Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WB68

n WB68 West Bldg 105C Computational Services in Cybermanufacturing Systems Sponsored: Quality, Statistics and Reliability Sponsored Session

n WB69 West Bldg 106A Statistical Learning and Modeling of Smart and Connected Systems Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Chao Wang, University of Wisconsin-Madison, Madison, WI, 53706, United States 1 - Cardiac Surface Potential Prediction Based on Gaussian Process and Partial Differential Equation Zhiyong Hu, Texas Tech University, 2500 Broadway, Room 240, Lubbock, TX, 79415, United States, Dongping Du Reconstructing the surface potential in cardiac chambers plays an essential role for precise diagnosis and treatment planning. Gaussian Process (GP) is a popular method that learns the spatial temporal correlations of available data for surface mapping. However, GP kernel functions cannot capture the physiological correlation among cardiac tissue. In this study, we integrate GP with physical models of heart to characterize the covariance structure of surface potential recordings. The resulting model enables an accurate surface prediction, thus providing more insight into cardiac disorders. 2 - Fault Diagnosis in Multistage Manufacturing Processes Based on Bayesian Linear Random Effects Model with Sparse Variance Components Prior Jaesung Lee, University of Wisconsin, 3255, Mechanical Engineering Building, 1513 University Ave, Madison, WI, 53706, United States, Junbo Son, Shiyu Zhou We present a Bayesian linear random effects model for the fault diagnosis in the multistage manufacturing processes with a prior for the sparse variance components. A modified horseshoe+ prior is used to tackle the low sample size and high dimensional problems with sparse faults. Furthermore, we introduce the informed-horseshoe+ that incorporates the positional fault likelihood information. To estimate the faults from the horseshoe+ prior, Gibbs sampler was developed. In simulation and numerical experiments, the horseshoe+ outperformed the inverse gamma prior in sparse fault cases and the informed- horseshoe+ achieved advantages given additional information on process faults. 3 - When are Markets out of Control? Monitoring Financial Networks with Online Hurdle Models Mostafa Reisi Gahrooei, Georgia Tech, Atlanta, GA, 30318, United States, Samaneh Ebrahimi, Shawn Mankad, Kamran Paynabar The interconnectedness of financial institutions can function as a mechanism for the propagation and amplification of shocks throughout the economy, thus contributing to financial crises. As such, network analysis has become a critical tool to assess interconnectedness and systemic risk levels. In this paper we create a formal monitoring system to detect changes within a sequence of sparse networks constructed from an interbank lending market in the European Union. 4 - Transfer Learning of Structures of Ordered Block Graphical Models Using Informative Priors Chao Wang, University of Wisconsin Madison, Rm. 3255, ME building 1513 University Ave., Madison, WI, 53706, United States, Shiyu Zhou We focus on the structure learning of ordered block model (OBM) based on data- driven informative prior. A novel informative structure prior distribution is represented by a categorical random variable. In this way, the informative prior distribution construction is equivalent to the parameter estimation of the graph random variable distribution using historical data. The sample space inconsistency between historical and new OBMs is addressed by adding pseudo nodes with probability normalization, then removing extra nodes through marginalization. The simulation and real case study show the effectiveness of the proposed method, especially when the data from the new OBM is small.

Chair: Ran Jin, Virginia Tech, Blacksburg, VA, 24061, United States Co-Chair: Xiaoyu Chen, Virginia Tech, Blacksburg, VA, 24061, United States 1 - Statistical Transfer Learning for Profile Monitoring Ziyue Li, The Hong Kong University of Science and Technology, Kowloon, Hong Kong, Fugee Tsung, Ke Zhang In the practical manufacturing process, there exist quite amounts of similar procedures or stages. When a new task comes, data in new task is too limited to get an accurate parameter estimation. Transfer learning outperforms by joint estimation with Gaussian Graphical Model. Some profiles are observed with sparse anomaly. To deal with it, we propose a new method to capture the relatedness across multiple profiles and detect their anomalies efficiently. Profile is decomposed into smooth background, sparse anomaly and random noise. Smooth background is treated with transfer learning model and group-lasso is introduced for anomaly. Coordinate Descent is adapted to solve the optimization problem. 2 - A Multitask Learning Based Decomposition Approach for Modeling Distributed Machines Hao Yan, Tempe, AZ, 85281, United States, Yifu Li, Ran Jin Quality prediction is traditionally conducted on the system within a single machine. In this work, we consider the data are collected from multiple machines with different input variables. The machines share similarity but may not be identical. We decompose the variation patterns of multiple machines into explainable variation patterns by input variable, latent variation patterns shared by multiple machines, and noise. Furthermore, we utilize multi-task learning to model the explainable variation patterns. This decomposition can be used for quality prediction and process monitoring. Simulation and case study are conducted to demonstrate the advantage of the proposed algorithm. 3 - A Physics-specific Change Point Detection Method Using Torque Signals in Pipe Tightening Processes Juan Du, Xi Zhang, Jianjun Shi In pipe tightening processes, the torque signals presents with various quasi- periodic profiles,introducing a bunch of false change points while using existing change point detection methodsfor condition monitoring. To address this issue, this presentation proposes to consider the profilegenerating mechanism to improve the detection performance. 4 - Personalized Recommendation for Information Visualization Xiaoyu Chen, Virginia Tech, 302 Heartwood Xing, Blacksburg, VA, 24060, United States, Ran Jin Information visualization help acquire insights of complex datasets in system engineering. However, a uniform visualization format may be ineffective to bring the insights to different users at different contexts. In this research, a personalized recommendation approach is proposed to dynamically adjust visualization with respect to different tasks and contexts for different users. It can automatically select the significant features from unobtrusive measures to support further diagnostics of the visualization designs. 5 - Clustering-based Data Filtering for Manufacturing Big Data System Yifu Li, Virginia Tech, 302 Heartwood Crossing, Blacksburg, VA, 24061, United States, Xinwei Deng, Ran Jin, Shan Ba, William Myers As the sensing technology advances in manufacturing, various sensors enable real-time modeling and monitoring of in situ covariates. However, sensors collect massive data, which cause high computational load and excessive data storage in manufacturing. To address these issues, we proposed an unsupervised data filtering method based on index-segmented and clustered datasets. Furthermore, a filtering information criterion is proposed to automatically determine the proportion of data filtered for further data analysis, which effectively balances the trade-off between the sample size of the filtered datasets and information preserved.

466