Creating an Auto-Encoder Based Predictive Maintenance Tool for Offshore Annulus Wells

Abstract

A key component of well integrity is annular integrity. Much of the focus on this has been on establishing maximum and minimum pressure limits and designing envelopes under various well conditions encountered during well construction and subsequent production and injection operations. Many operators have established systems for operating wells within this design envelope to monitor for pressure excursions. However, abnormal annulus pressure behavior within the design envelope could be overlooked using a system that relies on limit monitoring and excursions. We propose a modeling workflow that combines novel deep learning techniques with statistical analysis to create online models which predict potential asset failures and alert on abnormal behavior such as abrupt pressure build up in producer and water injection wells’ A-Annulus. The model uses autoencoder architecture to learn the behavior of the wells during normal operating periods and generates alerts when it encounters new or abnormal behavior. The autoencoder architecture outputs a risk score aggregated over the residuals from all input features. Sequential Probability Ratio Test (SPRT) is performed on the risk score to determine abnormal regime during operation to raise alerts. These alerts can be used for root cause analysis based on the top contributors to the risk score. In our approach, we use feature thresholds as filters to determine normal operating periods for training the model. To simulate live conditions during model training, the historical time series data is divided into training and prediction windows. The model is trained on each training window and risk scores are created for the prediction window using a sliding window technique. To find the optimum model, a grid search is performed over a wide distribution of autoencoder and SPRT hyper-parameters. The models are scored based on recall, precision and lead time provided before a failure. We demonstrate this workflow using annulus pressure, downhole pressure, upstream choke pressure and upstream choke temperature as input to the model. The model does not require the physical properties of a well but uses historic well data lending itself to be applicable to already existing well stock. Next, we demonstrate using engineered features and synthesized data to efficiently train and score the models. During our experiments, we have explored several engineered features across multiple platforms and found that the correct set of engineered features can deliver a model that accurately alerts on asset abnormalities and potential failures. Our approach combines the strengths of deep learning techniques, statistical analytics and subject matter expertise to provide a framework that has demonstrated efficient scaling across multiple assets and sites and has potential application on a variety of oil and gas equipment.