Automated defect detection for Fluorescent Penetrant Inspection using Random Forest

Abstract

Fluorescent Penetrant Inspection (FPI) is the most widely used NDT method in the aerospace industry. Inspection of FPI is currently done visually and diculties arise distinguishing between penetrant associated with defects and that due to insucient wash-o or geometrical indications. This, in addition to the nature of the inspection process, means inspection is largely inuenced by human factors. The ability to perform automated inspection would provide increased consistency, reliability and productivity. The Random Forest algorithm was used to detect defects in a number of at titanium plates which had been processed with FPI and photographed to produce digital images. This method has demonstrated the ability to correctly distinguish between defects and other non-relevant indications with accuracy comparable to a human inspector with a very small number of training examples. These results show the potential for the Random Forest algorithm to be used to detect defects in aerospace components, allowing the entire FPI line to become autonomous.

Performance Based Modifications of Random Forest to Perform Automated Defect Detection for Fluorescent Penetrant Inspection

“…Previous work has investigated the use of the established Machine Learning method Random Forest (RF) to perform automated defect detection for FPI. Results of this work found that RF was able to correctly identify 76% of defects with a false call rate of 0.42, demonstrating capability comparable to that of a human operator [3]. Whilst this work showed promising results, in order for this method to be successfully implemented in industry it will be necessary to improve the results to provide a higher true positive rate and lower false call rate.…”

“…This section has provided an overview of the methodology more thoroughly laid out in Sect. 3 of [3].…”

“…Previous work considered the use of Random Forest (RF) to perform automated inspection of a number of test pieces which had been processed using FPI [3]. A brightness threshold was first performed on the images to obtain regions of interest (ROIs) which formed the training data set.…”

“…The final subset is used to analyse the overall performance of the reduced RF. An RF of 100 trees is trained using the 21 features laid out in [3] and the images from data subsets A and B. Subset C is then passed down the RF. Each of the six performance evaluation methods in Sect.…”

“…2.1), Machine Learning specifically RF (Sect. 2.2) and an overview of the proceeding work, using RF for FPI [3] (Sect. 2.3).…”