Machine learning tecniques applied to hydrate failure detection on production lines

Abstract

The present work proposes a methodology that covers the whole process of classifying hydrate formation-related faults on production lines of an offshore oil platform. Three datasets are analyzed in this work, where each one of them is composed of a variety of sensor measurements related to the wells of a different offshore oil platform. Our methodology goes through each step of dataset cleaning, which includes: identification of numerical and categorical tags, removal of spurious values and outliers, treatment of missing data by interpolation and the identification of relevant faults and tags on the platform. The present work designs a framework that puts together many Machine Learning classic techniques to perform the failure identification. The system is composed of three major blocks: the first block performs feature extraction: as the input data is a set of time-series signals we represent each signal using its statistical metrics computed over a sliding window; the second block maps the previous block output to a more suitable space, this transformation uses the z-score normalization and the Principal Components Analysis (PCA); the last block is the classifier, the one we adopted was the Random Forest classifier due to its simple tuning and excellent performance. We also propose a technique to increase the reliability of the normal operation data. When handling a database composed by real data, it is usual to face a lot of mislabeled data, which can significantly jeopardize the model performance. Therefore, we deploy a technique to reduce the mislabeled samples, which presented an improvement of 7.93%, on average, reaching over 80% of accuracy in all single-class scenarios.