Volume CT Data Inspection and Deep Learning Based Anomaly Detection for Turbine Blad

Volume CT (VCT) scan is being used in various industrial and medical applications. One example of such is the inspection of turbine blades in jet engines. Even with the complete 3D data, it can be difficult to directly inspect potential anomaly regions (e.g. drilled holes) in the internal of the blade. The aim of this thesis is to design methods to help manual visual inspection and automatic detection of anomalies in turbine blade VCT data.;In the first part, an unwrapping method inspired by medical research was designed and tested on the dataset. The method applies thresholding methods to the CT scans and extracts the skeleton of the major internal cavities. The skeleton branches are then used to perform unwrapping on the VCT data to transform it into a flattened 3D representation. The result reveals the internal of the blade and allows much more convenient visual inspection.;The second part then explores deep learning techniques to automate anomaly detection. This chapter begins with an introduction to fundamental ideas of deep learning and popular frameworks, followed by a discussion of dataset composition and training/validation/testing set partitioning strategies. Finally, four Convolutional Neural Networks (CNN) were designed and trained to identify a specific type of anomaly. The result shows that an optimal combination of set partitioning strategy and network design allows the system to reach high accuracy in automatic anomaly detection.;In conclusion, the thesis demonstrated a novel view to inspect VCT data and investigated the application of deep learning techniques to capture 3D features in the industrial environment. Both demonstrate great practical value and may inspire and stimulate more discussions and researches in this area.