Developing a level-1B qualifiable CNN for in-situ ultrasonic damage classification of aerospace composite structures

Abstract

This paper examines the end-to-end development process for a Convolution Neural Network (CNN) based damage classification tool for ultrasonic inspection of aerospace-grade composite structures. The recent advent of Artificial Intelligence (AI) and Machine Learning (ML) has piqued the interest of the aerospace industry since it has the potential to improve performance and alleviate the burden on personnel. The big question in the industry right now is how and where to introduce this technology while assuring the safety and reliability of its implementation. Guidelines drafted by the European Aviation Safety Agency (EASA) for the development of AI showed that maintenance and training were the most accessible points of entry for this technology as it did not have the same stringent requirements that a flying system would have. This paper proposes a research methodology which allows for the cost-effective development of ultrasonic data for the training and testing of data-driven tools. This was partly achieved by using a novel eFlaw technique which has been implemented for the first time in composite structures. The method allows for significant augmentation and generalisation of datasets, resulting in a model with the ability to detect features potentially smaller than one-quarter of a wavelength. This improved performance paves the way for more sensitive low-frequency ultrasonic inspection in thick composites. To evaluate these models, various evaluation techniques were compared and showed that Receiver operator curves and confusion matrix-derived metrics provided comparable results. Explainable methods found that the GradCam and the inspection of feature maps showed the most interpretable results on the features that were being identified. Using the feature maps it was possible to generate a new type of C-scan, called an F-scan (Feature-scan) which provides an inspector with a view of the C-scan from the perspective of a feature map from the model providing an interpretable view of the model’s classifications. In addition to these positive results, this thesis provides readers with a cost-effective methodology for developing data-driven tools for maintenance applications within the aerospace industry.