Droplet jetting behavior largely determines the final drop deposition quality in the inkjet printing process. Forming such behavior is governed by the fluid flow pattern. Therefore, a measurement of the flow pattern is of great importance for improving the printing quality of the inkjet printing process. Most of the current works use static images for the study of the drop evolution process. The problem of the static images is that the images cannot recognize the motion information (i.e., temporal transformation) of the droplet. Thus the information of the jetting process in the temporal domain will be lost. Instead of using the images, this paper takes the video data as the study subject to investigate the droplet evolution behavior in the inkjet printing process. Moreover, this paper introduces a deep learning method for the study of such video data. Compared to most of the current learning approaches conducted in a supervised/semi-supervised manner for manufacturing process data, we propose an unsupervised learning method for studying the flow pattern of the droplet, which does not require well-defined ground-truth labels. Regarding the spatial and temporal transformation of the droplet in video data, we apply a deep recurrent neural network (DRNN) to implement the proposed unsupervised learning. To verify the hypothesis that the proposed method can learn a latent representation for reproducing original data, the proposed DRNN is trained and tested on both simulation and experimental datasets. Experimental results demonstrate that the proposed method can learn latent representations of the droplet jetting process video data, which is very useful for the prediction of the droplet behavior. Furthermore, through latent space decoding, the learned representations can infer the droplet forming stimulus parameters such as material properties, which would be very helpful for further understanding of the process dynamics and achieving real-time in-situ droplet deposition quality monitoring and control.

墨滴喷射行为在很大程度上决定了喷墨打印过程中最终的墨滴沉积质量。形成这种行为取决于流体的流动方式。因此，流型的测量对于提高喷墨印刷工艺的印刷质量非常重要。当前大多数作品都使用静态图像来研究液滴的演变过程。静态图像的问题是图像不能识别液滴的运动信息（即，时间变换）。因此，在时域中的喷射过程的信息将丢失。本文以视频数据为研究对象，而不是使用图像来研究喷墨打印过程中的墨滴演变行为。此外，本文介绍了一种用于学习此类视频数据的深度学习方法。与当前大多数以监督/半监督方式进行的制造过程数据学习方法相比，我们提出了一种无监督学习方法来研究液滴的流型，该方法不需要明确定义的地面真相标签。关于视频数据中液滴的时空变换，我们应用深度递归神经网络（DRNN）来实现所提出的无监督学习。为了验证所提出的方法可以学习潜在表示以再现原始数据的假设，在模拟和实验数据集上对所提出的DRNN进行了训练和测试。实验结果表明，该方法可以学习液滴喷射过程视频数据的潜在表示，对于预测液滴行为非常有用。此外，通过潜在空间解码，学习到的表示可以推断出液滴形成的刺激参数，例如材料特性，这对于进一步了解过程动力学以及实现实时原位液滴沉积质量监测和控制非常有帮助。