Predictive design analytics for optimal system design

Ma, Jung Mok

Permalink

https://hdl.handle.net/2142/78365 Copy

Description

Title

Predictive design analytics for optimal system design

Author(s)

Ma, Jung Mok

Issue Date

2015-04-13

Director of Research (if dissertation) or Advisor (if thesis)

Kim, Harrison H.M.

Doctoral Committee Chair(s)

Kim, Harrison H.M.

Committee Member(s)

• Thurston, Deborah L.
• Allison, James T.
• Work, Daniel B.

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Data analytics for design
• Data-driven product design
• Trend mining
• Time series usage modeling

Abstract

“Predictive Design Analytics” proposed by this dissertation is a new paradigm to enable design engineers to extract important patterns from large-scale data characterized by four dimensions (volume, variety, velocity and veracity), and combine the extracted knowledge and its trend with complex systems optimization for various design decision making problems such as economical life cycle design, product family design and sustainable design. The goal of this research is the development of predictive design analytics methods for optimal systems design: Demand Trend Mining, Continuous Preference Trend Mining, Predictive Data-Driven Product Family Design, and Predictive Usage Mining for Life Cycle Assessment. To the best of the author’s knowledge, this is one of the first attempts to provide a systematic framework of predictive analytics for design, which comprises data preprocessing, data representation, predictive analytics algorithms, mathematical formulation of design problems, and design decision making. Demand trend mining (DTM) is developed to link pre-life (design and manufacturing) and end-of-life (remanufacturing and recycling) stages of a product for the improvement of initial product design. In order to capture hidden and upcoming trends of product demand, the algorithm combines three different models: decision tree for large-scale data, discrete choice analysis for demand modeling, and automatic time series forecasting for trend analysis. DTM dynamically reveals design attribute patterns that affect demands. A new design framework, Predictive Life Cycle Design (PLCD), is formulated, which connects DTM and optimal product design. The DTM algorithm interacts with the optimization-based model to maximize the total profit of a product through its life. For illustration, the developed model is applied to an example of smart-phone design, assuming that used phones are taken back for remanufacturing after one year. The result shows that the PLCD framework with the DTM algorithm identifies a more profitable product design over a product’s life cycle when compared to traditional design approaches that focus on the pre-life stage only. Continuous Preference Trend Mining (CPTM) is developed to generate multiple profit cycles of product design while addressing some fundamental challenges in previous studies. The CPTM algorithm captures a hidden trend of customer purchase patterns from accumulated transactional data. Unlike traditional, static data mining algorithms, the CPTM does not assume stationarity, but dynamically extracts valuable knowledge from customers over time. By generating trend embedded future data, the CPTM algorithm not only shows higher prediction accuracy in comparison with well-known static models, but also provides essential properties that could not be achieved with previously proposed models: utilizing historical data selectively, avoiding an over-fitting problem, identifying performance information of a constructed model, and allowing a numeric prediction. Furthermore, the formulation of the initial design problem is proposed, which can reveal an opportunity for multiple profit cycles. This mathematical formulation enables design engineers to optimize product design over multiple life cycles while reflecting customer preferences and technological obsolescence using the CPTM algorithm. For illustration, the developed framework is applied to an example of tablet PC design in the leasing market, and the result shows that the determination of optimal design is achieved over multiple life cycles. Predictive, data-driven product family design (PDPFD) is proposed as one of the predictive design analytics methods to address the challenge of determining optimal product family architectures with large-scale customer preference data. The proposed model expands clustering based data-driven approaches to incorporate a market-driven approach. The market-driven approach provides a profit model in the near future to determine the optimal position and number of product architectures among product architecture candidates generated by the k-means clustering algorithm. Unlike discrete choice analysis models which were used in previous market-driven approaches, a market value prediction method is proposed as a dynamic model which can capture and reflect the trend of customer preferences. Prediction intervals provide market uncertainties of the dynamic profit model for product family architecture design. A universal electric motors design example is used to demonstrate the implementation of the proposed framework with large-scale data. The comparative study shows that the PDPFD algorithm can generate more profit than pure clustering based data-driven models, which shows the necessity of combining data-driven and market-driven approaches. Predictive usage mining for life cycle assessment (PUMLCA) is developed to provide the usage modeling in life cycle assessment (LCA) which has been rarely discussed despite the magnitude of environmental impact from the usage stage. The PUMLCA algorithm can serve as an alternative of the conventional constant rate method. By modeling usage patterns as trend, seasonality, and level from a time series of usage information, predictive LCA can be conducted in a real time horizon, which can provide more accurate estimation of environmental impact. Large-scale sensor data of product operation is suggested as a source of data for the proposed method to mine usage patterns and build a usage model for LCA. The PUMLCA algorithm can provide a similar level of prediction accuracy to the constant rate method when data is constant, and the higher prediction accuracy when data has complex patterns. In order to mine important usage patterns more effectively, a new automatic segmentation algorithm is developed based on the change point analysis. The PUMLCA algorithm can also handle missing and abnormal values from large-scale sensor data, identify seasonality, formulate a predictive LCA for existing and new machines. Finally, the LCA of agricultural machinery demonstrates the proposed approach and highlights its benefits and limitations.

Graduation Semester

2015-5

Type of Resource

text

Permalink

http://hdl.handle.net/2142/78365

Copyright and License Information

Copyright 2015 by Jungmok Ma

Owning Collections