Recent years have seen the widespread utilization of vision-based noncontact methods for measuring rotor vibrations, but the measurement accuracy of such approaches is still substantially constrained by both the acquisition environment and the equipment, for which improving the quality and clarity of the captured sequence frames would be an effective solution strategy. In this article, a progressive video super-resolution (VSR) reconstruction network is thus constructed to enhance the image feature information during the preliminary phase of vibration displacement measurement, elevating the measurement accuracy while increasing the capture accuracy of the object detection algorithm. To address the challenge of the impractical application of VSR reconstruction methods in diverse industrial conditions, our approach employs pixel displacements between adjacent frames as a reference for motion estimation, ensuring effective feature alignment through a prealignment module. Additionally, a deep feature extraction module is implemented to capture long-range dependencies in multiscale feature representations, crucial for preserving structural image information. To further enhance reconstruction optimization, a feature fusion module (FFM) is introduced, integrating information from diverse rotor images. The experimental results demonstrate that the proposed network surpasses current advanced multiple comparison networks in reconstructing rotor datasets across diverse conditions and rotational speeds and achieves this with a modest parameter count and short run-time, striking a trade-off between computational cost and performance. Specifically, the network proposed in this article achieves peak signal-to-noise ratio (PSNR) values of 41.07, 26.11, 25.05, and 44.96 respectively, with less than half the parameter count of BasicVSR++ across four distinct rotor datasets. In comparison to other VSR networks, the reconstructed image frames in our network exhibit a smooth vibration displacement curve and minimal deviation.

中文翻译：

---

通过渐进式视频超分辨率网络提高转子振动位移的测量精度

近年来，基于视觉的非接触式方法被广泛用于测量转子振动，但这种方法的测量精度仍然受到采集环境和设备的严重限制，因此提高了捕获序列帧的质量和清晰度将是一个有效的解决策略。本文构建了渐进视频超分辨率（VSR）重建网络，以增强振动位移测量初始阶段的图像特征信息，提高测量精度，同时提高目标检测算法的捕获精度。为了解决 VSR 重建方法在不同工业条件下应用不切实际的挑战，我们的方法采用相邻帧之间的像素位移作为运动估计的参考，通过预对准模块确保有效的特征对准。此外，还实现了深度特征提取模块来捕获多尺度特征表示中的远程依赖性，这对于保留结构图像信息至关重要。为了进一步增强重建优化，引入了特征融合模块（FFM），集成了来自不同转子图像的信息。实验结果表明，所提出的网络在不同条件和转速下重建转子数据集方面超越了当前先进的多重比较网络，并以适度的参数数量和较短的运行时间实现了这一目标，在计算成本和性能之间取得了权衡。具体来说，本文提出的网络在四个不同转子数据集上的峰值信噪比 (PSNR) 值分别为 41.07、26.11、25.05 和 44.96，参数数量不到 BasicVSR++ 的一半。与其他 VSR 网络相比，我们网络中的重建图像帧表现出平滑的振动位移曲线和最小的偏差。