AVRRC Specialised Test Facilities

The AVRRC undertakes a range of research activities ranging from visual inspection to identification of defects and other issues during wing manufacture and in-service. To support this research the AVRRC have obtained a section of A320 Wing taken from a retired operational aircraft. This provides us with a representative real-world example of an actual wing. Research includes the use of AI and other visual inspection techniques to test the effectiveness of these techniques against more traditional human based inspection.

Detailed inspection of large, manufactured components such as aircraft wings and other complex structures is necessarily a highly intensive manual process requiring careful inspection, checking for missing or misaligned components, out of tolerance sub-components, surface defects, damage to paint surface, poor bounding of components, etc. Depending on the inspection requirements this can take a significant time, and, in many cases, there cannot be any margin for error.

AVRRC are developing several automated visual inspection systems ranging from AI (machine learning) to more conventional image-based inspection techniques.

However, one of the major challenges with any Machine Learning solution is the requirement for vast amounts of training data. This can be difficult if there are very few (if there are any manufacturing defects). Also, it is not possible to use the many libraries of labelled images in the public domain since these are not representative and in any case care has to be taken when using open-source AI tools because of commercially sensitive requirements. The AVRRC have been investigating the use of multi-modal sensory solutions in conjunction with AI (Machine Learning) to determine whether an AI system can be trained to recognise defects of various kinds. This research has spanned traditional and novel imaging systems along with new AI training methods to see if it is possible to replace/augment the human inspection process.

The image above illustrates a typical automated AI inspection process used in the AVRRC. This has involved the use of an Nvidia DXG Station along with edge computing solutions to act as inference engines. Furthermore, to reduce the requirements for massive test data sets based on real defects novel techniques have been investigated such as generating virtual or synthetic images of possible defects and using these as part of the training activity. Results are extremely promising but remain commercially sensitive. Further research is underway to build on the automated inspection technique with additional focus on different machine learning algorithms.

Other non-AI based visual inspection solutions have been shown to work in certain circumstances. The example shown here is where a reference image of a good component can be used with image processing to show various manufacturing anomalies. The novelty here is that the camera alignment does not need to be in the precise position/orientation of the camera used to take the reference image. This technique can be extremely fast and can operate in real-time. 

These form a single large display area of (bezel-corrected) 6160x1080 pixels through NVIDIA Surround graphics drivers. A single 22" (1920x1080) LCD touchscreen provides flight (instrument) displays.