AbstractSurface reconstruction, or reverse engineering of a surface, is the practice of combining data collection and analysis techniques for retrieving measurement data from an object with the aim of generating a digital representation of the object, where the original is either missing or incomplete. Recovering the digital representation of a scanned physical shape will, in most cases, be in the form of a three-dimensional (3D) point cloud. However, the process of capturing the data will inevitably lead to a point cloud contaminated by the uncertainty of measurement and imperfections in the measured object. To reverse engineer an object, especially a used one, consideration must be made for any manufacturing defects (tolerances or imperfections) and in-service damage. Given the significant challenges in designing a surface reconstruction algorithm, this project aims to develop a framework for reverse engineering/ performing surface reconstruction for damaged objects and develop a method for identifying and localising damage on the surface of a model with an accuracy of a few tens of micrometres. This accuracy level is significant due to the resolution of the generated data.
The workflow for reverse engineering considered in this thesis is capturing point cloud data (PCD) using a laser scanner mounted on an Articulated Arm Coordinate Measuring Machine (AACMM). Preprocessing of the raw points is performed to mitigate noise, sparse data, and contamination from the measuring system and identify properties that influence the data file structure, such as interoperability. Before performing detection, it is necessary to identify geometric features from the PCD, since an artefact's “nominal” design geometries are unknown. Hence, a linear fitting method is applied to the input point cloud data, and the system is optimised to obtain a good level of robustness of fitting the line of best fit for estimating “nominals”. A combination of edge detection, profile monitoring, model sectioning, and machine learning are used to identify potential damage locations on a model’s surface.
The processes are ideally designed to be reproducible, so with reduced human influence. Slices produced from model sectioning of the PCD are unwrapped to create planar (X-Y) data at the micrometre-level, suitable for using machine learning (ML). The ML is trained and tested to classify the spatial-sequences of the data as either “good” or” damaged”. When applied to a potentially damaged surface, the sequence classification can then identify the location of damage along the height axis of a model. Future work would generalise this into three dimensions.
The performance of the proposed system is evaluated through mathematical simulation of noisy point cloud data on objects with simulated defects and by physical validation using measurement data. The results indicate that the method can detect and localise areas of damage. This is a stepping-stone to automated repair of damaged surfaces by adaptively reproducing a copy of the original part, considering deviation of the “as manufactured” from the design intent. The thesis contribution to knowledge is a novel method of damage detection and localisation at the micrometre level as a step towards implementing a surface reconstruction framework, which has also been developed in this study.
|Date of Award||2023|
|Supervisor||Andrew Longstaff (Main Supervisor) & Simon Fletcher (Co-Supervisor)|