SUMMARY

The ability to automatically identify manufacturing processes and their sequences capable of transforming a product design into its physical counterpart is a key requirement of cyber manufacturing services, and more generally in automated manufacturing process planning. This function is usually achieved using Computer-Aided Process Planning (CAPP) systems with hard-coded rule-based knowledge of manufacturing capabilities, which are usually derived from user experience. However, process capability knowledge embedded in most of the reported CAPP systems is usually incomplete and limited to one or very few manufacturing processes. In addition, the process capability knowledge may vary from system to system and is often difficult to scale as new knowledge becomes available. Modern data-driven or machine learning methods offer a potential solution to these challenges by providing computational methods to infer the core manufacturing process capability knowledge from digital design and manufacturing data such as 3D part geometry, material properties, and part quality. Motivated by this potential, this dissertation presents (i) data-driven methods for learning the capability of different discrete manufacturing processes and their sequences from existing design and manufacturing data, and (ii) illustrates the utilization of the said capability knowledge in process selection and sequencing. Specifically, both traditional machine learning methods (e.g., Decision Tree) and deep learning methods are investigated for learning the manufacturing process capabilities. Training and testing datasets are synthetically generated to evaluate the performance of these methods. Based on comparisons of the performance of the methods on the same datasets, it is concluded that the deep learning method, although computationally more expensive, outperforms the traditional machine learning methods. In addition, a data-driven framework is proposed to utilize a sequence mining algorithm to learn the manufacturing process sequence capabilities from design and manufacturing data. Finally, a data-driven framework that combines the aforementioned functions is proposed to suggest suitable manufacturing sequences for more realistic 3D part designs composed of multiple manufacturing features. An implementation example is presented to show the predictive strength of process selection and sequencing of the proposed framework. It is expected that the proposed research will lead to computational methods necessary to acquire process capability knowledge that can be easily updated automatically to adapt to the rapid technology advancements in manufacturing. Such adaptability is necessary to enable cyber manufacturing services.