Inferential measurement for integrity and security of cyber-physical systems.

Asset integrity is vital for the safety and reliability of Cyber-Physical Systems in industries like energy, transportation, and healthcare. It prevents hazards, minimizes downtime, and protects against cyber threats while ensuring economic efficiency. Additionally, it promotes sustainability by reducing environmental damage and extending asset lifespan. A key element of asset integrity management is the inferential measurement system, which includes a fault diagnostics model for identifying the root causes of faults, a soft sensing model for predicting hard-to-measure variables, and a security framework to safeguard the sensors and Internet of Things network. There are two main techniques for developing an inferential measurement system: model-driven and data-driven approaches. Data-driven models generally outperform model-driven ones by leveraging data for decision-making. However, they struggle to capture the complete spatiotemporal relationships in time series sensor data, which can lead to degraded performance. When developing a data-driven model, it's also important to ensure data quality, extract meaningful features, and manage the complexities (spatial relations among the sensors) of industrial control systems. Additionally, model interpretability is a challenge, particularly when identifying faulty equipment based on the model's decisions. Finally, securing the sensor and Internet of Things enabled industrial control system network is vital to prevent cyber threats that could produce false results. This PhD thesis aims to develop a data-driven inferential measurement system comprising a fault diagnostic model, a soft sensing technique, and a malware prevention model to guarantee asset integrity and maintenance of cyber-physical systems. As part of the inferential measurement system, we have developed a fault diagnostic model to detect and diagnose the root cause of the fault by integrating a Variational Autoencoder with a Graph Convolutional Network. The developed fault diagnostic model is then evaluated on a solar or PV array dataset and compared with conventional Autoencoder based models by detecting five different types of array faults. The experimental outcome shows that the developed model can learn the spatiotemporal characteristics of the sensor data more effectively than the conventional Autoencoder models. We have also developed a soft sensing model depending on the spatiotemporal attention LSTM encoder-decoder model for predicting the formation pressure of oil/gas wells. A novel dataset is generated from a digital twin of the oil/gas well drilling rigs to evaluate the performance of the developed soft sensing model. Comparison outcomes with alternative Machine learning and Deep learning models reveal that the proposed soft sensor can outperform the baseline LSTM models for predicting hard-to-measure formation pressure variables from the secondary variables. Finally, we have proposed a machine learning-enabled micro-segmentation process to mitigate the lateral movements of malware. A security analysis is performed to demonstrate the effectiveness of the proposed automated micro-segmentation model in reducing the propagation of malware through the sensor/IoT integrated industrial control network.

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2025

Author dc:creator

• Arifeen, Murshedul

Advisor dc:contributor.advisor

• A. Petrovski

• Cyber-physical systems
• Inferential measurement systems
• Fault diagnosis
• Soft sensing
• Microsegmentation
• Intrusion detection

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:2934857
https://doi.org/10.48526/rgu-wt-2934857

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:2934857