Description
TitleStatistical learning for lightweight materials manufacturing
Date Created2019
Other Date2019-10 (degree)
Extent1 online resource (xvi, 170 pages) : illustrations
DescriptionLightweight materials such as Carbon fiber-reinforced polymer (CFRP) composites and aluminum alloys have been increasingly used in many industries for their high strength-to-weight ratio. However, due to their high cost, many industries, especially the automobile sector, usually mix dissimilar materials such as joining CFRP with aluminum or magnesium alloys. Friction stir blind riveting (FSBR) and micro friction stir welding (µFSW) are two solid state joining processes that have been recently extended to or developed for joining dissimilar lightweight materials, including CFRP and aluminum alloys. The overarching goal of this dissertation research is to contribute to these two advance joining techniques, FSBR and µFSW, from a process analytics perspective. For FSBR, the specific objective is to develop new methods in sensor fusion and process monitoring to enable in-situ non-destructive evaluation (NDE). For µFSW, the specific objective is to develop a new method to characterize and monitor the dynamic tool wear propagation in µFSW.
In the first part of this dissertation, a systematic data-driven approach is developed for monitoring FSBR and evaluating the quality of CFRP joints using in-situ sensor signals. The proposed method extracts hidden information from the multi-sensor, high-dimensional, heterogeneous in-situ signals by unsupervised tensor decomposition; informative features are selected by sparse group lasso; the process is then monitored in real-time through the selected features by classifier fusion. A case study with FSBR experiments demonstrates the effectiveness of the proposed method in real-time evaluation of CFRP joint quality.
Since the number of training samples is usually limited in manufacturing experiments due to time and cost, the second part of this dissertation focuses on addressing the small sample size problem in data-driven quality evaluation. A novel in-situ NDE method is developed by integrating tensor decomposition and ensemble learning. Regularized supervised tensor decomposition extracts discriminant features from in-situ signals; ensemble learning is adopted to stabilize the tensor decomposition and provide better quality evaluation results. The proposed method optimizes the integration of tensor decomposition and ensemble learning by developing a novel diversity-based feature generation and selection approach: a diversity measure is defined to evaluate the extracted features; a heuristic adaptive algorithm is developed to determine the optimal parameters for integration; optimal features are selected via clustering to maximize the diversity measure, which is expected to strengthen ensemble learning performance. Numerical studies and case study are performed to demonstrate the superiority of the proposed method over existing methods.
Tool condition in manufacturing plays a significant role on process dynamics and part quality. Effective modeling and monitoring of tool condition deterioration can provide the technical basis for maintaining production efficiency and quality. Inspired by the need of tool condition monitoring in joining processes, the third part of this dissertation aims to model and monitor the spatial and temporal patterns in μFSW tool surface measurements. A hybrid hierarchical spatio-temporal model is developed for the time-ordered, high-dimensional tool surface measurement images to characterize the dynamic tool wear propagation in μFSW. The model is developed in a hierarchical Bayesian structure with the first level being a data-driven regression model for the high-resolution tool pin profile images and the second level being a physics-based advection-diffusion model for the welding temperature distribution. Kalman filter is adopted to estimate the posterior distributions of the state variable (temperature distribution) and the error between the measured tool surface image and the predicted images. Regularized Mahalanobis distance is proposed to monitor tool wear progression. Numerical studies on three abnormal tool wear progression patterns demonstrate the effectiveness of the proposed spatio-temporal modeling method, as well as the timeliness, confidence, and power of detection.
The methodological development in this dissertation is expected to enable near real-time non-destructive evaluation of product quality, facilitate early detection of abnormal tool wear progressions, reduce the efforts in manual inspection, and support sustainable advanced manufacturing. The proposed methods can be easily extended to other manufacturing processes with online sensing and tool measurement capabilities.
NotePh.D.
NoteIncludes bibliographical references
Genretheses, ETD doctoral
LanguageEnglish
CollectionSchool of Graduate Studies Electronic Theses and Dissertations
Organization NameRutgers, The State University of New Jersey
RightsThe author owns the copyright to this work.
Statistical Profile