Due to rapid advancements in sensing and computation technology, multiple types of sensors have been embedded in various applications, on-line automatically collecting massive production information. Although this data-rich environment provides great opportunity for more effective process control, it also raises new research challenges on data analysis and decision making due to the complex data structures, such as heterogeneous data dependency, and large-volume and high-dimensional characteristics. This thesis contributes to the area of System Informatics and Control (SIAC) to develop systematic data fusion methodologies for effective quality control and performance improvement in complex systems. These advanced methodologies enable (1) a better handling of the rich data environment communicated by complex engineering systems, (2) a closer monitoring of the system status, and (3) a more accurate forecasting of future trends and behaviors. The research bridges the gaps in methodologies among advanced statistics, engineering domain knowledge and operation research. It also forms close linkage to various application areas such as manufacturing, health care, energy and service systems. This thesis started from investigating the optimal sensor system design and conducting multiple sensor data fusion analysis for process monitoring and diagnosis in different applications. In Chapter 2, we first studied the couplings or interactions between the optimal design of a sensor system in a Bayesian Network and quality management of a manufacturing system, which can improve cost-effectiveness and production yield by considering sensor cost, process change detection speed, and fault diagnosis accuracy in an integrated manner. An algorithm named “Best Allocation Subsets by Intelligent Search” (BASIS) with optimality proof is developed to obtain the optimal sensor allocation design at minimum cost under different user specified detection requirements. Chapter 3 extended this line of research by proposing a novel adaptive sensor allocation framework, which can greatly improve the monitoring and diagnosis capabilities of the previous method. A max-min criterion is developed to manage sensor reallocation and process change detection in an integrated manner. The methodology was tested and validated based on a hot forming process and a cap alignment process. Next in Chapter 4, we proposed a Scalable-Robust-Efficient Adaptive (SERA) sensor allocation strategy for online high-dimensional process monitoring in a general network. A monitoring scheme of using the sum of top-r local detection statistics is developed, which is scalable, effective and robust in detecting a wide range of possible shifts in all directions. This research provides a generic guideline for practitioners on determining (1) the appropriate sensor layout; (2) the “ON” and “OFF” states of different sensors; and (3) which part of the acquired data should be transmitted to and analyzed at the fusion center, when only limited resources are available. To improve the accuracy of remaining lifetime prediction, Chapter 5 proposed a data-level fusion methodology for degradation modeling and prognostics. When multiple sensors are available to measure the degradation mechanism of a same system, it becomes a high dimensional and challenging problem to determine which sensors to use and how to combine them together for better data analysis. To address this issue, we first defined two essential properties if present in a degradation signal, can enhance the effectiveness for prognostics. Then, we proposed a generic data-level fusion algorithm to construct a composite health index to achieve those two identified properties. The methodology was tested using the degradation signals of aircraft gas turbine engine, which demonstrated a much better prognostic result compared to relying solely on the data from an individual sensor. In summary, this thesis is the research drawing attention to the area of data fusion for effective employment of the underlying data gathering capabilities for system modeling, performance assessment and improvement. The fundamental data fusion methodologies are developed and further applied to various applications, which can facilitate resources planning, real-time monitoring, diagnosis and prognostics.

A fusion system sometimes requires the capability to represent the temporal changes of uncertain sensory information in dynamic and uncertain situation. A Bayesian Network can construct a coherent fusion structure with the hypothesis node which cannot be observed directly and sensors through a number of intermediate nodes that are interrelated by cause and effect. In some BN applications for observing a hypothesis node with the number of participated sensors, rank and select the appropriate options (different combination of sensors allocation) in the decision-making is a challenging problem. By user interaction, we can acquire more and useful information through multi-criteria decision aid (MCDA) as semi-automatically decision support. So in this study, Multi Attribute Decision Making (MADM) techniques as TOPSIS, SAW, and Mixed (Rank Average) for decision-making as well as AHP and Entropy for obtaining the weights of indexes have been used. Since MADM techniques have most probably different results according to different approaches and assumptions in the same problem, statistical analysis done on them. According to results, the correlation between applied techniques for ranking BN options is strong and positive because of the close proximity of weights suggested by AHP and Entropy. Mixed method as compared to TOPSIS and SAW is ideal techniques; moreover, AHP is more acceptable than Entropy for weighting of indexes.