Towards an adaptive and data-efficient framework for deep learning feature recognition

Abstract

Engineering design can be described as creating and manipulating geometry that influences the physical world and fulfils needs. Key to this is abstraction; engineers take a top-down approach since many specific forms can fulfil a function to varying degrees. In modern computer-aided design (CAD), the geometry is represented digitally and abstraction is difficult on the basic entities - vertices, faces, etc. This thesis is focused on the process of mapping these primitives to more useful abstractions for automated decision-making - termed feature recognition. Critically, features of interest are context dependent, even given the same geometry. Traditional rule-based approaches and current deep learning approaches struggle against changing contexts and novel geometries and features. This thesis presents work towards a data-efficient framework that can adapt to novelty and produce neural network (NN) instances trained for a given geometric feature recognition task. A point-based NN approach is presented as the core pattern recognition component. Often stated drawbacks of point-based approaches are addressed such that it is competitive with other approaches while preserving potential application to other 3D geometry representations. An empirical survey of NN uncertainty estimation techniques for the current setting is also presented. This capability allows the system to quantify its lack of knowledge and enables the decision-making underlying the adaptability to novelty. Two approaches are presented for training a NN given different data-scarce scenarios of engineering relevance. First, an active learning workflow for iteratively curating the minimum amount of labelled data is formalised and validated – leveraging the uncertainty estimates as a surrogate for unseen accuracy. Second, a real-time workflow is investigated where the system learns from iterative and partial labelling of large CAD geometries. This is work towards a collaborative interface for accelerating manual tagging with no pre-training or previous data curation. Overall, a system-wide view is presented with proof-of-concept components being investigated and validated. A reduction in user effort has been suggested by the results and the work stands as a guide for further development of deep learning approaches in the space of feature recognition for computer-aided engineering.

More information

Identifiers

ORCID for David Toal: orcid.org/0000-0002-2203-0302

ORCID for Andy Keane: orcid.org/0000-0001-7993-1569

ORCID for Xu Zhang: orcid.org/0000-0002-6918-1861

Catalogue record

Export record