Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SB71

4 - Monitoring of User-generated Reviews via a Sequential Reverse Joint Sentiment-topic Model Qiao Liang, Tsinghua University, Beijing, China, Kaibo Wang User-generated reviews can serve as an efficient tool for evaluating the customer- perceived quality of online products and services. This article proposes a joint control chart for monitoring the quantitative evolution of topics and sentiments in online customer reviews. A sequential model is constructed to convert the temporally correlated review documents to topic and sentiment distributions, which are subsequently used to monitor the topics and topic-specific opinions in an ongoing product and service process. Simulation studies on various data scenarios demonstrate the superior performance of the proposed control chart in terms of both shift detection and diagnosis. 5 - Optimal Design of Reliability Demonstration Tests with Risk-adjusted Costs Suiyao Chen, University of South Florida, 5017 Patricia Court, Tampa, FL, United States, Lu Lu, Qiong Zhang, Mingyang Li Conventional optimal design of reliability demonstration tests (RDTs) mainly minimizes the testing costs within a RDT, but neglects its impacts on subsequent reliability assurance activities, such as reliability growth and warranty. This work investigates the influence of RDT design on its subsequent reliability activities and further proposes an optimal RDT design strategy by jointly considering cost components at both design and operational phases in a more holistic manner. A comprehensive case study is given to demonstrate the benefits of the proposed work under different cost scenarios and prior elicitation settings. n SB70 West Bldg 106B Joint Session QSR/DM: Condition-based Maintenance Optimization Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Anahita Khojandi, University of Tennessee, 521 Tickle Building, 851 Neyland Drive, Knoxville, TN, 37996, United States Co-Chair: Mahboubeh Madadi, Louisiana Tech University, Ruston, LA, 71272, United States 1 - Understanding Wind Turbine Performance for Better Condition Monitoring Hoon Hwangbo, Texas A&M University, College Station, TX, 77840, United States, Yu Ding For wind turbines, a direct monitoring of power output data may not be sufficient for tracing the performance change of wind turbines, due to the strong dependency of power output on wind conditions. As such, we model power output as a function of wind variables and provide an approach to monitor the change of the function via a performance metric defined over the function. Since the functional model, albeit data driven, is developed in a way it follows physical restrictions, it is better representative of wind turbine performance and more suitable for the condition monitoring of wind turbines. 2 - Selective Maintenance of Systems with Multiple Dependent Failure Modes Cesar Ruiz, University of Arkansas, 4183 Bell Engineering Center, Fayetteville, AR, United States, Edward A. Pohl, Haitao Liao Selective maintenance may be conducted by choosing a subset of possible actions during the downtime of a system. We propose a selective maintenance optimization framework for complex systems with components that have more than one failure mode. In particular, both non-catastrophic and catastrophic failures are considered. The first occurs when the physical characteristic of a component exceeds a predetermined threshold. On the other hand, the time-to- failure of a catastrophic failure is modeled by a Weibull distribution where the scale parameter depends on the state of the non-catastrophic failure process. A genetic algorithm is utilized to determine the optimal maintenance actions. 3 - Dual-stage Attention-based Recurrent Neural Networks for Prognostics and Smart Maintenance Jos Carlos Hernández Azucena, PhD Student, University of Arkansas, 800 W. Dickson St., Fayetteville, AR, 72701, United States, Haitao Liao The present work proposes the use of a Dual-Stage Attention-Based Recurrent Neural Network (DA-RNN) for the estimation of remaining useful life of an individual complex system with multivariate measures. Using the C-MAPSS turbine engine run-to-Failure datasets (PHM08 data challenge) as a benchmarking setting, an initial implementation is carried out using Python. The potential of the current algorithm and technical modifications that may arise as research progresses are expected to be of great benefit for the advancement of prognostics and smart maintenance.

4 - Joint Optimization of Resource Allocation and Maintenance Planning for a Multi-Facility Infrastructure System Yue Shi, Yisha Xiang Maintenance management is intriguing when considering deteriorating facilities at a system level, which represents more challenges on modeling the interdependencies among facilities. Meanwhile, maintenance policies are usually made by agencies responsible for an entire network, leading to an increasing need to optimally allocate the limited budget. This research develops an integrated resource allocation and maintenance planning for a deteriorating transportation infrastructure system that consists of multiple facilities with complex maintenance effects. The proposed method is illustrated using real-world pavement deterioration data from the state of Florida. 5 - Optimizing Condition-based Maintenance for Systems with Degrading Sensors Anahita Khojandi, University of Tennessee, 521 Tickle Building, Knoxville, TN, 37996, United States, Mahboubeh Madadi We consider a degrading system with non-silent failures. System failure is costly and it requires an immediate system replacement. The system status can be partially observed using a series of heterogeneous sensors at a given cost. The sensors are not only noisy, but also are prone to degradation themselves, and hence, the sensor noise level is a function of the sensor status. We develop a POMDP model to determine the optimal condition-based maintenance scheme to minimize the total expected discounted cost over an infinite horizon. n SB71 West Bldg 106C New Methods and Algorithms in Statistical Learning Sponsored: Computing Sponsored Session Chair: Weijun Xie, Virginia Tech, Blacksburg, VA, 24060, United States 1 - The Computational Operations Research Exchange (CORE) Project for NSF Yunxiao Deng, University of Southern California, Los Angeles, CA, 90089, United States, Carl Kesselman, Suvrajeet Sen We will introduce a new platform named Computational Operations Research Exchange (CORE) with several illustrative examples. The CORE platform allows OE (Operations Engineering) researchers to leverage data, models, and software created by an ecosystem of researchers so that they are able to use the exchange to demonstrate that the value of their research results using the cyber- infrastructure developed, maintained, and updated by this platform. 2 - Solving Structured Nonconvex Problems in Statistical Learning Rahul Mazumder, Massachusetts Institute of Technology, Sloan School of Management, 100 Main Street, Cambridge, MA, 02139, United States Nonconvex problems arise frequently in machine learning, posing challenges from a computational and statistical viewpoint. Continuous especially convex optimization, has played a key role in our understanding of these problems. However, some other well-grounded techniques in mathematical optimization (for example, mixed integer optimization) have not been explored to their fullest potential. I will demonstrate how such techniques can be used to address problems in structured sparsity. I will outline instances in robust statistics, nonparametric function estimation and low-rank factor analysis where such techniques seem to be promising. 3 - Online Active Set Algorithm for Generalized Solution Paths in Machine Learning Ammon Washburn, University of Arizona, 2552 N. Geronimo, Tucson, AZ, 85705, United States, Neng Fan, Hao Helen Zhang Parameter tuning is computationally difficult in machine learning due to cross- validation on a large search space. Solution path algorithms (or online active set in optimization) alleviate this difficulty by analytically solving each successive optimization program from previous optimal solutions. The current theory for general online active set assumes certain constraint qualifications and a strictly convex objective that are not met in general for machine learning problems. We extend the theory of online active set to deal with parametric convex QPs with no constraint qualifications and show a particular implementation with a new model of Support Vector Machines.

57