Anomaly Detection of Gas Pipeline Operational Data Using TCN-Autoencoder and LSTM-Autoencoder Models

Anomalies in gas pipeline systems, such as undetected small leaks, are often the primary triggers of major failures that impact operational safety and the environment. Early detection of such anomalies is crucial to prevent larger risks. This study aims to compare the performance of two deep learnin...

Full description

Saved in:
Bibliographic Details
Published in:2025 International Conference on Data Science and Its Applications (ICoDSA) pp. 1347 - 1352
Main Authors: Gde Pradnyana, Anak Agung, Ihsan, Aditya Firman, Hasmawati
Format: Conference Proceeding
Language:English
Published: IEEE 03.07.2025
Subjects:
Anomaly detection
autoencoder
Autoencoders
Computational modeling
Computer architecture
Data models
Euclidean distance
Long short term memory
LSTM
operational data
Pipelines
Sensitivity
TCN
Training
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Anomalies in gas pipeline systems, such as undetected small leaks, are often the primary triggers of major failures that impact operational safety and the environment. Early detection of such anomalies is crucial to prevent larger risks. This study aims to compare the performance of two deep learning-based autoencoder models-Temporal Convolutional Network Autoencoder (TCN-AE) and Long Short-Term Memory Autoencoder (LSTM-AE)-in detecting anomalies in unlabeled operational data from gas pipelines. The data used in this study comprises four key features: pressure, temperature, energy rate, and volume rate, all collected on a time-series basis at one-hour intervals over a five-year period. Each model was tested through six architectural variants (cases) with consistent training parameters. The experimental results demonstrate that the baseline TCN-AE model provides the most efficient outcome with the lowest Mean Squared Error (MSE) of 1.43×10 −6 and the fastest training time. Meanwhile, the bidirectional LSTM-AE architecture exhibits better generalization capabilities in recognizing complex anomaly patterns, although it requires a longer training duration. Euclidean distance was also used to determine the anomaly threshold during the evaluation phase. With these findings, the study aims to provide strategic insights into selecting the best model for time-series-based anomaly detection in gas pipeline systems. Hopefully, this research will help the oil and gas industry enhance surveillance and prevent significant losses caused by undetected damage.
DOI:10.1109/ICoDSA67155.2025.11157467