Continuous video stream pixel sensor: A CNN‐LSTM based deep learning approach for mode shape prediction
Continuous video stream pixel sensor: A CNN‐LSTM based deep learning approach for mode shape prediction
复制标题
连续视频流像素传感器：基于 CNN-LSTM 的模式形状预测深度学习方法
DOI：
10.1002/stc.2892
复制
发表时间：
2021-11
影响因子：
5.4
通讯作者：
Ruoyu Yang;S. Singh;Mostafa Tavakkoli;Nikta Amiri;M. Karami;Rahul Rai
中科院分区：
文献类型：
--
作者：
Ruoyu Yang;S. Singh;Mostafa Tavakkoli;Nikta Amiri;M. Karami;Rahul Rai
模态分析已成为全球公认的制定和优化工程结构行为功能的工具，有助于评估结构故障并制定维护计划。模态分析旨在确定激励下系统的频率、阻尼比和模态形状。然而，传统的振型测量方法（如接触传感器）由于传感器的重量和低空间分辨率而容易出现精度和准确度问题。在本文中，我们改进了各种现有的模态形状确定方法，并引入了用于模态预测的全场像素传感器的思想。所提出的基于计算机视觉的深度学习架构可以非常精确地预测振动结构的振型。此外，还策划了一个由振动记录视频和基于有限元分析 (FEA) 的标签组成的 ModeShape 数据集。具体来说，我们引入了一种基于卷积神经网络、长短期记忆（CNN-LSTM）计算机视觉的非接触式振动测量技术，用于自动模态预测。关键思想是使用 RGB 相机的每个像素作为传感器并处理捕获的时空数据以实现模态形状预测。我们的 CNN-LSTM 模型将振动结构的视频流作为输入并产生基本模态振型。所提出的技术是非侵入性的，可以以相对较高的空间密度提取信息。 CNN-LSTM 模型充分利用了实验结果。通过利用各种不同材料和波动尺寸的样本，对深度学习模型的稳健性进行了仔细检查。
Modal analysis has emerged as a globally accepted tool to formulate and optimize the behavioral functions of engineering structures, which assists in assessing structural failure and laying out a plan for their maintenance. Modal analysis aims at determining the frequencies, damping ratios, and mode shapes of the system under excitation. However, conventional mode shape measurement methods like contact sensors are prone to precision and accuracy issues owing to the sensor's weight and low spatial resolution. In this paper, we improve upon various existing methods for mode shape determination and introduce the idea of a full‐field pixel sensor for mode shape prediction. The proposed computer vision‐based deep learning architecture predicts the mode shape of a vibrating structure with significant precision. Besides, a ModeShape dataset consisting of the vibration recording video and finite element analysis (FEA) based label has been curated. Specifically, we introduce a convolutional neural network, long short‐term memory (CNN‐LSTM) computer vision‐based non‐contact vibration measurement technique for automated mode shape prediction. The key idea is to use each pixel of a RGB camera as a sensor and process the captured spatio‐temporal data to enable mode shape prediction. Our CNN‐LSTM model takes the video streams of a vibrating structure as input and yields the fundamental mode shapes. The proposed technique is non‐invasive and can extract information at relatively high spatial density. The CNN‐LSTM model is proficient by utilizing experimental outcomes. The robustness of the deep learning model has been scrutinized by utilizing specimens of an assortment of different materials and fluctuating dimensions.