Model and Data-Based Approach for Leakage Localization using Multiple Flow Sensors for Processing of Fiber Reinforced Plastics

Kurzfassung

Composite fibre reinforced plastics of superior quality are manufactured for the usage in the aviation industry. To maintain the degree of quality to its highest order, a hermetic seal of the composite laminates from the environment with the help of auxiliary elements leads to the complete evacuation of the system which is also referred to as the vacuum bag. Mandatory inspection of the vacuum bag during the evacuation stage is expected to avoid impairing the properties of the composite part. Leakage identification on a later stage require more effort and resources to repair a vacuum bag. An improvement in a two stage leakage detection process developed at the German Aerospace Center in Stade is introduced. Prior research on this process describes the use of flowmeters connected to the vacuum lines to identify leak regions with the help of volumetric flow analysis and further localize them with the help of thermographic camera. In this thesis, the possibility of integrating machine learning concepts to already present sensor systems for leakage localization is examined. In a model-based approach, several experiments are conducted to understand the flow distribution inside vacuum bags of several sizes with a different orientation of several combination of flow meters. Based on the findings, it is realized that a principle of localization for a four flowmeter set up in a symmetrical vacuum bag stands true and localizes the leak to a smaller region. A data-based approach was then introduced to record flow distribution data over a set time to introduce the theory of recurring neural networks. This allows the machine to train and learn from the reference data and predict leakage locations based on the precision and the recall value chosen from the trained model. Based on the conclusions, machine learning concepts could be further refined with new architectures and constantly updating the training models to improve the accuracy of the localization process. This process could also help provide an insight to more refined information in the future.