INFORMS 2021 Program Book

INFORMS Anaheim 2021

MD43

4 - Deep Spatio-temporal Anomaly Detection in Laser Powder Bed Fusion Sepehr Fathizadan, Arizona State University, Tempe, AZ, United States, Feng Ju, Yan Lu, Zhuo Yang Parametric and regression-based anomaly detection methods often fall short when faced with high-dimensional data containing rich spatio-temporal correlations. The multitude of unrealistic assumptions renders the decisions made by such methods unreliable and prone to large errors. It has been shown that relying on a single melt pool image to detect anomalies in the process is usually misleading and can result in significant inflation of false alarm rates. In this study, we propose a configuration of convolutional long short-term memory auto- encoders to learn a deep spatio-temporal representation from the sequence of melt pool images and perform anomaly detection.

MD41 CC Room 212A In Person: Electrical Grid Analysis/Transmission Planning General Session Chair: Afzal Siddiqui, Stockholm University, Stockholm, Sweden 1 - Integration of Optimization and Machine Learning for Improving Electrical Grid Operation Carl Laird, Professor, Carnegie Mellon University, Pittsburgh, PA, 87123-3453, United States, Jordan Jalving, Logan Blakely, Fani Boukouvala, Zachary Kilwein, Michael Eydenberg Reliable, optimal operation of the electrical grid requires consideration of generator costs and operating feasibility. Machine learning can be a valuable tool to approximate complex operating constraints. There has been an increase in research focused on the ability to optimize models that include neural network constraints with general nonlinear or piecewise linear activation functions. In this presentation, we demonstrate the use of machine learning to replace challenging constraints in electrical grid optimization problems. 2 - Sustainable Transmission Planning Afzal Siddiqui, Stockholm University, Stockholm, Sweden, Makoto Tanaka, Yihsu Chen Increased penetration of intermittent variable renewable energy sources (VRES) requires variability management, which often refers to storage and transmission investment. However, the cost of damage from emissions is overlooked in favour of VRES targets. We use a bi-level framework to devise transmission plans that directly include the cost of damage from emissions. Our upper level comprises a welfare-maximising transmission planner who internalises the damage cost. At the lower level, profit-maximising firms invest in VRES capacity and operate their fleet of assets. We implement problem instances for a Western European test network in order to examine how storage and transmission are complements or substitutes in integrating VRES. MD42 CC Room 212B In Person: Deep Learning/Machine Learning II Contributed Session Chair: Sepehr Fathizadan, ASU, Tempe, AZ, 85281, United States 1 - Learning a Continuous Search Space for Discrete Routing Problems Using Autoencoders Kevin Tierney, Bielefeld University, Bielefeld, 33615, Germany, André Hottung, Bhanu Bhandari Methods for automatically learning to solve routing problems are rapidly improving. However, most methods are unable to effectively utilize longer run times because they lack a sophisticated search component. We present a learning- based optimization approach that learns to map instance-specific routing problem solutions to points in a continuous space, thus turning a discrete search space into a continuous search space that can be explored using any unconstrained continuous optimization method (e.g., differential evolution). Our approach outperforms existing machine learning based approaches for the traveling salesperson problem and the capacitated vehicle routing problem. 2 - Deep Learning-based Disease Diagnostic Biomarker Detection With Metabolomics Data Seonyoung Kim, Chungnam National University, Daejeon, Korea, Republic of, Taewon Go, Dongil Kim Biomarkers play important roles in the early diagnosis of disease to improve the survival rate of patients. To determine the type and stage of diseases and the corresponding treatment options, the target biomarkers should be identified. Metabolomics provides information on cellular metabolic processes that drive tumorigenesis and tumor progression. In this study, we propose deep learning methods to detect disease diagnostic biomarkers with metabolomics. Through experiments, we compare the performance of the proposed method and other machine learning methods. 3 - Interpretable Trees Zheng Zhang, University of Tennessee, Knoxville, Knoxville, TN, United States Tree- and rule-based models have been known to have good interpretability, as the trained predictive models form a set of decision rules that are easy to understand in practice. Existing methods select the best model by minimizing the prediction errors on both child nodes (i.e., too much or too little) and predict values as the average of all observations in each terminal node. In this paper, we propose a split criterion based on rank concordance and a rule by fitting each terminal node with a linear function. The proposed method rather than the traditional performed better in accuracy and interpretability (e.g., capture the variations and associations between target variable and predictors).

MD43 CC Room 213A In Person: Neural Networks Contributed Session

Chair: Angela Avila, San Antonio, TX, 78240, United States 1 - Analysis for the Continuous Version of the Alternative Fuel Refueling Station Location Problem Sara F. Abu Aridah, PhD Student, Pennsylvania State University, University park, PA, United States, Omar Abbaas, Jose Antonio Ventura We address the deviation-flow refueling station location problem. We start with a continuous network where any point is considered as a candidate station location. Then the network is discretized by rounding distances and the driving range to the closest integer multiple of a common divisor value. This reduces the required computational effort to solve the problem. In this network, we prove that given any feasible solution with refueling stations located at non-integer distances from network vertices, it is always possible to find an equivalent integer solution. These results are used to discretize the network and propose an efficient polynomial time algorithm to locate a set of refueling stations. 2 - Predicting Utility Power Line Risk From Tree Failure via an Interpretable Convolutional Neural Network Nasko Apostolov, Graduate Research Associate, University of Massachusetts Amherst, Amherst, MA, United States, Ryan Suttle, Jimi Oke, Sanjay Arwade, Brian Kane Automating tree risk assessments, which are critical to the integrity of utility power lines, could potentially yield significant cost savings while boosting community resilience. We train a novel convolutional neural network to predict tree failure likelihood categories using augmented inputs from expert-assessed tree images. Via cross-validation and hyperparameter optimization, we obtain a binary classifier with an accuracy of 0.94 (SD = 0.1). We assess the visual interpretability of the classifier using techniques such as gradient-weighted class activation maps. Our framework demonstrates the potential of artificial intelligence for sustainable infrastructure management. 3 - Classifying Soil Moisture Levels of Grazeland Using UAV Imagery Data Based on a Convolutional Neural Network Method Angela Avila, University of Texas at Arlington, Arlington, TX, United States, Jianzhong Su, Heidi Taboada, Huihui Zhang In field management, soil moisture is a key predictor in yield success. With acres of fields and mindfulness of water conservation, our goal is to maximize crop production at minimal cost. Unmanned Aerial Vehicles can be used to take images and provide information on ground moisture. We can then derive vegetation indices and use Convolutional Neural Network to distinguish well maintained areas from areas that are in water deficit. With imaging data collected in Fort Collins ARS, we will train the CNN system to successfully reach about 89% accuracy in predicting an image class correctly. With UAV data and CNN method we can precisely predict soil moisture levels and tend to areas with more irrigation needs.

85