Automatic Detection and Classification of Relevant Events in Oil Wells Using Numerical Derivatives and Rule-Based Segmentation

Resumo

Anomaly detection in oil wells is a complex task, especially when identifying critical events that affect well operations. In this context, the oil industry has increasingly relied on computational techniques to enhance operational safety, leveraging time series data recorded by pressure and temperature sensors along wells. Significant variations in these signals over time may indicate relevant events. While all anomalous events are relevant, the converse is not necessarily true — not all relevant events are anomalies. However, focusing anomaly detection exclusively on relevant events simplifies the process. This work proposes an automated approach for detecting relevant events in production/injection oil wells, using a rule-based decision algorithm combined with statistical analysis tools. The objective is to identify the start and end points of relevant events, segmenting them into well-defined intervals. The technique applies numerical differentiation and Gaussian smoothing to sensor data, in order to reduce noise and highlight peaks caused by abrupt variations. Decision rules are then used to group these variations into coherent intervals. Each detected event is represented by a compact set of attributes, such as absolute, relative, and temporal differences between its start and end points. The main advantage of this approach lies in its ability to filter out noise and focus on relevant segments throughout the wells’ lifecycle, excluding periods of low variability that add little value and may introduce bias into machine learning processes. This focus on high-variability segments results in a cleaner, more representative dataset for modeling purposes. This initial detection serves as a starting point for distinguishing operational events from anomalies and preparing the data for more complex models. Initial results indicate that the statistical features extracted from the segmented intervals are promising for detecting anomalies in Inflow Control Valves (ICVs). Experiments using real well data, under different operational conditions and with recorded ICV failures, enabled the evaluation of the supervised classifier K-Nearest Neighbors (KNN) in distinguishing between normal and anomalous relevant events. The results indicate that the proposed method is effective in automating the identification and classification of anomalies, contributing to the continuous monitoring of oil well operations.