Development of an Artificial Intelligence-Based Well Integrity Monitoring Solution

Abstract

Abstract The question of how to safeguard well integrity is one of the most important problems faced by oil and gas companies today. With the rise of Artificial Intelligence (AI) and Machine Learning (ML), many companies explore new technologies to improve well integrity and avoid catastrophic events. This paper presents the Proof of Concept (PoC) of an AI-based well integrity monitoring solution for gas lift, natural flow, and water injector wells. AI model prototypes were built to detect annulus leakage as incident-relevant anomalies from time series sensor data. The historical well annulus leakage incidents were classified based on well type and the incident relevant anomalies were categorized as short and long-term. The objective of the PoC is to build generalized AI models that could detect historical events and future events in unseen wells. The success criteria are discussed and agreed with the Subject Matter Experts (SMEs). Two statistical metrics were defined (Detected Event Rate – DER – and False Alarm Rate – FAR) to quantitively evaluate the model performance and decide if it could be used for the next phase. The high frequency sensor data were retrieved from the production historian. The importance of the sensor was aligned with the SMEs and only a small number of sensors was used as input variable. The raw data was pre-processed and resampled to improve model performance and increase computational efficiency. Throughout the PoC, the authors learnt that specific AI models needed to be implemented for different well types as generalization across well types could not be achieved. Depending on the number of available labels in the training dataset, either unsupervised or supervised ML models were developed. Deep learning models, based on LSTM (Long-Short Term Memory) autoencoder and classifier were used to detect complex anomalies. In cases where limited data were available and simplistic anomaly patterns were present, deterministic rules were implemented to detect well integrity-relevant incidents. The LIME (Local Interpretable Model-Agnostic Explanations) framework was used to derive the most important sensors causing the anomaly prediction to enable the users to critically validate the AI suggestion. The AI models for gas lift and natural flow wells achieved a sufficient level of performance with a minimum of 75% of historical events detected and less than one false positive per month per well.