Embedding measured data within a computational framework for vibro-acoustic design

产品设计以达到可接受的噪音和振动水平是一系列行业的主要关注点。在许多情况下，成本和性能之间存在很大的权衡，这意味着实现高效的设计对商业成功至关重要。原则上，设计优化可以通过测试和改进物理原型来实现，但原型的生产既耗时又昂贵。由于这个原因，迫切需要虚拟设计方法，其中计算模型用于在构建物理原型之前生成接近最终的设计。用于噪声和振动分析的计算模型必须能够在较宽的频率范围内预测系统的性能，可能范围从几赫兹的低频振动问题到几千赫兹的高频噪声问题，这带来了严重的困难。高频运动需要非常详细的计算机模型，这导致长时间的运行，这不是迭代设计的理想选择。此外，系统的高频性能可能对微小的制造缺陷非常敏感，因此预测的性能可能与实际系统的性能不匹配。这些困难可以通过采用噪声和振动建模的最新进展来很大程度上克服，其中一种称为统计能量分析（SEA）的技术与更传统的分析方法（如有限元法（FEM）或边界元法（BEM））相结合；这种方法被称为混合方法。混合方法大大减少了模型的运行时间，同时还提供了由制造缺陷引起的性能差异的估计。然而，这种方法并不能完全解决预测问题，因为进一步的主要困难仍然存在：系统中的一些组件可能非常复杂，以至于不可能产生组件的详细计算模型，因此某种程度的物理测试是不可避免的。经常使用实验测量来验证计算模型，或更新计算模型中的参数，但这里的要求完全不同：必须使用测量数据通过将缺失的复杂组件的表示耦合到模型的其他部分来完成计算模型。这个问题构成了当前研究计划的核心。当前工作的目的是将“实验”组件添加到混合方法中，其中一种方法是将组件建模为灰盒或黑盒：灰盒模型由具有实验确定参数的数学方程组成，而黑盒模型纯粹基于测量的输入输出属性。这些模型必须能够耦合到FEM、BEM或SEA组件模型，项目将解决这个问题。一项主要挑战是确定在复杂振动声学组件环境下生成此类模型所需的适当实验测试和机器学习算法。第二个主要挑战是量化这些模型中的不确定性，并将这种不确定性包括在组合系统模型中。模型必须预测对设计者有用的输出，这些输出包括噪声和振动水平，以及预测的不确定性界限。在某些情况下，“声音质量”而不是整体噪音水平是值得关注的，该项目将开发技术，使组合模型的输出“听觉化”。将与工业伙伴一起进行若干案例研究，以探讨拟议方法的应用。目前的研究计划将产生一种有效和可靠的振动声学“科学设计”预测工具，以满足广泛的工业部门的需求。

The design of products to achieve acceptable levels of noise and vibration is a major concern across a range of industries. In many cases there is a large trade off between cost and performance, and this means that achieving an efficient design is crucial to commercial success. In principle design optimisation can be achieved through testing and improving physical prototypes, but the production of a prototype is time consuming and costly. For this reason there is a pressing need for virtual design methodologies, in which computational models are used to produce a near-final design before a physical prototype is built. Computational models used for noise and vibration analysis must be able to predict the performance of the system over a wide frequency range, potentially ranging from low frequency vibration problems at several hertz to high frequency noise problems at several kilohertz, and this presents severe difficulties. High frequency motions require a very detailed computer model, and this leads to long run times that are not ideal for iterative design. Furthermore, the high frequency performance of a system can be very sensitive to small manufacturing imperfections, and hence the predicted performance may not match the performance of the actual system. These difficulties can be largely overcome by employing recent advances in noise and vibration modelling in which a technique known as Statistical Energy Analysis (SEA) is combined with more conventional analysis methods such as the finite element method (FEM) or the boundary element method (BEM); this approach is known as the Hybrid Method. The Hybrid Method leads to a very large reduction in the run time of the model, while also providing an estimate of the variance in the performance caused by manufacturing imperfections. However, this approach does not fully solve the prediction problem, as a further major difficulty remains: some components in a system can be so complex that it is not possible to produce a detailed computational model of the component, and hence some degree of physical testing is unavoidable. Frequently experimental measurements are used to validate a computational model, or to update the parameters in a computational model, but the requirement here is quite different: the measured data must be used to complete the computational model by coupling a representation of the missing complex component to the other parts of the model. This issue forms the core of the current research proposal. The aim of the present work is to add "experimental" components to the Hybrid Method, and one way to do this is to model a component as a grey or black box: a grey box model consists of mathematical equations with experimentally determined parameters, while a black box model is based purely on measured input-output properties. These models must be capable of being coupled to either FEM, BEM, or SEA component models, and the project will address this issue. A major challenge is to determine the appropriate experimental tests and machine learning algorithms that are required to produce such models in the context of complex vibro-acoustic components. A second major challenge is to quantify the uncertainty in such models, and to include this uncertainty in the combined system model. The model must predict outputs that are useful to the designer, and such outputs include noise and vibration levels, together with uncertainty bounds on the predictions. In some cases "sound quality" rather than the overall noise level is of concern, and the project will develop techniques for the "auralisation" of the output of the combined model. A number of case studies will be developed with industrial partners to explore the application of the proposed approach.The present research programme will produce an efficient and reliable vibro-acoustic "design by science" prediction tool that meets the needs of a wide range of industrial sectors.