Abstract Determination of the model order is a challenging problem in system identification, especially in output-only or operational modal identification where some modes are weakly excited. Although existing methods such as the stabilization diagram method (spectral information) are effective, they do not scale to high-dimensional data, which is usually needed for high-fidelity characterization of structural dynamics and has been made available in the emerging full-field measurement techniques using optical methods such as photogrammetry and laser vibrometers. In this proof-of-concept study we present a new non-parametric, data-driven approach for robust output-only identification of high-spatial-dimensional modal parameters of basic structures by efficiently processing and interactively exploiting the full-field measurement (i.e., very dense spatial measurement points). Specifically, we first over-estimate the system model once, producing a pool of candidate modes associated with their modal frequencies and full-field, high-spatial-dimensional mode shapes. This is accomplished by a data-driven method termed affinity propagation clustering (APC), where the active clusters, which are the active modes in our formulations, emerge from the “message-passing” procedure and does not require a pre-determination of the cluster number (mode or model order). Next, rather than using the spectral information to distinguish the physical and spurious modes in the stabilization diagram method, we exploit and visualize the spatial, full-field mode shape associated with each candidate mode to do so. We conduct extensive experiments on basic structural models with comparisons to a few existing methods. The results indicate that the new method is computationally efficient for identifying high-spatial-dimensional modal parameters, and robust to identify weak modes by exploiting the full-field measurement. We also discuss its applicability and limitations for structures with complex geometry (shapes).

中文翻译：

---

用于鲁棒仅输出模态识别的全场、高空间维测量的亲和传播聚类：概念验证研究

摘要 模型阶数的确定是系统辨识中的一个具有挑战性的问题，特别是在某些模态弱激发的仅输出或操作模态辨识中。尽管稳定图法（光谱信息）等现有方法是有效的，但它们不能扩展到高维数据，而高维数据通常是结构动力学的高保真表征所需要的，并且已经在新兴的全场测量中可用使用光学方法的技术，如摄影测量和激光测振仪。在这项概念验证研究中，我们提出了一种新的非参数、数据驱动的方法，通过有效地处理和交互利用全场测量（即, 非常密集的空间测量点）。具体来说，我们首先高估了系统模型一次，产生了与其模态频率和全场、高空间维模式形状相关联的候选模式池。这是通过称为亲和传播聚类 (APC) 的数据驱动方法实现的，其中活动簇，即我们公式中的活动模式，从“消息传递”过程中出现，不需要预先确定簇号（模式或模型顺序）。接下来，我们不是在稳定图方法中使用光谱信息来区分物理模式和杂散模式，而是利用和可视化与每个候选模式相关的空间全场模式形状来做到这一点。我们对基本结构模型进行了广泛的实验，并与一些现有方法进行了比较。结果表明，新方法在识别高空间维模态参数方面具有计算效率，并且通过利用全场测量来识别弱模式具有鲁棒性。我们还讨论了它对具有复杂几何形状（形状）的结构的适用性和局限性。