| This thesis investigates the use of latent variable models for the analysis of large datasets, and the development of novel techniques for the design of robust manufacturing processes and new products. The thesis consists of two distinct sections: (i) Latent variable methods for the development of robust manufacturing processes, and (ii) developing novel applications using latent variable and machine learning methodologies to learn from sound and vibration signals.; In the first section, a new paradigm on how to achieve robustness by design is presented. The new framework involves two main steps: the first is a learning task where methods to identify and quantify the effect of disturbances on the product quality are developed and new multivariate measures of the degree of robustness of industrial processes are presented. In the second step, latent variable models are used in an optimization framework to define new raw material specifications and to design new operating conditions for batch and continuous processes. A new technique using derivative-augmented models that describe the sensitivity of product quality to disturbances has been developed. The design approach is formulated as an optimization problem to achieve an optimal desired product quality that's insensitive to historical disturbances. The design methodology, which allows the inclusion of constraints, is applied successfully on two industrial processes with different degrees of complexity.; In the second section, a methodology to extract knowledge from sound and vibration is proposed and applied to develop two patent-pending applications. The framework consists of four steps: (i) signal acquisition, (ii) feature extraction, (iii) feature reduction using latent variable methods, and (iv) predictive modeling.; In the first application, an innovative method is proposed to measure surface roughness of machined surfaces in real time. The method consists of an apparatus that directs air at high pressure to machined surfaces; the output signal is then analyzed using wavelet and PLS to predict surface roughness. The method is simple, non-contact, relatively cheap, and robust to outliers. The method can be seen as an evolutionary step in measuring machining surfaces. The second application aims at developing a method to detect/classify knee osteoarthritis in human patients using vibration signals measured by placing accelerometers on the knee joints. Features are extracted from the vibration signals using wavelets; the results are presented in reduced latent variable space and the patients are then classified by Support Vector Machine (SVM) according to the type of their knee problems. Partial Least Squares (PLS) is also used to relate these vibration features to joint characteristics measured from Magnetic Resonance Imaging (MRI). MRI is a diagnostic tool that is considered to be the gold standard in medical imaging of the knee.; Throughout the thesis, Multivariate statistical methods such as Principal Component Analysis (PCA) and Partial Least Squares (PLS) are shown to be key tools for extracting useful information from large datasets and for dealing with the high correlation found in both multivariate industrial processes and sound/vibration signals. |
