AUTOMATIC DETECTION AND CLASSIFICATION OF RELEVANT EVENTS IN OIL WELLS USING NUMERICAL DERIVATIVES AND RULE-BASED SEGMENTATION

Abstract

Anomaly detection in oil wells is challenging, especially when it comes to identifying critical events impacting operations. This study proposes an automated methodology to segment relevant events from multivariate time series, improving anomaly detection in Inflow Control Valves (ICVs) by isolating periods with significant signal variations. The approach uses a rule-based algorithm that processes sensor data via numerical differentiation, Gaussian smoothing, and peak detection. Decision rules based on signal stability and temporal proximity filters enable precise segmentation. Evaluation shows an 88% overlap (IoU) with manual annotations, confirming the method’s effectiveness in identifying abrupt operational changes. One of our key contribution is to reduce data complexity and bias by excluding low-variability segments, producing a concise dataset focused on meaningful variations. To validate this preprocessing, a Random Forest classifier trained on the segmented data achieved an F1-score and accuracy of 0.80, correctly identifying 8 out of 10 anomalies. This performance surpassed the same model trained on the full dataset, which scored 0.60 on F1. These results validate the methodology’s effectiveness and suggest its potential for broader use in detecting not only ICV anomalies but also other significant events reflected by notable variations in well data.