Information-Theoretical Approaches to Condition-Monitoring Using Physics Driven Models (Project 1)
The team will lead explorations of machine performance monitoring and prediction algorithms based on the fundamentals of physics and information theory. The methods will be based on detailed physics models of the monitored system to interpret complex sensor signals, identify faults, and check observability of sensor systems to see if the data actually contain the needed diagnostic information. The model parameters will be tuned to match the data measured from the healthy system, after which they will be updated to match the newly arrived sensor data from the monitored system. System degradation will be assessed based on the view of the monitored system as a communication channel and use of the physics based model to evaluate the information content dissipated in it. Prognosis will be based on fitting thermodynamics-inspired differential equations to the continuously updated system parameters and predicting their behavior based on those equations.
Operating Regime Dependent Condition Monitoring and Prediction (Project 2)
In this project, Dr. Djurdjanovic will lead the explorations of a novel data-driven approach to monitoring of complex systems operating under variable operating conditions. The method is based on characterizing the degradation process via a set of operation specific Hidden Markov Models (HMMs), whose hidden states represent the unobservable degradation states of the monitored system, while its observable symbols represent the sensor readings. The new research will focus on understanding of uncertainties of the identification of degradation HMMs from sensor readings emitted by the monitored machine, enabling the HMM-based degradation modeling and prediction paradigm to systems undergoing continuously variable operating conditions (batteries, automotive engines) and implementation of the new paradigm to such systems.
Immunity Inspired Approaches to Condition-Monitoring (Project 3)
In complex machines with a large number of interconnected dynamic subsystems, faults in one subsystem can cascade and affect the behavior of numerous other subsystems. This complicates the traditional fault monitoring procedures because of the need to train models of all the faults that the monitoring system needs to detect and recognize. Traditional diagnostic methods are based on recognizing or predicting previously anticipated and modeled faults, which results in limited diagnostic coverage and inability to deal with new faults. To circumvent these difficulties, a novel approach inspired by the functionalities of a natural immune system will be pursued in this project. The methodology is based on the use of advanced dynamic models that are capable of discerning abnormalities caused by unusual inputs into the system from those caused by changes in internal dynamics of the system (real fault). Based on those models, distributed anomaly detection will be used to encircle subsystems and components that are causing anomalous behavior with alarming anomaly detectors, similarly to how biological immune systems label antigens with appropriate antibodies, after which leukocytes dispose of anything labeled with antibodies. The newly proposed research will be led by Prof. Djurdjanovic and will be conducted in several directions. Firstly, we will focus on facilitating dynamic modeling of systems of interconnected dynamic systems in which interrelations between subsystems are not fully understood. In addition, the UT team will pursue optimal diagnostic coverage of the system via distributed anomaly detection schemes that achieve a tradeoff between detection speed and accuracy on one side, and computational load on the other. Finally, the team will apply the newly developed diagnostic approaches to actual systems in which interactions between subsystems are not completely understood, such as PECVD or lithography tools in semiconductor manufacturing.
Integrated Condition Monitoring and Process Control in Semiconductor Manufacturing Equipment (Project 4)
Exponentially Weighted Moving Average (EWMA) run-to-run (RtR) control is the standard control method in semiconductor manufacturing. Current RtR control relies on metrology results that are subjected to high measurement costs and poor availability. Virtual metrology (VM) is able to predict end-of-batch wafer properties from process conditions and could be used to improve EWMA controller performance at reduced costs. This project aims to develop methods to incorporate VM predictions into standard EWMA controllers. Using a Reliance Index (RI) metric, VM-assisted EWMA controllers can incorporate virtual metrology predictions in the RtR decision-making process. EWMA RtR control loop controller performance assessment needs to be able to handle non-threaded EWMA RtR control under high-mix manufacturing. The developed technique will first be verified using an industrial dataset and then implemented in an online environment and interfaced with a standard factory fault detection system.
Predictive Maintenance Paradigms in Biomedicine (Project 5)
The proposed research will yield an effective modeling strategy to relate the electromyogram (EMG) signatures of muscles to the joint kinematics in human limbs. The modeling strategy is based on locally tractable autoregressive with exogenous input (ARX) models. These models will use features derived from Cohen’s class of time-frequency distributions of EMG signals and will be used as inputs for the growing structure multiple model system (GSMMS), modeling the joint kinematics and dynamics. The local tractability of the ARX models and the GSMMS structure offer significant benefits in terms of anomaly detection and localization in the neuro-musculo-skeletal (NMS) system. The use of a family of models for each degree of freedom (each joint) permits joint-level anomaly detection and localization. The physical interpretability of the ARX models and the features that drive them will facilitate methods to localize NMS anomalies at the muscle level as well.
Maintenance Logistics and Operations (Project 6)
In this project, the team that will explore incorporation of new models of degradation and maintenance into logistics models including spares stocking, technician dispatching, consumables replenishments and maintenance tool acquisition and siting. We plan to expand the focus of the maintenance decision-making models to include multiple sites, multiple equipment, and geographically dispersed install base. Spatial considerations will be explicitly modeled (distances between maintenance personnel, tools, and target systems) along with considerations of limited quantity and capacity of maintenance and repair resources. The contribution of this project will be incorporation of new degradation dynamics models into maintenance logistics models and development of scalable solution methodologies for industrial use. We plan to build on our previous research on multi-location multi-indenture service parts logistics network design, level of repair analysis and inventory stocking models and corresponding optimization techniques. We expect to achieve the following milestones: (1) Incorporate new degradation models into maintenance planning, and scheduling, (2) Integrate maintenance logistics for parts and tools, and consider spatial issues and resource capacities and quantities, (3) Develop scalable solution methods for increasingly more integrated models, and (4) Start implementing the proposed tool at an actual system. We intend to coordinate our efforts with the degradation dynamics research efforts so that the new maintenance logistics models are able to integrate them. Collaboration with industrial members will provide a use case with data and insights. The companies should benefit from improved and more quantitative maintenance decision making, integrated and coordinated activities across maintenance areas including, but not limited to logistics and hardware support, up-time maximization with controlled maintenance logistics costs.