Towards automated damage detection for in-situ remanufacturing

Remanufacturing is the restoration of used products to a like-new condition, with warranty matching the original product. It is labour intensive, and faces problems of increasing labour costs, labour shortage, and varying quality of work between individuals, but can be solved by automation. However, in the marine and offshore industry, the large parts are impractical to disassemble and transport to remanufacture offsite, and hence it should be carried out in-situ. With automated in-situ remanufacturing, there is increased safety, efficiency in material use, reduced turnaround time, and in turn reduced costs, and is the motivation for it. However, there is a lack of a formal framework for such remanufacturing to ensure repeatability and accountability, and is first established in this thesis. A laborious remanufacturing process that benefits from automation is the inspection process, specifically damage detection and classification. It is preferable to automate damage detection from 3D data over 2D images, as the 3D data of the detected damage region is required for the hybrid manufacturing process in the downstream repair procedure. Although the ideal 3D damage detection method is comparing the scanned 3D model to the as-built model, some as-built models of legacy parts may be missing, whilst other damaged parts may differ from the as-built model due to undocumented modifications, or allowable deformations. As such, this research focuses on automating damage detection from only the 3D scan data of the part, for use in downstream repair processes. The unoptimized computational efficiency for geometric learning, and lack of training data inhibits the direct application of machine learning for damage detection from 3D data. A literature review reveals that current 3D damage detection methods are unable to detect different damage types, have assumptions on the geometry of the undamaged part, or require additional inputs, representing a significant gap within the knowledge, and is filled by this research. Selected algorithms for 3D damage detection reviewed are implemented, alongside several proposed novel algorithms for dimensionality reduction. The extracted features are then processed via machine learning, with different machine learning models compared. The models were trained and demonstrated on synthetic scan data due to the limited access to real-world data, followed by validation of the trained model on real scan of physically damaged parts. The proposed approach can detect different types of damage from 3D scan data without assumptions on the undamaged part geometry or requiring additional inputs, with a sensitivity of 95.9%, representing a substantial improvement over current 3D damage detection methodologies. Based on the visual results, damaged regions are largely identified, although the proposed approach has difficulties identifying the exact borders of the damage region. The same approach is also implemented to classify the damage type, attaining an accuracy of 73.2%. However, it should be noted that the proposed approach may wrongly identify certain intended features as damage, such as debossed text, and is also sensitive to noise in the data. While there are other enabling technologies required to fully automate in-situ remanufacturing, this thesis only focuses on the processes leading up to the damage detection and classification.Doctor of Philosoph