| dc.description.abstract | The oil and gas industry faces significant operational risks from anomalous events, necessitating effective Abnormal Event Management (AEM) to prevent production losses, environmental damage, and safety hazards. Automation of AEM using machine learning, particularly Long Short-Term Memory (LSTM) networks, has shown promise in detecting anomalous patterns in real-time sensor data. This study applies LSTM for time-series classification of anomalies using the real incstances of 3W Dataset, a collection of over 2000 events from offshore wells. The research aims to evaluate LSTM's effectiveness in identifying anomalies and comparing to other deep learning models i.e., RNN and GRU. The models were trained with data from sensors of pressure and temperature in production choke valve and temperature-pressure transducer. To show comparison the model were trained with different period of observations: 60, 120, and 180 seconds. The results of this research indicate that LSTM models is capable and most effective to classify anomaly with the observation of 2 minutes, proven by the findings of 92% F1-score. Moreover, LSTM model shows consistency of balanced F1 score across all period of observations comparing to RNN or GRU. LSTM-based anomaly classification has competitive performance compared to previous work in anomaly classification of 3W Dataset. | en_US |
