High-dimension nonlinear functional data are commonly encountered in many manufacturing systems. It is an important and challenging topic to conduct feature extraction, monitoring, and data analytics from these datasets. Conventional statistical or machine learning methods, such as mixed-effects model and tensor decomposition, have several limitations including inability to separate detailed components, computational inefficiency, and high-dimensionality in data analysis. In this talk, we propose two novel methods: (1) a penalized mixed-effects decomposition (PMD) method, which can decompose the profiles into four components and enable in-line multi-channel profile detection; (2) a tensor mixed effects (TME) model, which can handle massive high dimensional datasets with complex correlation structures. An accelerated proximal gradient (APG) based optimization algorithm and an iterative double Flip-Flop algorithm are proposed to realize efficient parameter estimations. The properties of these two methods are explored. Using surrogated data analysis and real case studies, we evaluated the performance of the proposed PMD and TME method, and demonstrated their effectiveness and efficiency with applications to nano-manufacturing process. These methodologies can also be applied to data analytics for other smart manufacturing systems.
- Understand the multi-channel profile monitoring in manufacturing systems
- Understand a new technique for engineering-driven data analytics of smart manufacturing systems
- Be able to determine the scope of dimension reduction of big data in manufacturing systems
Why Is It Important?
Real-time data collection at various levels of manufacturing systems becomes easier and more cost-effective, by benefitting from new sensing and information technologies. Massive data with complex spatial and temporal structures (“big data”) provide great opportunities for improving system operations. Meanwhile, it raises significant challenges on data analytics, system pattern recognition, feature learning, and in-line decision making, because the data is usually high-dimensional (HD), large-size, noisy and heterogeneous. Thus, there is a pressing need to develop systematic methodologies for data analytics of smart manufacturing.My research focuses on engineering-driven data mining and analytics for smart manufacturing. The objective is to develop new methodologies for system-level fusion of data-driven models and physical models to effectively extract features from high-dimensional data, in order to realize real-time monitoring of system operations, accurate detection of system faults, quick diagnosis of root causes, predictive control and automation.