Name: Benjamin Peters 

Title: Real-time Data Analytics for Condition Monitoring of Complex Industrial Systems 

Advisor: 

• Dr. Nagi Gebraeel, School of Industrial and Systems Engineering (Georgia Tech) 

Committee Members:  

• Dr. Kamran Paynabar, School of Industrial and Systems Engineering (Georgia Tech) 
• Dr. Jianjun Shi, School of Industrial and Systems Engineering (Georgia Tech) 
• Dr. Nicoleta Serban, School of Industrial and Systems Engineering (Georgia Tech) 
• Dr. Timothy Lieuwen, School of Aerospace Engineering (Georgia Tech) 

Time: Friday, December 3rd at 10:00 AM ET 

Location: Microsoft Teams (Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTIwOTY2ZTgtMjJlYy00YjExLWE1YTgtYjQxNGYwMmM5MTI1%40thread.v2/0?context=%7b%22Tid%22%3a%22482198bb-ae7b-4b25-8b7a-6d7f32faa083%22%2c%22Oid%22%3a%22d5cedaf4-c523-4101-a5de-f989a67b9e47%22%7d

Join conversation

teams.microsoft.com

) 

Abstract: 

Modern industrial systems are now fitted with several sensors for condition monitoring. This is advantageous because these sensors can provide mass amounts of data that have the potential for aiding in tasks such as fault detection, diagnosis, and prognostics. However, the information valuable for performing these tasks is often clouded in noise and must be mined from high-dimensional data structures. Therefore, this dissertation presents a data analytics framework for performing these condition monitoring tasks using high-dimensional data. To collect this data, the projects discussed utilize either simulated data sets or state-of-the-art, industry-class test-rigs. This enables the validation of data-analytics techniques on real-world data. In Chapter 1, the proposed data analytics framework is outlined and various challenges associated with high-dimensional data are discussed. 

Chapter 2 develops a severity-based diagnosis framework for electric-powered vehicle electric systems with multiple, interacting fault modes. This framework leverages functional data profiles collected from a 4-cylinder automotive engine test-rig. The profiles are collected for various combinations of battery-motor state-of-health. Three regularized multinomial regression models are fitted, each using a unique feature extraction technique, to map profile features to the fault mode combination. Then ensemble methods are proposed that merge the results of these three models to create a more accurate model. The space of fault-mode combinations is partitioned into degrees of overall degradation. Thus, predictions of the fault-mode combination can be mapped to the overall system degradation. 

Chapter 3 discusses two projects related to the gas turbine combustor. These projects utilize industry-class combustors located in the Ben T. Zinn Combustion Laboratory at Georgia Tech. The first project involves utilizing a control chart to detect precursors to lean blowout, an important operational fault that can cause costly powerplant outages and compromise the safety of aircraft passengers. A probability distribution is fitted to these precursors to develop a measure of blowout risk that can be utilized by turbine or aircraft operators. The second project describes a hierarchical feature selection methodology for diagnosing the degradation state of the combustor centerbody, a component that provides stability for the combustor flame and protects the hardware of the combustor from the flame. A combustor test-rig is equipped with ten acoustic sensors which can be categorized as longitudinal (sensors mounted on the tubes which carry the fuel/air mixture to the combustion chamber) and transverse (sensors mounted on the support beams perpendicular to the tubes. This methodology utilizes multi-class logistic regression with adaptive group lasso penalty to select an optimal subset of sensors for monitoring. Then multi-class logistic regression with adaptive lasso is used to select features from the selected sensors to predict centerbody degradation. The results indicate that the proposed methodology is robust to reduction of the size of the training set. Furthermore, the longitudinal sensors are shown to be better predictors of centerbody degradation than the transverse sensors. 

Chapter 4 discusses two projects related to the gas turbine. These projects utilize an industry-class, single-stage gas turbine located at Pennsylvania State University. The first project proposes two approaches for predicting sealing effectiveness, defined as the ability of the combination of turbine geometry and coolant system to prevent ingestion of hot gas into the wheelspace region. The first approach utilizes Linear Regression with Lasso to map features from time-resolved pressure signals recorded either near the rim seal or on the outer casing to the sealing effectiveness. In the second approach, information regarding the nonlinear relationship between sealing effectiveness and the dominant frequency of the pressure signal was utilized to first make a coarse prediction of the sealing effectiveness. Then, given the coarse prediction, a more granular prediction is made. The result was a much more accurate diagnostic model when utilizing the two-step approach. The second project utilizes IR imaging of the gas turbine blade to diagnose faults in the cooling system. Various feature extraction techniques are analyzed for their predictive ability. Upon analyzing the IR image as a predictor, sparse regions corresponding to hot and cold regions of the turbine blade were found to be highly correlated with the cooling system faults. This demonstrated potential for optimal placement of more cost-effective sensors for implementation with fielded gas turbines. 

Chapter 5 discusses a data analytics framework for predicting remaining useful life in systems with multiple failure modes. This framework assumes that the causes of the observed failures are unknown. Therefore, an unsupervised model is proposed. Given that the cause of failure is known, it is assumed that the natural logarithm of the time-to-failure can be modeled as a Gaussian regression problem where the predictors are features extracted from the sensor signals. Thus, the marginal distribution of the natural logarithm of the time-to-failure can be modeled as a finite mixture of Gaussian regressions. To fit the parameters of this model, the log-likelihood is appended with the adaptive sparse group lasso penalty and optimized using a proposed Expectation-Maximization method. This method enables simultaneous clustering of the failures into separate failure modes and optimal sensor selection within each cluster. For prediction of the remaining useful life, a supervised mixture of gaussian model is fitted using the cluster labels and Multivariate Functional Principal Component (MFPC) scores from the selected sensors. Given the diagnosed failure mode, a Linear Regression with Lasso model is fitted to map MFPC scores from sensors selected for the diagnosed failure mode to the natural logarithm of the time-to-failure. By subtracting the current observation time from the time-to-failure prediction, the remaining useful life is predicted.