At the 2025 SPE Europe Energy Conference and Exhibition in Vienna, Fabio Concina from Kwantis presented an innovative AI-based solution aimed at enhancing the quality of surface logging data. His work, titled “Detecting Low-Quality Intervals in Surface Logging Data Using AI-Based Anomaly Detection,” demonstrates how advanced machine learning techniques can effectively identify and address low-quality intervals within surface logging datasets. By applying anomaly detection algorithms, this approach offers the potential to significantly improve data reliability, streamline analysis, and support better decision-making in drilling operations.
The Challenge: Data Quality in Surface Logging
Surface logging data is essential for safe and efficient drilling operations, but this time-series data often contains anomalies caused by sensor errors or processing faults. Detecting these low-quality intervals early is crucial to ensure reliable, real-time decision-making and to maintain operational integrity.
The Solution: Combining Classical Methods with AI
At the SPE Europe Energy Conference & Exhibition, Kwantis presented a powerful AI-based anomaly detection pipeline. Their method integrates traditional statistical techniques for simple anomaly detection with AI models capable of uncovering more complex, multivariate issues that conventional rule-based systems typically overlook.
The Approach: Anomaly Detection Pipeline
The process starts with classical preprocessing methods such as threshold filtering and analyzing variable relationships. The pipeline then applies an autoencoder—a type of neural network—to reconstruct normal data patterns. Deviations between the original and reconstructed signals highlight potential anomalies. To accurately classify these, the team developed a distribution-based thresholding technique, outperforming standard percentile or Z-score methods.
The Impact: Enhancing Data Reliability
This AI-driven solution enables more effective identification of low-quality intervals in surface logging data, improving the accuracy of operational decisions and reducing the risk of misinterpretation. Looking ahead, the team plans to explore probabilistic models like Variational Autoencoders and algorithmic techniques like Matrix Profile to further advance anomaly detection in this critical field.