Abstract
Progress in the product development and production industry has taken a step toward automation because the competition among producers forces complexity and the fabrication of advanced products. Additive Manufacturing (AM) is considered a unique and competitive technology for producing complex geometries with material diversity. Unlike traditional manufacturing (TM) methods (e.g., casting, machining, forming), which often have limitations due to tool requirements, AM allows building parts layer-by-layer, which is key to the fabrication of complex geometries. AM technology allows designers to optimize shapes for performance and aesthetics rather than manufacturability, leading to complex internal structures, varying thickness, and organic shapes. The geometric freedom that comes with AM opens new possibilities for innovative products across various industries. Even though design methods and principles hold particular importance in engineering, graphic design, and architecture, a clear and formal definition of the characteristics that establish a suitable geometry for AM has not been formalized. The absence of this standardized definition poses a challenge for manufacturers as it creates biases while selecting a manufacturing process. Simultaneously, the introduction of Artificial Intelligence (AI) in industrial applications has paved the way for more accurate, adaptive, and efficient products through automation. Machine Learning (ML), a branch of AI, is a valuable tool for advancing industrial applications, including manufacturing process selection (MPS). AI and ML are believed to be crucial for the evolution of traditional manufacturing systems into Industry 4.0 and 5.0. Traditionally, MPS is performed by experts who evaluate multiple factors and parameters; however, research indicates that AI systems could potentially outperform expert judgment in this task. Supervised learning algorithms, in particular, have shown promise in coping with labelled data in supporting MPS, while unsupervised learning is advantageous for working with unlabeled datasets, a common scenario in manufacturing data. To overcome these challenges, this Ph.D. research aims to investigate the role of part geometry in MPS by following, comparing and possibly integrating two distinct approaches: unsupervised ML and human decision-making. Both approaches are set to evaluate the suitability of parts for AM or TM. The first approach is based on a hierarchical clustering algorithm that uses unlabeled images/CAD models as inputs to generate two clusters of AM and TM parts as a support tool for MPS. The algorithm can use images directly, while an additional process is required for CAD models to be used as inputs. CAD models are converted into images via CAD processing software (SolidWorks was used in this research). The external geometry of each part is highlighted using a feature extraction methodology named “histogram of oriented gradients”. The second approach, which included two surveys and one semi-structured interview, was used to gather information from experts and novices in the manufacturing and engineering field. The first survey was conducted with design and manufacturing experts who were asked to evaluate parts’ suitability to AM or TM, considering each product’s image and CAD model. The analysis of survey results showed that the agreement level among experts was poor, which was the primary motivation behind conducting the semi-structured interviews and the second survey. The final AM and TM clusters generated by the developed algorithm were compared to the first survey results. A 76% convergence was achieved between experts’ designations in terms of parts suitable for AM and TM, and algorithm clusters. The convergence level reached 90% for those parts where experts’ decisions were significantly aligned. Furthermore, this research values the input of both experts and novices in understanding the influence of part geometry and the rationale behind MPS. The interview was a key tool in gathering expert insights, while the second survey provided valuable perspectives from novices. The analysis of interview data gave rise to seven manufacturing properties crucial to MPS, with part geometry being of secondary importance unless specific features were present on the part. The findings of the thesis suggest that it is effective to use unsupervised learning to support decision-making in MPS since there exists a codependent relationship between part geometry and AM, with each affecting the other. The results of the thesis confirm that specific design characteristics bend the manufacturing choice towards AM or disregard it as the primary choice. The main reason behind this is that a clear boundary among manufacturing processes cannot be defined as various characteristics influence the choice. There is a “grey zone” where significant disagreements were found among experts and novices and between the two cohorts in which no salient design characteristics were present. The findings of this research support the use of unsupervised learning algorithms as a tool to enhance decision-making processes in the engineering field at large.