Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MC70

3 - A Feature Ranking Method for Process Data Based on Distance Correlation

n MC68 West Bldg 105C Data-driven Approaches to Predictive Analytics under Uncertainty Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Abdallah A. Chehade, University of Michigan-Dearborn, Dearborn, MI, 48128, United States 1 - Sensor Fusion via Statistical Hypothesis Testing for Prognosis and Condition Monitoring Abdallah A. Chehade, University of Michigan-Dearborn, 4901 Evergreen Road, 2280 HPEC, Dearborn, MI, 48128, United States The rapid development of sensing technologies made multiple sensors available to real-time monitor the degradation status of machine systems. However, almost no method provides a statistical metric to evaluate the quality of degradation signals for prognosis and condition monitoring. To fill this literature gap, the paper constructs a health index via a sensor fusion framework for prognosis and degradation analysis through a series of statistical hypothesis tests. The performance of the proposed is evaluated and compared to benchmark methods on a publish C-MAPPS dataset. 2 - Data-Driven Sensitivity Indices for Models with Dependent Inputs using the Polynomial Chaos Expansion Zhanlin Liu, University of Washington, Seattle, WA, United States Variance-based sensitivity analysis characterizes how the variance of a model output is propagated from the model inputs. Conducting such analysis for models with dependent inputs requires strong assumptions in the literature. We propose data-driven sensitivity indices for models with dependent inputs using the polynomial chaos expansion where the strong assumptions are not required. Two numerical examples, in which sensitivity indices cannot be directly obtained by the existing methods, are used to validate our proposed method. 3 - Sensitivity Analysis on a Network using the Polynomial Chaos Expansion Zhanlin Liu, University of Washington, 8033A 45th Ave NE, Seattle, WA, United States Network structures exist in various engineering applications. Conducting variance-based sensitivity analysis on a network helps to control the quality of the network output. We propose the network sensitivity indices based on the sparse network polynomial chaos expansion to measure the sensitivity indices for the controllable independent inputs to the network output. The proposed method is more computationally efficient and requires fewer observations of the inputs and the output than using the Monte Carlo method. Two manufacturing examples are implemented to validate the proposed method. High-Dimensional Data Analysis for System Prognostics and Performance Improvement Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Xiaolei Fang, Georgia Institute of Technology, Atlanta, GA, 30329, United States 1 - Reliability Analysis of Repairable Systems with Dynamic Covariate Information Wujun Si, Wichita State University, 11316 E. Pine Meadow Ct, Wichita, KS, 67206, United States, Qingyu Yang With the advancement of sensoring technology, dynamic covariate information is often recorded simultaneously with reliability data. We propose a novel reliability model for repairable systems when dynamic covariate information is presented. A maximum likelihood estimation method is developed to estimate the model parameters. A simulation study is implemented to illustrate the developed methods, and a real world case study is conducted to demonstrate the proposed model. 2 - A Prognostic Framework for Complex Engineering Systems with Multi-failure Modes Liexiao Ding, Georgia Institute of Technology, Atlanta, GA, United States, Xiaolei Fang, Nagi Gebraeel Many systems and subsystems usually have more than one cause of failure. Therefore, in real world applications, it is important to distinguish between these different causes of failures also known as Multi-failure Modes. To address the challenges, we develop a prognostic methodology, utilizing multi-sensor stream, to identify failure mode and then predict remaining useful life(RUL) of partially degraded systems. We propose a finite mixture of (log)-location-scale regression to model inhomogeneous sensor streams a grouped-penalized maximum likelihood estimator to perform sensor selection. n MC69 West Bldg 106A

Andi Wang, Georgia Institute of Technology, H. Milton Stewart School of, Industrial and Systems Engineering, Atlanta, GA, United States, Juan Du, Xi Zhang, Jianjun Shi In advanced manufacturing systems, multiple features are extracted from the data obtained from each working stage to characterize the process condition. Some of the features relate to the quality of the final products, and may provide valuable information on the quality of the product before inspection. Based on distance correlation metric, we propose a method that ranks the importance of the features obtained from the process based on their general dependency with the product quality. Theoretical properties of our method is studied, and a real case study in a solar cell manufacturing process validates the effectiveness of our method. 4 - A Supervised Dimension Reduction-Based Prognostics Model for Applications with Incomplete Signals and Censored Failure Times Xiaolei Fang, Georgia Institute of Technology, Atlanta, GA, 30329, United States, Kamran Paynabar, Nagi Gebraeel, Weijun Xie The incompleteness of degradation signals and censoring of historical failure times pose significant challenges for prognostics. To address these challenges, this paper develops a novel method that combines a feature extraction term and a regression term. The first term is capable of extracting features using multi-stream incomplete signals, while the second term regresses the features against censored failure times. By simultaneously optimizing the two terms, the extracted features are guaranteed to be most informative for predicting failure times. To solve the optimization problem, a Block Prox-Linear Coordinate Descent algorithm with a global convergence property is developed. n MC70 West Bldg 106B New Advanced Deployment Features of Modeling Languages for Optimization Sponsored: Computing Sponsored Session Chair: Bjarni Kristjansson, Maximal Software Inc., Arlington, VA, 22201, United States 1 - Optimizing in the Cloud - Deploying Optimization Models on the Cloud with Web Services REST API’s Over the past decade the IT has been moving steadfastly towards utilizing software on clouds using Web Services REST API’s. In this presentation we will demonstrate the new MPL REST Server, which allows optimization models to be easily deployed on the cloud. By delivering optimization through a standard REST API, which accepts data in JSON and NoSQL formats, the optimization becomes purely data-driven. Client applications can now be implemented relatively easily on different client platforms such as mobile/tablets or web sites, using just standard HTML/CSS with Javascript, or any other preferred programming language. 2 - New Programming Tools and Interfaces for Deploying AMPL Models Robert Fourer, AMPL Optimization Inc., 2521 Asbury Ave, Evanston, IL, 60201, United States, Filipe Brand π o Though fundamentally declarative in design, optimization modeling languages are invariably implemented within larger modeling systems that provide a variety of programming options. Although programming is not used to describe models, it facilitates the integration of models into broader algorithmic schemes and business applications. This presentation surveys ways in which a programming interface can be useful for development and deployment, with examples from the AMPL modeling language and system. The focus is on new APIs (application programming interfaces) for controlling AMPL from programs in Python and in R, and on new facilities for invoking Python programs from within AMPL. 3 - Enhanced Model Deployment and Solution in GAMS Steven P. Dirkse, GAMS Development Corporation, Washington, DC, United States, Fred Fiand The new capabilities recently added to GAMS make it easier than ever to develop, solve, and deploy models in non-traditional and powerful ways. Our cross- platform GAMS Studio provides a fresh look that will be welcomed by Mac users. Our embedded Python facility opens up a wealth of possibilities and makes data movement between GAMS and Python fast and painless, while providing a seamless union of Python code in a GAMS model. Finally, we have multiple ways to solve & deploy models to take advantage of parallel computing and opportunities to solve a group of similar models with only a single model generation. In this talk, we’ll introduce these capabilities and demonstrate their use with a series of examples. Bjarni Kristjansson, Maximal Software Inc., 2111 Wilson Boulevard, Suite 700, Arlington, VA, 22201, United States

209