Deep learning-based defect detection in industrial CT volumes of castings

Abstract

Industrial X-ray computed tomography (CT) has proven to be one of the most powerful non-destructive testing (NDT) methods for the inspection of light metal castings. The generated CT volume allows for the internal and external geometry of the specimen to be measured, casting defects to be localised and their statistical properties to be investigated. On the other hand, CT volumes are very prone to artefacts that can be mistaken for defects by conventional segmentation algorithms. These artefacts require trained operators to distinguish them from real defects, which makes CT inspection very time consuming if it is to be implemented on the production line. Foundries using this inspection method are constantly looking for a module that can perform this interpretation automatically. Based on CT data of aluminium alloy automotive and aerospace specimens provided by industrial partners, an automated approach for the analysis of discontinuities inside CT volumes is developed in this paper based on a two-stage pipeline: 2D segmentation of CT slices with automatic deep segmentation using U-Net to detect suspicious greyscale discontinuities; and classification of these discontinuities into true alarms (defects) or false alarms (artefacts and noise) using a new convolutional neural network classifier called CT-Casting-Net. The choice of each model and the training results are presented and discussed, as well as the efficiency of the approach as an automatic defect detection algorithm for industrial CT volumes using metrics relevant to the field of non-destructive testing. The approach is tested on six new CT volumes with 301 defects and achieves an object-level recall of 99%, a precision of 87% and a voxel-level intersection-over-union (IoU) of 62%.