with a reference GR log. Evaluation metrics, including NMSE (Normalized Mean Squared Error) and MAD (Mean Absolute Deviation), underscore the efficacy of our approach, with the optimal performance recorded at NMSE=0.3148 and MAD=21.7658. This achievement secured the top position in the aforementioned competition.This research signifies a pioneering application of frequency information derived from well logging data to address the intricate task of depth alignment across disparate well log types. Notably, the complexity introduced by aligning logs of distinct types is surmounted, demonstrating the robustness and foresight of our proposed method. The outcomes presented herein affirm the efficacy of our approach, heralding a significant advancement in the realm of well-log depth correction and alignment.