An ensemble machine vision system for automated detection of surface defects in aircraft propeller blades

Abstract

Visual inspections comprise the majority of inspections for large transport aircraft. The traditional inspection process is time-consuming, inconsistent, and subject to human errors. Automated defect detection systems have been developed to leverage computer vision and deep learning to decrease inspection times and improve detection performance. Prior methodologies have used convolutional neural networks to detect defects from image data. The performance of these methods is insufficient for critical aircraft inspection. Furthermore, the tradeoff between error rates of false alarms and missed detections has not been well addressed in the literature. This thesis presents a novel application of deep learning ensembles to automated aircraft visual inspection and provides a methodology for using ensembles to manage the error rates of the algorithm. Stacked ensembles are constructed from three deep learning base learners and a logistic regression meta-learner is used to combine their predictions. The performance of the stacked ensembles is evaluated, and it is found that the stacked ensembles outperform the current state-of-the-art defect detection approaches. Furthermore, it is shown that with sufficient error diversity, ensembles can be constructed to eliminate the missed detections that may lead to critical failures in aircraft inspection.