Fluorescent Penetrant Inspection (FPI) is a popular Non-Destructive Testing (NDT) method which is used extensively in the aerospace industry. However, the nature of FPI means results are susceptible to the effects of human factors and this can lead to variable results, making automation desirable. Previous work has investigated the use of established machine learning method Random Forest to perform automated defect detection for FPI. Whilst good results were obtained, there was still a significant number of false positives being identified as defective. This paper presents work done to investigate the potential of using deep learning methods to perform automated defect detection.

A dataset was obtained from a set of 99 titanium alloy test pieces with cracks induced using thermal fatigue loading. These test pieces were repeatedly processed and using data augmentation a large dataset was obtained. This data was used to train a ResNet34 and ResNet50 architecture as well as a Random Forest. Two significant results were obtained. Firstly, the ResNet50 is able to create a network capable of detecting 95% of defects with a false call rate of 0.07. This result far exceeded that obtained using the Random Forest method despite both methods only having access to a small dataset. This demonstrated the strong capability of deep learning architectures. The second result was that increasing the amount of data obtained from non defective regions significantly increases performance. This result is encouraging as this data, obtained from non-cracked parts, can be quickly and cheaply obtained by reprocessing test pieces.