Automated model complexity monitoring and adjustment for simulation-based design of dynamic systems

当使用不适当复杂性的模型时，基于仿真的设计是不必要的昂贵或不准确的。目前，还没有方法来检查和调整整个仿真机动中的模型复杂性。在仿真过程中，所需的复杂性在瞬变过程中与在稳态条件下不同。因此，所提出的研究的目标是：1)开发在线模型监测方法，以预测模型何时应该更新；2)预测哪些模型元素必须被移除或恢复；3)切换到更新后的模型而不会出现数值不稳定。因此，这种“主动建模”对车辆、机器人和机电系统设计行业非常重要。如果使用单一模型进行基于仿真的设计实践，如在不同地形上优化车辆悬架参数，则会产生重大的折衷。*如果模型过于复杂(例如，平坦地形上车辆的高度非线性多体模型)，则计算时间将不必要地高，并且优化可能成本高昂。*如果模型过于简单，无法捕捉复杂的瞬变(在崎岖地形上的点质量车辆模型)，则预测能力将不足。此外，拟议的主动建模研究将建立在申请者的博士和博士后工作的基础上，在这些工作中，系统内的功率流被用来分离动态系统模型，并创建更小、计算效率更高的子模型用于设计。此外，模型性能、模型输入的类型和位置、模型响应对输入的敏感度以及响应本身的变化将用于启动模型复杂性调整。通过使用键合图等图形模型形式化，将促进工作，并最大限度地提高对过程的物理洞察。这项工作在车辆设计和机器人技术中有直接的应用，但新的方法将适用于任何集总参数动态系统模型。

Simulation-based design is needlessly expensive or inaccurate when models of inappropriate complexity are used.  At present, no method exists to check and adjust model complexity throughout a simulation maneuver.  Over the course of a simulation, the required complexity is different during transient as opposed to steady state conditions.  The objectives of the proposed research are thus to 1) develop methods of on-line model monitoring to predict when the model should be updated, 2) predict which model elements must be removed or reinstated, and 3) switch to the updated model without numerical instability.    Such "Active Modelling" is important to the vehicle, robotic, and mechatronic system design industries.  A significant compromise results from using a single model for simulation-based design exercises such as optimization of vehicle suspension parameters over varying terrain.  If the model is too complex (e.g. highly nonlinear multibody model of a vehicle on flat terrain), then computation time will be needlessly high and optimization may be cost-prohibitive.  If the model is too simple to capture complex transients (point mass vehicle model on rough terrain), then predictive ability will be inadequate.    The proposed Active Modelling research will build on the applicant's doctoral and post-doctoral work wherein intra-system power flow was used to decouple dynamic system models and create smaller, more computationally efficient submodels for design.    Metrics based on power flow among system elements will be monitored and used to predict required model changes.  In addition, the model performance, type and location of model inputs, the sensitivity of model responses to the inputs, and variation in the responses themselves will be used to initiate model complexity adjustment.  The work will be facilitated, and the physical insight into the process maximized, by use of graphical model formalisms such as bond graphs.  The work has immediate applications in vehicle design and robotics, but new methodologies will be applicable to any lumped-parameter dynamic system model.