Abstract
Pertamina EP Cepu (PEPC) Jambaran Tiung Biru is a major gas producer and a strategic national project in Indonesia. In 2023, the solvent pump, crucial in the selexol solvent regeneration process to reduce H2S levels, experienced several trips, causing plant shutdowns and production losses. To prevent such failures, PEPC aims to develop an early warning system using anomaly detection to prevent equipment failures and mitigate production losses.
The Leading Equipment Anomaly Detection System (LEADS) was developed using historical solvent pump operating data to model normal conditions. It leverages a Deep Learning LSTM model with an Auto Encoder to reconstruct these conditions and compare them to real-time data. If the delta of Loss MAE between historical and current data exceeds a set threshold, it flags an anomaly. Alerts are sent to the JTB field operations team for proactive measures to prevent failures and shutdowns. LEADS also features plant visualization, historical anomaly trends, operational data mirroring, and automated email notifications.
Using the LSTM model with an autoencoder, the system learns normal operating conditions to establish threshold values for anomaly detection. These anomalies were then clarified and totaled for the number of anomalous results obtained in the four pumps where at P-9027A there were 263 anomalies with a threshold value of 0.108, at P-9027B there were 48 anomalies with a threshold value of 0.128, P-9025A found 3682 anomalies with a threshold value of 0.059, P-9025B found 116 anomalies with a threshold value of 0.145, P-90211A found 12 anomalies with a threshold value of 0.127, and finally at P9011B found 1504 anomalies with 0.103.
The results show a strong correlation between the detected anomalies and actual plant conditions, with each anomaly carefully verified by the operations team. LEADS' ability to reconstruct normal operating conditions in multivariate time series data significantly improves early detection of issues that could lead to solvent pump failures, minimizing the risk of shutdowns and production losses. To further enhance accuracy, it is recommended that specific models be developed for each type of equipment, even if they share the same design and operating parameters. This approach ensures that subtle differences in operation are captured, enabling more precise anomaly detection. The system's flexibility and scalability make it applicable to other critical equipment, offering a robust solution for proactive maintenance and operational reliability.
This paper presents a novel application of the LSTM autoencoder model to detect anomalies in solvent pump operations in gas plants, using multivariate time series data. LEADS reconstructs normal conditions and categorizes anomalies based on context, offering a proactive solution to prevent equipment failures and production losses. This approach enhances reliability in gas processing, particularly for H2S removal systems, and can be applied to other critical equipment, contributing significantly to operational efficiency and safety.
