Date of Award
8-1-2026
Degree Name
Doctor of Philosophy
Department
Mechanical Engineering
First Advisor
Chu, Tsuchin
Abstract
A fourth industrial revolution is currently taking place in the manufacturing sector, called “Industry 4.0,” where material production and fabrication processes are becoming fully automated using integrated sensors, closed-loop feedback mechanisms, and artificial intelligence/machine learning (AI/ML). The nondestructive evaluation (NDE) field has begun to follow suit by integrating AI/ML techniques in material inspection processes to automate defect detection for many NDE methods. Since material inspection processes are paramount in ensuring safety and reliability in critical manufactured products, it is important that new AI/ML techniques are trustworthy for NDE data analysis. The current approach to building trust in new AI/ML methods is through investigative research, which this dissertation intends to contribute to.The research presented in this dissertation applies various AI/ML techniques for automated data analysis for two different NDE methods. In the first portion of research, described throughout Chapter 2, shallow convolutional neural networks (CNNs) are trained and tested for classification of in-situ optical tomography (OT) images of selective laser melting (SLM) samples containing seeded (intentional) defects. The in-situ OT sensor captures images of the SLM process, thus providing insight on the behavior of the melt pool during fabrication. The images were first pre-processed by cropping each of the defective samples, stacking the layerwise images, and applying a squared difference operator to elucidate defective indications in the images. The processed images were then labeled as either “defective” or “nominal” based on whether defects were present in the image, and image augmentation techniques were applied to balance each class in the training set. The images were then used to train and test seven variant shallow CNN architectures for binary classification. Results showed that each of the models were able to achieve classification accuracy of over 90%, with five of the models achieving an accuracy greater than 95%. Investigating some of the misclassifications revealed that edge cases, where seeded defects were first appearing or disappearing in their layerwise progression, were problematic for the models. The second portion of research, described in Chapter 3, investigates the application of support vector machines (SVMs) for automatic classification of porosity severity in cast aluminum components. This track of research was inspired by current standardized inspection protocols for cast aluminum materials, where inspectors visually compare production radiographs to reference images to determine the severity of porosity in the material. To carry out this research, the eight standard digital reference radiographs provided by ASTM E2422 for gas porosity (round) were used to test and train various types of SVM classifiers. First, feature extraction was performed on each of the eight images, which depict varying degrees of porosity, to build a feature dataset for training and testing the SVMs. Nine statistical features were extracted from each horizontal row of pixels in each image. Two separate training approaches were used, with the first splitting the original feature dataset using an 80/20 split for training and testing, respectively. The second training approach employed the original feature dataset for training, while the testing dataset was created by rotating each of the reference radiographs by 90° and repeating the feature extraction process on each row of pixels, thus creating a separate dataset for testing. Results from the first training approach showed that the quadratic SVM (One-vs-One) achieved the highest test accuracy of 92.52%, while most other models were able to achieve accuracy values greater than 90%. Results from the second training approach showed that the fine Gaussian SVM (One-vs-All) was able to achieve the greatest accuracy at 86.75%, while all other models yielded accuracy values greater than 80% in testing. The slight drop in accuracy from the second approach was expected since the test dataset was distinctly different than that of the training dataset. The research presented in this dissertation provides valuable examples of how AI/ML techniques can be applied to various types of NDE data for automated analysis and classification of defects. By implementing automated approaches for NDE data analysis, the risk for human subjectivity during the inspection process can be eliminated. However, it is important to understand that trained AI/ML models will only be as good as their training datasets allow them to be. Therefore, with a diverse amount of training examples, robust AI/ML models can be trained to perform highly accurate inspections, ultimately reducing the time and effort required by inspectors to carry out NDE protocols.
Access
This dissertation is Open Access and may be downloaded by anyone.