Development of a Classification Algorithm Based on Graph Neural Networks for CAD Gear Assemblies as a Basis for a Rule-Based Ontology

Kurzfassung

Computer-Aided Design (CAD) systems are indispensable tools in modern engineering, enabling the creation of complex mechanical assemblies such as gear systems for manufacturing, simulation, and design analysis. These assemblies are critical components in various industries, including automotive and aerospace [SRN08, p.25]. As the timeto-market pressure increases [CRL05] so does the challenge of ensuring reliability and performance of their real-world application in the design phase, respective to conditions such as thermal and mechanical stresses. However, exhaustively simulating all potential real-world scenarios is computationally prohibitive due to the combinatorial complexity of all operational variables involved [ML22]. Therefore, developing a methodology to distinguish propagation paths of errors caused by stresses that are function-critical to the assembly, compared to cases where error propagation does not compromise functionality (e.g., a bearing wearing out due to excessive meshing force between gears versus a non-critical issue such as a screw loosening in the housing), is a non-trivial task. The process of identifying such critical paths during the design process for further analysis (e.g. simulation) often requires extensive domain expertise and manual eort. Recent advancements in Artificial Intelligence (AI), particularly in Graph Neural Networks (GNNs), oer promising solutions for automating and enhancing this process. Graph structures excel at modeling relational data, making them well-suited for representing and analyzing CAD assemblies [SMV20]. In such representations, nodes can represent individual parts (e.g., gears, shafts, bearings), while edges capture geometric and functional relationships (e.g., contact, distance, constructional constraints). By generating graphs based on parametric data based on native CAD files, which inherently encapsulates the design procedure history and therefore the domain-specific knowledge of its engineer, it becomes possible to represent assemblies in a structured manner while preserving their functional semantics. These graphs, initially naive to the data, can be enriched with external sources of physical information, such as material databases, and integrated with rule-based ontologies that capture mechanisms like implicit degrees of freedom. For instance, a gear mounted on a shaft, while fully constrained as an individual part, exhibits a rotational degree of freedom when considered within the assembly context. Such enriched graphs enable comprehensive design analysis, validation, and optimization. However, the integration of ontologies, databases and simulation data are tied to the underlying part classes. 1. Introduction 2 1.1. Research Challenges and Problem Statement Transitioning from an information-naive, data-based graph to an information-aware graph representation of an CAD assembly that is suitable for analyses and reasoning, therefore requires robust methods for accurately identifying and classifying mechanical parts based on the data. This is where GNNs play a crucial role, as they enable the classification of nodes (as parts) within the information-naive graph, thereby enabling the connection to these enrichment sources. Despite the potential of GNNs in this domain, research gaps remain, particularly in the eective use of parametric CAD data (further just called CAD data). While previous studies have explored shape-based methods like Boundary Representation (B-Rep) [MB22] or Mesh [WSK15] for part classification, these approaches lack domain-specific insights and overlook the rich parametric information embedded in CAD data. This thesis aims to address this research gap by developing a robust and accurate classification algorithm for CAD gear assemblies represented as graphs using said data. Additionally, it explores the integration of B-Rep data into the graph-based classification process to improve accuracy and robustness further. 1.2. Thesis Objectives and Scope Embedded within the broader project scope, this work aims to create fully classified graph representations of generic CAD gear assemblies. These representations are designed to be seamlessly connected to a rule-based ontology, enabling advanced error propagation analysis and enhancing the design process. The tangible objectives of this thesis are threefold. First, it seeks to develop a methodology for constructing graph representations of CAD gear assemblies using CAD data and B-Rep data. Second, it aims to maximize the classification performance on various GNN architectures with respect to classification metrics separately on both types of data. Finally, it explores hybrid approaches for combining both types, including a hierarchical GNNs and confidence-based classification, to achieve optimal results in terms of accuracy and robustness.