Abstract
In recent years, manufacturing has pushed the boundaries on what is possible and practical to produce. New technologies, such as additive manufacturing, have enabled the creation of intricate, specialized parts. Additionally, new optimization methods such as machine learning have further strengthened traditional manufacturing by reducing the unnecessary waste of time, energy, and money. However, these advancements have not been without their drawbacks. The increased technological complexity seen in additive manufacturing results in parts that are difficult to accurately predict and can have internal defects on many different size scales. This often necessitates expensive and slow post-production inspection or treatment before they can be certified. Additionally, optimizing traditional manufacturing and utilizing complex machinery has resulted in processes that can be hard to monitor and vary little from run to run. This obstructs the development of robust predictive models that could otherwise be used to investigate alternatives or adaptations of current processes.
This dissertation encompasses techniques and methods to mitigate obstacles introduced by these advancements in manufacturing while enabling their advantages. The first research thrust addresses defect detection in advanced manufacturing, while the second research thrust addresses challenges that may be faced in the utilization of neural networks in advanced manufacturing and other physical systems. In the first part of research thrust one, a stereoscopic camera array coupled with three-dimensional digital image correlation (3D-DIC) technology was utilized to monitor an additively manufactured part in-situ. Using 3D-DIC, the 3D topology of each layer was measured, which was then compared against the expected topology created through a simulated virtual print. This technique allows for the detection of defects both in and out of plane, in situ, and in real-time. In the second part of this research thrust, a novel data preprocessing method was developed to enable the utilization of 3D-DIC to make fine measurements in the presence of unstable lighting or large part deformations. This method creates simulated defects with random sized noise in a generic self-supervised manner, eliminating the need for large datasets of labeled, pre-failed parts. This preprocessing method was used with a proven state-of-the-art image classification neural network and managed to detect localized cracking in an object undergoing an expanding plug test, despite a low signal-to-noise ratio.
The second research thrust focuses on investigating and compensating for the relatively small and tightly clustered data found in advanced manufacturing contexts. These data characteristics leads to training datasets that are not representative of the data their models will be used to predict. Along with incorrect predictions resulting in significant real-world loss, this has helped delay the full adoption of neural networks and advanced machine learning in the manufacturing industry. The first part of this second research thrust investigates the effect that dataset size and distribution can have on the ability to create accurate neural network based prediction models. Seventy-two different machine learning models were trained using various dataset sizes sampled from three physical scenarios of varying complexity. This work showed that while the commonly held belief that more data can create better predictors holds true, the underlying scenario's complexity and the prediction variable's distribution significantly impact how much data can be considered enough to train an adequate model. Other difficulties arising from these types of datasets, such as convergence to local minima, were also highlighted. In the second part of this thrust, two deep learning architectures were created, focusing on data derived from manufacturing scenarios and other physical datasets. The first is a predictor-critic network, which utilizes basic a priori knowledge about the underlying system (such as systems with a constant consumption over time will have a higher total consumption with higher total durations) to create a critic network to guide the fitting of a predictor model. The second is a quasi-adversarial network that helps guide the fitting of a predictor model through a discriminator, which attempts to learn data trends and then discriminate between random predictions, predictions generated by the predictor model, and actual answers from the training dataset. These models outperformed traditional machine learning and basic feed-forward networks at the cost of increased training time and complexity.
