CAREER: Transfer Learning Based Quality Improvement in Spatially - Temporally Complex Systems

With mass customization having clearly established itself in the marketplace, high product variety and short life cycles have now become key features of many manufacturing industries. Manufacturing systems having these features are called spatially-temporally complex systems in this proposal. Here, spatial complexity refers to the many diverse capabilities the system must be able to provide; temporal complexity refers to the constant evolving of the system over time. Quality control of such systems faces enormous challenges, as the systems usually have enormous manufacturing complexities, reflected by numerous processing steps/operations and intricate interaction between them. In addition, these systems have a very complicated and unique data structure. On the one hand, the spatial-temporal complexity leads to the system to consist of many configurations each having a different data distribution and variable relationship. On the other hand, the respective data of these configurations may share an underlying unified representation due to the fact that these configurations may be equipped with the same types of machines, operate under the same manufacturing environment, produce products belonging to the same product family or generations of the same product, or use the same types of sensors to collect data, all of which imply that these configurations are similar in some ways. Existing methods fail to convert such data into useful knowledge for quality improvement. Common pitfalls are to either treat the different configurations as a single one and pool their data together; or model them independently, which fails to take the advantage that the knowledge gained for one configuration may be useful for other similar configurations. Therefore, the objective of this project is to develop, implement, and teach a transfer learning based methodology for modeling and decision-making of spatially-temporally complex systems, for quality improvement. This methodology features a novel idea of transfer learning, aiming to leverage the knowledge gained during quality control of a configuration for quality control of other spatially-temporally similar configurations. As a result, quality control of all configurations, i.e., the entire system, can be achieved effectively and efficiently.