No abstract

The authors present an approach for verifying computer simulations for spatiotemporal systems. Specifically, the goal is to confirm that the computer simulation for a linear controller of an Unmanned Underwater Vehicle (UUV) correctly captures key elements of a conceptual model for the UUV controller. The key elements considered are stability and robustness of the UUV’s control system model. Here, stability refers to asymptotic stability of the model’s state variables, assuming bounded input parameters. Robustness refers to the model’s ability to handle uncertainty in these parameters. The verification is performed in two steps. First, for a fixed UUV controller, a maximum range of uncertain parameters is found and for that entire range confirmed that the computer simulation remains stable as required by the conceptual model. Next, a combination of control cost function and range of uncertain parameters is optimized while maintaining UUV stability, again, for the entire range of uncertainty and as required by the conceptual model. In both steps, independent Monte-Carlo simulations are performed to confirm the verification.

No abstract

Data-driven prognostics typically requires sufficient offline training data sets for accurate remaining useful life (RUL) prediction of engineering products. This paper investigates performances of typical data-driven methodologies when the amount of training data sets is insufficient. The purpose is to better understand these methodologies especially when offline training datasets are insufficient. The neural network, similarity-based approach, and copula-based sampling approach were investigated when only three run-to-failure training units were available. The example of lithium-ion (Li-ion) battery capacity degradation was employed for the demonstration.

Model bias can be normally modeled as a regression model to predict potential model errors in the design space with sufficient training data sets. Typically, only continuous design variables are considered since the regression model is mainly designed for response approximation in a continuous space. In reality, many engineering problems have discrete design variables mixed with continuous design variables. Although the regression model of the model bias can still approximate the model errors in various design/operation conditions, accuracy of the bias model degrades quickly with the increase of the discrete design variables. This paper proposes an effective model bias modeling strategy to better approximate the potential model errors in the design/operation space. The essential idea is to firstly determine an optimal base model from all combination models derived from discrete design variables, then allocate majority of the bias training samples to this base model, and build relationships between the base model and other combination models. Two engineering examples are used to demonstrate that the proposed approach possesses better bias modeling accuracy compared to the traditional regression modeling approach. Furthermore, it is shown that bias modeling combined with the baseline simulation model can possess higher model accuracy compared to the direct meta-modeling approach using the same amount of training data sets.