Open-world anomaly detection for oil-well production systems under evolving operating conditions

Reliable identification of abnormal production events is essential for safe and efficient oil-well operation. Undetected or misclassified anomalies may lead to delayed intervention, production loss, and increased operational risk. In practical production systems, well conditions may change over time due to flow variations, sensor noise, and complex production disturbances. These changes make abnormal-event identification challenging, especially when emerging anomalies deviate from previously observed production patterns. To address these challenges, this study proposes an adaptive open-world anomaly detection framework for intelligent oil-well monitoring. MiniROCKET is used to extract temporal features from multivariate production data. Based on these features, an adaptive anomaly identification strategy is developed by combining prediction uncertainty, feature-space similarity, and temporal variation. This strategy enables the model to flag production behaviors that deviate from known anomaly patterns as unknown events. To support model updating in evolving production environments, an incremental learning mechanism is further introduced. It combines replay, knowledge distillation, and parameter regularization to incorporate newly observed anomaly categories while reducing catastrophic forgetting of previously learned classes. Experiments on oil-well monitoring data demonstrate that the proposed method improves anomaly classification and unknown-event detection under evolving operating conditions. The proposed framework provides an effective solution for intelligent production monitoring and adaptive anomaly-event management in oil-well production systems.