Search

This thesis proposes three big data-predictive control (DPC) approaches that integrate theoretical development and practical applications of behavioral systems theory with machine learning tools. This integration allows for using extensive historical process data effectively in designing control systems for processes with complex dynamics. The first approach develops a multi-level DPC for linear multi-timescale processes, utilizing data resampling and a novel design that features multi-level optimization horizons to approximate the dynamics of different timescales. This strategy expands the horizon length exponentially with optimization steps, thereby facilitating the simultaneous consideration of near-to-distant future dynamics during online control. A trajectory-based dissipativity condition is also formulated to ensure stabilizing tracking control while achieving disturbance rejection. The second strategy focuses on controlling nonlinear processes with a single dominant timescale by introducing the Dynamic Latent Variable Autoencoder (DLVAE). The DLVAE transforms nonlinear physical space into a linear latent variable space, enabling direct control of physical nonlinear process variables through their corresponding linear latent variables. Based on the behavioral systems theory, the proposed data-driven control framework does not require knowledge of the causality of the latent variables. The robustness of this method is ensured by incorporating sensitivity bounds between the latent and physical variables into the stabilizing control design using dissipativity theory. This integration allows stabilizing control of nonlinear processes with a dominant timescale, entirely within the linear latent space. The third approach presents a multi-block DLVAE strategy to control a nonlinear process with multi-timescale dynamics. Each DLVAE projects the nonlinear dynamics across different timescales onto corresponding linear latent spaces, which are trained on low-pass filtere

Haze conditions have reportedly reduced visibility to about 3km in some of the busiest shipping lanes in the world. Haze conditions, including inclement weather conditions, are identified as a key challenge for autonomous vehicle operations. However, field data on poor weather conditions and ship images under hazy conditions may not be readily available to support research work aimed toward overcoming such challenges for autonomous vehicles. In this thesis, synthetic ship images are rendered under hazy conditions to augment a baseline dataset of haze-free ship images, in order to support our research on ship vessel classifications in a hazy environment using machine learning. The proposed feature extraction involves the counting of corner points detected using the Kanade Lucas Tomasi (KLT) technique to characterize the pattern of specific ship classes and computing of higher-order moments on the color planes on the ship structure detected in the images. Results show that the average ship classification accuracy rate is about 40% higher when the model is trained using a dataset augmented with synthetic hazy ship images; the classifier can classify for ship classes such as container ships, cargo ships, and sailing vessels, with an 80% average accuracy rate.

Haze conditions have reportedly reduced visibility to about 3km in some of the busiest shipping lanes in the world. Haze conditions, including inclement weather conditions, are identified as a key challenge for autonomous vehicle operations. However, field data on poor weather conditions and ship images under hazy conditions may not be readily available to support research work aimed toward overcoming such challenges for autonomous vehicles. In this thesis, synthetic ship images are rendered under hazy conditions to augment a baseline dataset of haze-free ship images, in order to support our research on ship vessel classifications in a hazy environment using machine learning. The proposed feature extraction involves the counting of corner points detected using the Kanade Lucas Tomasi (KLT) technique to characterize the pattern of specific ship classes and computing of higher-order moments on the color planes on the ship structure detected in the images. Results show that the average ship classification accuracy rate is about 40% higher when the model is trained using a dataset augmented with synthetic hazy ship images; the classifier can classify for ship classes such as container ships, cargo ships, and sailing vessels, with an 80% average accuracy rate., Major, Republic of Singapore Navy, Approved for public release. distribution is unlimited