such as Auto-Encoders (AE), Long Short-Term Memory networks (LSTM), and Dynamic Principal Component Analysis (DPCA) in terms of accuracy and reliability. To further enhance fault detection capabilities, a data-agnostic approach using Graph Neural Di"erential Auto-Encoders (GNDAE) is introduced. This model integrates Graph Convolutional Networks (GCNs) with a Graph Neural Di"erential Equation (GNDE) solver, providing continuous parameterization of hidden states. The GNDAE model treats the neural network as a di"erential equation, allowing it to learn values at infinitesimal points and capture both spatial and temporal dependencies in the data. This continuous modeling approach improves the understanding of underlying datasets and leads to highly accurate fault detection results.
Change point detection is addressed by applying the GNDAE model within an unsu-pervised learning framework. By modeling time-series data as graphs and incorporating contrastive learning techniques, the model e"ectively identifies significant changes in process parameters, indicating the onset of faults or transitions in operational states. The model’s ability to forecast feature values at subsequent time steps enhances its capability to detect magnitude changes, outperforming traditional methods such as CUSUM and LSTM-based approaches in chemical process case studies.
The importance of multi-objective optimization in industrial processes is emphasized, particularly in hydrocracking operations—a critical process in petroleum refining that con-verts heavy hydrocarbons into lighter, more valuable products like diesel, kerosene, and aviation turbine fuel (ATF). Optimizing the hydrocracking process is essential for maximiz-ing yields of desirable products while minimizing resource consumption, such as hydrogen, thereby enhancing economic e!ciency and contributing to sustainable resource utilization. To address this, the thesis develops an Hydrocarbon Graph Neural Di"erential Equation (HCGNDE) model that represents the hydrocracker unit by modeling the underlying partial di"erential equations as graphs. Variables are represented as nodes, and their relationships form the edges, capturing the axial dispersion of components through catalyst beds. The continuous parameterization inherent in the HCGNDE model allows for precise modeling of concentration changes across the process. The model is then utilized for multi-objective optimization to maximize the yields of ATF and diesel while minimizing hydrogen consump-tion. Results demonstrate significant improvements, with hydrogen usage reduced by 25.5% and yields of ATF and diesel maintained at high levels. This showcases the practical benefits of applying graph-based machine learning to complex industrial optimization problems.
In summary, this thesis provides comprehensive solutions for enhancing the safety, re-liability, and e!ciency of industrial process units through advanced graph-based machine learning techniques. By focusing on data-driven approaches aligned with Industry 4.0 objectives, the research addresses critical challenges in fault detection, change point detection, and multi-objective optimization. The proposed methodologies leverage the structural and relational properties inherent in industrial processes, o"ering robust, accurate, and scalable solutions that outperform traditional methods. The findings have practical implications for improving operational performance, reducing risks, and optimizing resource utilization in critical industrial processes like hydrocracking.