Automated Defect Analysis using Optical Sensing and Explainable Artificial Intelligence for Fibre Layup Processes in Composite Manufacturing

In modern aircraft, structural lightweight composite components are increasingly used. To manufacture these components in a cost-effective way, robust production systems for the manufacturing of complex fibre composite components are necessary. A rather novel but already established process for fibre material deposition is the Automated Fibre Placement (AFP) technology, which automatically places several narrow, parallel fibre tows on a mould. Typically, a component consists of several, often hundreds of stacked layers of these fibre material strips. However, when these narrow fibre tows are placed in position, layup defects can occur and reduce the mechanical properties of the component. Hence, in safety critical applications, such as aircraft manufacturing, a visual inspection of every single ply is mandatory. This inspection step is currently carried out by an expert via a visual examination, which requires up to 50 % of the total production time. An automation of this inspection process using suitable algorithms offers great potential for increasing process efficiency. However, with the growing complexity of these respective defect analysis algorithms, their performance potentially increases, but the comprehensibility of the machine decision and the behaviour of the algorithm become more challenging. This is problematic especially in safety critical applications. In addition, the data quality of recorded images is influenced by the very matte, low reflective Carbon Fibre Reinforced Plastic (CFRP) material which raises the uncertainty of an inspection. To address these issues, this thesis covers the modelling of a Laser Line Scan Sensor in accordance with the European Machine Vision Association 1288 standard, with which the models accuracy and uncertainties are estimated. Besides sensor parameters, the model takes optical properties of CFRP as input parameters, which are investigated in detail in the spectral band of the used laser. Subsequently, different conventional computer vision algorithms for defect segmentation are compared, where the results indicate that defect detection scores greater than 97 % for suitable segmentation techniques are possible. In order to get a sufficient amount of image data for training and testing of more complex algorithms, a conditional Deep Convolutional Generative Adversarial Network is introduced for synthesising images of AFP layup defects from less than 50 representative original input images. Making use of this data, a Convolutional Neural Network deep learning classifier is considered very suitable for fibre layup defect classification reaching an accuracy greater than 90 % when trained with sufficiently large training data sets. A rather simplistic model­based Support Vector Machine classifier also yields high classification rates when trained with meaningful input features but a much smaller data set is needed. Combining both approaches in a parallel classification architecture allows the investigation of their machine decision­ making behaviour. This demonstrates that the features of a feature vector can be linked to the outcomes of a suitable Explainable Artificial Intelligence procedure as well as be mapped onto the raw defect input image. Hence the physical attributes of the input defect can be linked to individual image features and deep learning decisions, which enables the traceability of machine decisions and thus their utilisation in safety critical applications. The outcomes of this study are material, sensor and classification models which are especially valuable for developers and operators of camera ­based inspection devices in the aerospace sector. The results of this study can be applied for improving the efficiency and traceability of automated composite inspection procedures