Open Access

Carola Lenzen

• In the domain of assembly modeling, design engineers utilize computer-aided design (CAD) software to develop complex assemblies. However, the increasing volume and size of non-standardized part catalogs from various manufacturers pose significant challenges in the selection of suitable parts from the vast number of options. This thesis addresses these challenges by proposing a novel methodology that leverages machine learning techniques to support inexperienced designers in the assembly design process by extracting expert knowledge from previous assemblies. Specifically, it investigates the use cases of global part recommendation, localized part recommendation and handling anomalies in assemblies. The first contribution is a generic, data-driven approach to extract patterns of proven part combinations across multiple real-world assemblies. The core methodology utilizes a graph-based representation of assemblies, wherein parts are represented as nodes and their connections as edges.In the domain of assembly modeling, design engineers utilize computer-aided design (CAD) software to develop complex assemblies. However, the increasing volume and size of non-standardized part catalogs from various manufacturers pose significant challenges in the selection of suitable parts from the vast number of options. This thesis addresses these challenges by proposing a novel methodology that leverages machine learning techniques to support inexperienced designers in the assembly design process by extracting expert knowledge from previous assemblies. Specifically, it investigates the use cases of global part recommendation, localized part recommendation and handling anomalies in assemblies. The first contribution is a generic, data-driven approach to extract patterns of proven part combinations across multiple real-world assemblies. The core methodology utilizes a graph-based representation of assemblies, wherein parts are represented as nodes and their connections as edges. Leveraging this representation, the methodology provides means and guidelines for modeling arbitrary part recommendation tasks by applying graph machine learning. As second contribution, we developed an embedding technique to learn the similarity of parts in terms of their usage across multiple assemblies. This technique adapts a method from the field of natural language processing to general graph structures. The resulting embeddings provide features for the parts in all learning tasks to enhance the models’ generalization capabilities. The third contribution comprises an automated approach to generate learning instances for self-supervised machine learning from assembly data which can be tailored to a specific learning task on assemblies. This method addresses the scarcity of labeled data in real-world assembly datasets without involving domain experts. Moreover, this thesis introduces an algorithm for generating synthetic anomalous assemblies by extracting regular part combinations from a dataset of assemblies. Finally, this thesis is the first to address part recommendation during assembly modeling by analyzing previous assemblies through machine learning. It proposes a general framework for a recommendation task, modeled as a classification problem to generate a fixed number of recommendations. The experimental results across all use cases demonstrate that machine learning-based methods can greatly enhance the efficiency of assembly modeling, improve knowledge transfer among engineers, and reduce time for perusing extensive part catalogs.…