Automated defect detection for fluorescent penetrant inspection using machine learning

Fluorescent Penetrant Inspection (FPI) is a well-established NDT method used widely in the aerospace industry. The nature of FPI inspection, particularly in new manufacture, can lead to variable results influenced by human factors. This has led to the desire for an automated inspection system which would improve reliability, however, there is a lack of available training data. This, combined with accumulating surface penetrant due to roughness, geometry, insufficient wash off, etc. makes it challenging to train an automated inspection system capable of distinguishing penetrant which has arisen from defects from other non-defective indications. A dataset of test pieces containing cracks induced using thermal fatigue was manufactured, processed and imaged to obtain a dataset of 99 test pieces with 173 cracks. This dataset was used to train two Machine Learning (ML) algorithms. The first was the Random Forest (RF). A number of features were manually selected and used to train a number of decision trees with a subset of the full dataset. The RF method was able to detect 76% of defects with a false call rate of 0.42 per image. Subsequent work used performance-based methods to improve the RF. This work found that it was possible to remove poorly performing trees from the RF without affecting the results, allowing for a more computationally efficient solution. The second ML technique used was deep learning. The ResNet50 architecture was found to exceed the performance of the RF method. ResNet50 was able to detect 95% of the defects with a false call rate of 0.15. Moreover, it was also found it is possible to improve results by providing additional training data from non-defective indications which is quicker and cheaper to obtain. The work presented in this thesis demonstrates the feasibility of using ML techniques to perform automated defect detection for FPI.