Book Description

Information revolution is turning modern engineering systems into smart and connected systems. The smart and connected systems are defined by three characteristics: tangible physical components that comprise the system, connectivity among components that enables data acquisition and sharing, and smart data analytics and decision making capability. Examples of smart and connected systems include GM's OnStar® tele-service system and the InSite® tele-monitoring system from GE. The unprecedented data availability in smart and connected systems provides significant opportunities for data analytics. For example, since we have observations from potentially a very large number of similar units, we can compare their operations, share the information, and extract some common knowledge to enable accurate prediction and control at the individual level. In addition, for a complex system such as multistage manufacturing processes, we can collect synchronized data from multiple stations within the system so that we can identify the operational relationships among these stations. Such relationship can enable better process control. On the other hand, the tremendous data volume and types also reveal critical challenges. First, the high dimensional data with heterogeneity often poses difficulties in sharing common information within/across similar units/processes in the smart and connected systems. This problem becomes more severe when the system under the start-up period, where insufficient data and experience could result in the deficiency of data driven approaches. Second, the non-Gaussian data and non-linear relationship among various units impede the quantitative description of the inter-relationship of processes in the smart and connected systems. Although existing non-parametric methods, e.g., Kriging, can deal with these situations to some extent, limited description power (focus on mean value prediction) and lack of physical interpretation are the common drawbacks in these methods. Moreover, the real time monitoring and control for the smart and connected systems require efficient and scalability algorithms and strategies to meet the rapid and large scale response under advanced sensing and data acquisition environment. Lastly, the efficient control of the smart and connected systems also becomes challenging due to the complex relationship among units. Data-driven methods are required to meet the exigent demands for effectively formulating and solving the control problem. To address the issues listed above, four tasks are investigated in this dissertation under different applications in the smart and connected systems. [1] Transfer learning among heterogeneous multistage manufacturing processes. A series of data analytical methods for modeling and learning inter-relationships among product quality characteristics in multistage connected manufacturing processes are developed. The methods offer a rigorous way to reveal commonalities among heterogeneous data from different manufacturing processes to benefit the learning in complex connected manufacturing processes. [2] Statistical modeling and inference for Key Performance Indicators (KPI) in production systems. A surrogate model for inference and prediction at distribution level of different KPIs is developed. This model utilizes the pair-copula construction to capture the non-linear association in the non-Gaussian data. [3] Real time contamination detection in water distribution network. A contamination source identification framework is proposed for real time tracking and detection of contamination released in the urban water distribution network. The framework utilizes the Bayesian theory to sequentially update the posterior probability for determining the contamination source upon very limited sensor readings. [4] Control of KPIs in manufacturing production systems. The KPI control problem is formulated as a stochastic optimization problem, where the noise distribution in the cost function depends on the decision variables. The standard uniform distributions are employed to link the KPI relationship surrogate model and the objective function to efficiently solve the KPI control problem. The proposed methods can be applied to a broad range of data analytics problems, and the emerging challenges in modeling, monitoring and control of smart and connected systems can be effectively addressed.