Explicabilidade de algoritmos de classificação aplicados à detecção de anomalias em tubulações de gás e petróleo

Abstract

The oil industry understands that predicting anomalies in production wells can help reduce maintenance costs, prevent production losses, and mitigate the risk of accidents that could impact the environment and human safety. Similar to various industrial processes, the oil extraction and transportation system is also subject to different types of anomalies. To identify the occurrence of these anomalies during oil production and develop methods that enable differentiation of anomalies with similar production impacts, this work combines the use of the Random Forest classification algorithm with Explainable Artificial Intelligence (XAI) tools. The classification algorithm was applied to identify different classes present in a public database containing real production data from various oil wells. To ensure that the algorithm's results are understandable and trustworthy, XAI tools were used to develop methods that clearly and accessibly explain the internal workings of the models. This may include techniques that help understand which features or data influenced a decision, making the process more transparent. In this work, two tools were selected: SHAP and LIME. Both were tested and analyzed based on various evaluation criteria. Subsequently, these criteria were part of an evaluation conducted using the Analytic Hierarchy Process (AHP) for comparing the chosen tools. The obtained results demonstrated success in anomaly classification, with F1-Scores of 100%. The use of XAI libraries revealed that the classification algorithm was capable of establishing conclusions in accordance with thermodynamic principles. Both were able to highlight the pertinent importance of each variable in the classification. Although SHAP and LIME have distinct theoretical foundations, their explanations converged in most cases. Finally, a formal comparison via the Analytic Hierarchy Process (AHP) indicated SHAP's superiority for this application. [resumo fornecido pelo autor]