Towards Intelligent Architecting of Aerospace System-of-Systems: Part II

System-of-Systems (SoS) are composed of large scale independent and complex heterogeneous systems which collaborate to create capabilities not achievable by a single system, for example air transportation system, satellite constellations, and space exploration architectures. To support architecting of aerospace SoS, in this work we present a methodology to accurately predict different aspects of performance for design/operation and SoS architecting, expanding previous work on intelligent architecting of aerospace SoS, by adding rigorous Uncertainty Quantification via Bayesian Neural Networks. A Bayesian Neural Network is a neural network with a-priori distribution on its weights. In addition to solving the overfit problem, which is common to traditional deep neural networks, Bayesian Neural Networks provide automated model pruning (or reduction of feature design space), that addresses a well-known dimensionality curse in the SoS domain. We enable SoS design/operation by using modeling and simulation, quantifying the uncertainty inherently present in SoS, and utilizing Artificial Intelligence and optimization techniques to design and operate the system so that its expected performance or behavior when the unexpected occurs (for example, a failure) still satisfies user requirements. Much of the research effort in the field of SoS has focused on the analysis of these complex entities, while there are still gaps in developing tools for automated synthesis and engineering of SoS that consider all the various aspects in this problem domain. In this expansion of the use of Artificial Intelligence towards automated design, these techniques are used not only to discover and employ features of interest in a complex design space, but also to assess how uncertainty can affect performance. This capability supports the automated design of robust architectures, that can effectively meet the user needs even in presence of uncertainty. The SoS design and evaluation methodology presented in this paper and demonstrated on a synthetic modular satellites problem starts from modeling and simulation, and design of experiments to explore the design space. The following step is deep learning, to develop a model which relates SoS architectural features with performance metrics. Uncertainty Quantification techniques are then applied to assess the performance metrics for different architectures. Once the most critical features that affect the SoS performance are identified, stochastic optimization of the SoS on a reduced design space can be performed to determine Pareto optimal features. The final step is determining if any additional design/operation measures need to be explored to further maximize the SoS performance.