Data-Driven Systems Engineering Techniques for Advanced Manufacturing

Abstract

This dissertation presents research performed in the area of data-driven systems engineering to address some challenges of existing sensing and modeling technologies when applied in the emerging field of advanced manufacturing. One major contribution of this work is the development of data-driven machine learning techniques utilizing novel industrial internet of things (IIoT) sensors and applying them to industrial manufacturing processes. The research covers the following three areas: an IIoT Wi-Fi based woodchip moisture estimation system for the pulp and paper industry, a novel feature engineering enhanced virtual metrology methodology for the semiconductor manufacturing industry, and a process modeling and monitoring framework utilizing IIoT vibration sensors for the process industries. In recent years, IIoT has transformed industry by changing the way industries operate from day to day. Specifically, the advent of tiny low-cost IIoT sensors and high bandwidth wireless systems means even the smallest devices can be connected, monitored, and easily communicate and share data with other devices. Thanks to these advancements, industries have gained accesses to vast amounts of high-frequency data, i.e., so-called big data. The use of computationally efficient data-driven machine learning techniques to extract valuable information from the big data has led to a significant advancement in the manufacturing industry, such as more accurate view of the operations, enhancement in scalability and performance, and bridging the gap between production floors and control systems, all leading to more efficient data-informed decision-making.

Machine learning (ML) and artificial intelligence (AI) are at the core of data-driven decision-making for advancement in smart manufacturing. The use of collected data through sensors, when processed with robust machine learning algorithms, has changed the spectrum of real-time decision-making in industries. However, more often, the rote application of ML algorithms without considering domain knowledge leads to inadequate modeling of the process such as underfitting or overfitting. These models are undesirable in production environments due to poor performance and/or robustness. Through this work, the author addresses these challenges in various industrial settings and demonstrates that the key to robust and high-performance data-driven modeling is the synergistic integration of domain knowledge and machine learning.

In the first part of this work (Chapter 2), the author proposes a non-destructive, economic, and robust woodchip moisture content (MC) sensing approach utilizing channel state information (CSI) from IIoT based Wi-Fi to address the limitations of the existing technologies in the pulp and paper industry. An experimental design and an algorithmic technique were proposed to handle the confounding factors. To address the challenge that the raw CSI data is very noisy and sensitive to woodchip packing, a feature-based classification system based on statistics pattern analysis (SPA) was proposed in this work, which shows the advantages of domain knowledge combined with machine learning instead of the rote application of ML algorithms on the raw data collected. Specifically, the CSI data collected through IIoT Wi-Fi-based sensors is processed through SPA to extract not only robust and predictive but also physically meaningful features that enables accurate estimation the woodchip MC with the help of robust ML algorithms.

In the second part of this work (Chapter 3), the author proposes a feature-based virtual metrology (FVM) framework to address the limitations of the existing virtual metrology (VM) methods. In semiconductor manufacturing, VM, also known as soft sensors, predicts wafer properties using process variables. The author explores how batch features can better capture process characteristics and dynamic behaviors than the original process variables. This work also demonstrates how FVM can inherently handle and avoid some of the tedious and time-consuming data preprocessing steps and leads to better predictive models by extracting relevant features. In addition, the author shows how non-linearity and non-Gaussianity can be handled with FVM. The FVM based approach is compared with existing VM approaches to demonstrate its superior predictive ability through a simulated industrial case study and an actual industrial case study.

In the third part of this work (Chapter 4), the author demonstrates how IIoT sensors have great potential in advancing manufacturing process modeling through yet another type of IIoT sensor - accelerometer. This is an extension to a previous work where data from a centrifugal pump IIoT test bed was used to predict the motor speed and water flow rate inside a pipe using machine-learning techniques. In this work, the author compares different levels or extent of feature engineering and examines their impact on model performance. While the modeling of motor speed is relatively less challenging after appropriate feature engineering, efficiently predicting water flow rate requires a fusion of time-domain and frequency-domain features and relatively complex machine learning techniques as the relationship is not linear. This is the main contribution of this work. Through appropriate domain knowledge and feature engineering, superior models are proposed to predict the motor speed and water flow rate in comparison to the application of machine learning techniques on the raw data. The author demonstrates the performance of the predictive models for motor speed and water flow rate and shows that approaches that integrate feature engineering with human learning through exploration achieved superior performance.

The contribution of this work and potential future directions are summarized in Chapter 5.