Case Study in Improving the Health of a Remote Monitoring & Diagnostics Center

Abstract

This paper provides a case study where data analytics techniques are used to improve the health of a global remote monitoring & diagnostics (RM&D) center for power plants that is a key part of our industrial internet infrastructure. The “Industrial Internet” is being heralded as a transformative, disruptive technology that is part of the digitization of traditional industries all over the world. A key technical capability includes the ability to continuously monitor critical assets (like power generation equipment) with sensors and other measurements to get leading indicators of anomalous behavior and identify opportunities for optimizing the performance and life of this equipment. This allows our customers to maximize their asset’s availability, reliability, and performance and is usually achieved by connecting these assets to dedicated RM&D centers that aggregate and analyze the data streaming in from all over the world.
An RM&D center is a complex system with hardware (data acquisition boxes, communication hardware, processing servers), software and analytics in place to ensure the generation of timely notifications and recommendations. There is a need to understand the health of this complex system and to quickly diagnose and mitigate issues before any disruption occurs that impacts the ability to monitor. What makes this case study unique is the combination of qualitative and quantitative input variables that need to be considered, which is different from traditional PHM applications which tend to be based on sensor derived numeric or binary data.
Details in the paper include extraction of key system health features (server health, data integrity, analytic robustness, etc.) from computer logs. These parameters, once collected, can be analyzed using various statistical techniques (Multivariate Outliers, Process Control, Mixture Distributions) to develop system health indices using methods like Principal Component Analysis, Factor Analysis as well as kernel PCA methods. We explore the development of anomaly detection methods using regression (Multilevel regression, Logistic regression), Clustering (K-Means, Hierarchic, Normal Mixtures and Latent Class Analysis) and Classification (Decision Trees, Bootstrap Forest, Discriminant) techniques followed by their comparison in terms of computational cost and classification accuracy. We then go further to pinpoint the likely cause (Bayesian Diagnosis) of disruptions, and demonstrate how this approach can be generalized for a variety of industrial assets. All of results/analysis are based on representative data from a global Remote Monitoring and Diagnostic center that is used to monitor a significant portion of the world’s electric power generation fleet.