Graph Attention Network-Based Monitoring of Complex Operational Systems
(2024)Department of Automatic Control
- Abstract
- Operational systems, such as industrial automation, autonomous vehicles, and larger air/sea/landcrafts, often contain a large number of heavily connected systems with real-time requirements for their functionality. For such systems, detecting and responding to anomalies is both challenging and crucial. Until recently, such anomalies were monitored using heuristic methods, or even humans monitoring the systems. Such approaches often fail to detect anomalies accurately due to the complexity of the systems. A continuous development of the systems also poses a significant challenge, as the current anomaly detectors have to be updated, and staff trained. Geometrical deep learning is a well known tool used for anomaly detection in applications... (More)
- Operational systems, such as industrial automation, autonomous vehicles, and larger air/sea/landcrafts, often contain a large number of heavily connected systems with real-time requirements for their functionality. For such systems, detecting and responding to anomalies is both challenging and crucial. Until recently, such anomalies were monitored using heuristic methods, or even humans monitoring the systems. Such approaches often fail to detect anomalies accurately due to the complexity of the systems. A continuous development of the systems also poses a significant challenge, as the current anomaly detectors have to be updated, and staff trained. Geometrical deep learning is a well known tool used for anomaly detection in applications where the data can be represented as a graph. However, to our knowledge it is yet to effectively be used for complex operational systems, currently only being used for simpler cases such as fraud detection. Recently, a new architecture named Graph Attention Network (GAT) has been studied as an anomaly detection method. Its ability to incorporate information in large networks makes it potentially useful for complex operational systems. In this thesis we evaluate different machine learning based methods for anomaly detection, trained on data from real operational systems, focusing on submarines. The methods evaluated include GCNs, GATs and Autoencoders. We also evaluate which data preprocessing methods that are best suited for our case. The results of this thesis provide a basis for further research and show that GATs could be successfully implemented as anomaly detectors for complex operational systems, though the usage may not be justified without sufficient data and complexity. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9174455
- author
- Källander, Ivar and Swirski, Stanislaw
- supervisor
- organization
- year
- 2024
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6249
- other publication id
- 0280-5316
- language
- English
- id
- 9174455
- date added to LUP
- 2024-09-16 08:48:00
- date last changed
- 2024-09-16 08:48:00
@misc{9174455,
abstract = {{Operational systems, such as industrial automation, autonomous vehicles, and larger air/sea/landcrafts, often contain a large number of heavily connected systems with real-time requirements for their functionality. For such systems, detecting and responding to anomalies is both challenging and crucial. Until recently, such anomalies were monitored using heuristic methods, or even humans monitoring the systems. Such approaches often fail to detect anomalies accurately due to the complexity of the systems. A continuous development of the systems also poses a significant challenge, as the current anomaly detectors have to be updated, and staff trained. Geometrical deep learning is a well known tool used for anomaly detection in applications where the data can be represented as a graph. However, to our knowledge it is yet to effectively be used for complex operational systems, currently only being used for simpler cases such as fraud detection. Recently, a new architecture named Graph Attention Network (GAT) has been studied as an anomaly detection method. Its ability to incorporate information in large networks makes it potentially useful for complex operational systems. In this thesis we evaluate different machine learning based methods for anomaly detection, trained on data from real operational systems, focusing on submarines. The methods evaluated include GCNs, GATs and Autoencoders. We also evaluate which data preprocessing methods that are best suited for our case. The results of this thesis provide a basis for further research and show that GATs could be successfully implemented as anomaly detectors for complex operational systems, though the usage may not be justified without sufficient data and complexity.}},
author = {{Källander, Ivar and Swirski, Stanislaw}},
language = {{eng}},
note = {{Student Paper}},
title = {{Graph Attention Network-Based Monitoring of Complex Operational Systems}},
year = {{2024}},
}