Digital twins improve fault diagnosis in high-tech engineering systems
Farhad Ghanipoor defended his PhD thesis at the Department of Mechanical Engineering on March 13th
In the fast-paced world of high-tech industries, such as printer and semiconductor manufacturing, unexpected system failures can lead to costly downtime and repairs. To address this challenge, hybrid fault diagnosis methods have been developed, combining physics-based models with data-driven approaches. With his PhD research, Farhad Ghanipoor advances fault estimation methodologies and model updating techniques, specifically for nonlinear systems. By leveraging digital twin technology, this research aims to improve fault detection, enable predictive maintenance, and enhance system reliability.
Traditional fault diagnosis methods that rely solely on either mathematical models or operational data often struggle to handle the complexities of real-world, nonlinear systems operating in dynamic environments. Ghanipoor introduces a hybrid approach that merges the strengths of both paradigms, providing robust solutions for fault detection, isolation, and estimation. The study focuses on tackling key challenges such as modeling uncertainties, disturbances, and measurement noise, thereby enhancing the accuracy and reliability of fault diagnosis.
Innovative methodologies for fault estimation and model updating
The first part of the thesis presents novel fault estimation methodologies that can handle a comprehensive range of scenarios, including time-varying process and sensor faults. It develops a methodology that guarantees precise fault estimation for a class of time-varying faults in the absence of uncertainties and noise. Additionally, the approach provides explicit performance bounds in the presence of disturbances and noise. A computationally efficient algorithm is introduced, enabling performance trade-off analyses and enhancing fault estimation accuracy.
Moreover, the thesis introduces a model updating framework for locally Lipschitz nonlinear systems. By learning modeling uncertainties from input-output data, this method reduces computational costs and ensures stability properties. Two distinct approaches are proposed, based on cost and constraint modification, offering tractable convex programs for updating models while guaranteeing system stability.
Real-world applications in high-tech industries
The second part of the thesis focuses on practical applications of these hybrid fault diagnosis methods. One application is in wafer handler robots in semiconductor manufacturing. A hybrid fault diagnosis method combining model-based fault estimation with data-based classification is applied to robotic manipulators. Given the highly nonlinear behavior of these robots, a modified fault estimation method is developed to handle smooth nonlinearities. The simulation-based results show superior performance compared to purely data-driven approaches.
Another application is ink channel diagnosis in industrial printers. A hybrid method integrating model-based fault detection with linear regression and k-nearest neighbors classifiers is applied to self-sensing piezo actuators. Addressing the challenge of simultaneous actuation and sensing, the proposed fault detection method demonstrates excellent performance on both simulated and experimental datasets.
Advancing digital twin technology
This thesis is part of the nationwide digital twin program (Project 4.3), contributing significantly to the development of digital twins for fault estimation in high-tech systems. The digital twin framework not only enhances real-time fault detection but also lays the foundation for predictive maintenance strategies. By reducing unscheduled maintenance and improving system reliability, the research offers substantial economic and societal benefits.
Title of PhD thesis: . Promotor: Prof. Nathan van de Wouw. Co-promotors: Dr. Carlos Murguia Rendon, and Dr. (TU Delft).