Multisensorial self-learning systems for quality monitoring of carbon fiber composites in aircraft production

Currently in the aerospace industry is a change from the traditional nest production towards an automated mass production, which requires the fulfillment of very high demands on precision and availability of the new technologies. In a research cooperation between industrial companies and research institutions skills were developed for an automated production of carbon fiber composite parts for aircraft industries. A major challenge of the project was the enormous demands on the quality and the necessary evidence. While in automotive industries quality control will normally be carried out by samples, in the aircraft industry a 100%-monitoring is necessary to assure the safety of the aircraft and to reduce financial risks for the involved enterprises. The manufacturing process of assembling the wing-cover components made of carbon fiber composite-titanium- aluminum-structures consists primarily of the operations of drilling, reaming, removal of temporary rivets and placing the final rivets. With about 5,000 rivets the correct and flawless insertion of the holes is extremely important. Therefore a core area of the project was the line quality monitoring of a damage-free drilling of carbon-fiber composites. This included the automated measurement of all holes with respect to diameter, roundness and chamfer angle, but also the identification of possible delamination between layers of carbon fiber composites. Since the potential defects are not detectable using a single physical system, a platform for an open-system combination of various physical processes such as thermal imaging sensors, image processing and laser scanning systems was developed. In this way it is possible to compensate the disadvantages of the individual systems through the added properties of the corresponding systems. Via the developed system, the relevant features can be extracted from the raw data and fed into a generic rating classification system and a learning process. The addition of other sensor systems, such as eddy current or ultrasonic sensors, is possible at any time. The solution is based on a tripartite division of the data processing: data recording from almost any data acquisition system, data preprocessing for extracting the relevant features from the raw data and the classification system that performs the mapping of the data for the predefined and learned classes. As mathematical basis matrix operations, methods of exploratory data analysis (i.e. component analysis) and different classification methods (MLDA, MLP) are used. The paper shows the mathematical and technical procedures, explains the experimental conditions and shows the obtained results in detail.