Implementation of Ensemble Learning Methods in Autoencoder Architecture for Anomaly Detection in Natural Gas Pipeline Operational Data

Pipeline monitoring plays crucial role in ensuring safety and efficiency of natural gas transportation systems. Undetected anomalies may result in severe environmental and financial consequences, emphasizing need for robust monitoring systems. Research implements an ensemble of autoen-coders (EoAE)...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2025 International Conference on Data Science and Its Applications (ICoDSA) s. 281 - 286
Hlavní autoři: Azzaka, Muhammad Rivaditya, Firman Ihsan, Aditya, Astuti, Widi
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 03.07.2025
Témata:
anomaly
Anomaly detection
autoencoder
Autoencoders
Ensemble learning
Monitoring
Natural gas
Noise reduction
pipeline
Pipelines
Reliability
Sensitivity
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Pipeline monitoring plays crucial role in ensuring safety and efficiency of natural gas transportation systems. Undetected anomalies may result in severe environmental and financial consequences, emphasizing need for robust monitoring systems. Research implements an ensemble of autoen-coders (EoAE) approach for detecting anomalies in natural gas pipeline operational data through integration of three complementary autoencoder architectures. These include the standard autoencoder (AE) for baseline pattern recognition, the sparse autoencoder (SAE) for selective feature extraction, and the denoising autoencoder (DAE) for noise-resistant analysis. Implementation uses unlabeled dataset spanning two years containing 61,313 hourly records with 17 features categorized into operational parameters and gas composition parameters. Averaging mechanism combines reconstruction errors from three autoencoders, establishing balanced threshold for anomaly detection. Results demonstrate ensemble of autoen-coders (EoAE) successfully identified 41 anomalies (0.338% of test data) with mean squared error of 0.581, showing improved reliability compared to individual models which detected between 39-143 anomalies at varying thresholds. Analysis reveals anomaly propagation patterns across multiple operational and compositional parameters, demonstrating system capability to identify complex behavioral changes in pipeline operations. Experimental results validate effectiveness of ensemble approach in achieving balanced anomaly detection through complementary model integration. Research advances pipeline monitoring methodology through integration of complementary autoencoder strengths for robust anomaly detection.
DOI:10.1109/ICoDSA67155.2025.11157176