Dimension reduction modeling methods for high dimensional dynamic data in smart manufacturing and operations

Data from industrial operations are usually in time series forms. With the growing popularity of industrial internet of things (IIoT), time series data become massive, high-dimensional, dynamic, and collinear. These characteristics make the data contain both serially correlated and serially uncorrelated features. On the other hand, engineering fields have developed valuable first-principles models with many years of investment. 

This project aims to attack two challenges in developing time series data analytics for industrial systems: 

1. How to extract low dimensional dynamic features in parsimonious forms for accurate prediction and interpretation 

2. How to effectively incorporate first-principles models with latent variable analytics 

This project serves to develop fundamental understanding and algorithms to make use of high-dimensional, dynamic data for predictive analytics and monitoring. The project begins with surveying methods in engineering, financial time series, and statistics. We then integrate them with dynamic system theory to form a new framework of dimension reduction expression of state space (DRESS). The problem of dimension reduction with non-dynamic data has been well studied in the literature. On the other hand, the problem of time series analysis without dimension reduction has also been well studied. However, the massive dynamic data from IIoT and Industry 4.0 calls for a new framework, in which dimension reduction and dynamics must be analyzed simultaneously. With such an approach, high-dimensional data can be effectively utilized for predictive analytics. Further, the reduced-dimensional latent variables can be visualized for monitoring and decision-making. 

In addition, this project develops a method to incorporate first-principles models and known factors into latent variable data analytics. Therefore, prior knowledge accumulated in the first-principles models can complement the power of big data analytics. The integrated approach will facilitate fast adoption of data science by industries and provide sustainable and interpretable solutions. The developed algorithms will be demonstrated on industrial case with a goal of environmental emission monitoring and predictions. The proposed multi-dimensional dynamic data analytics can be extended to many other applications. The proposed data processing and analytic techniques could inspire solutions to similar problems in other fields, such as financial time series, dynamic system identification, and machine learning for dynamic feature analysis.