Vertaileva tapaustutkimus syväoppivista menetelmistä ja digitaaliseen kaksoseen perustuvasta menetelmästä dynaamista kynnysarvoa hyödyntäen aikasarjojen poikkeamien havaitsemisessa
Year: 2026
Type: dissertation
Abstract: Industrial Control Systems are central part for monitoring and controlling industrial processes. These however remain vulnerable to anomalies such as equipment faults, sensor errors, and cyberattacks. To detect these issues, anomaly detection methods are utilized, and during the last few years, specifically detection methods based on both deep learning and digital twin simulation have gained attention. However, comparisons between these fundamentally different approaches are scarce. This thesis presents a case study comparing these two types of anomaly detection approaches with Industrial Control System time-series data. Two deep learning methods are included, which learn normal system behavior and detects anomalies deviating from the learned behavior. The second type of approach uses adaptive thresholding to detect anomalies from residual differences between the digital twin and the physical system. This study evaluates and compares the anomaly detection performance, limitations, and sensitivity to operational changes between the approaches, under identical conditions. The experiments were conducted on a physical water process testbed in real-time and complemented with synthetic experiments, which simulated the operation of real-time anomaly detection. During the experiments, all methods performed anomaly classification using the same time-series sequences, which corresponded to the fixed-length window of sequential observations. Additionally, depending on the type of approach the methods utilized different subsets of measured features. The results show that deep learning based approaches achieve lower false positive rates and higher recall, which indicate better overall detection performance. In contrast, the digital-twin based approach enables earlier detection by identifying anomalies with a fewer anomalous observations. The deep learning methods are mainly constrained by the availability and quality of the training data, which can lead to reduced sensitivity or overfitting to normal behavior. On the contrary, the digital twin approach mainly depends on its digital twin, requiring it to accurately reflect the physical system dynamics to ensure reliable residual-based anomaly detection.
Source: Aaltodoc (Aalto University)
Author: Salminen, Teemu
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203728626