Abstract
Anomaly detection in surface logging data is crucial for improving efficiency and reliability, particularly in industries like oil and gas exploration. This paper introduces an advanced pipeline designed specifically for detecting low-quality intervals characterized by high anomaly densities in multivariate sensor data streams. The proposed methodology integrates traditional univariate anomaly detection checks with advanced deep learning techniques, employing autoencoders to effectively capture intricate interdependencies among sensors and temporal dynamics within data streams. Our structured approach includes preprocessing steps such as missing data handling and initial anomaly filtering, followed by training an autoencoder to differentiate normal patterns from anomalies using reconstruction error analysis. Additionally, a density-based clustering-inspired method is applied to group anomalous points and identify low-quality intervals. Experimental results validate the robustness and precision of the proposed approach, highlighting significant improvements of the novel thresholding technique compared to traditional approaches. This solution contributes to improved data integrity, facilitating informed decision-making processes in surface logging operations.
