Online droplet anomaly detection from streaming videos in inkjet printing

Abstract

Inkjet printing (IJP) has demonstrated its capabilities to produce high-quality and high-resolution parts, such as sensors, bio-chips, etc., with outstanding functionality. However, the quality of the printed parts can be endangered by the abnormal droplet jetting behaviors, which are substantially governed by the process dynamics and ambient conditions. Additionally, the droplet behaviors extensively define the final drop deposition quality in IJP. Timely capturing and identifying abrupt anomalies that the ejected droplets may suffer from are primordial for the quality assurance of the printed parts. Machine vision systems are able to record the droplet videos during IJP. Nevertheless, it is challenging to timely detect the anomalies from the collected droplet videos. The objective of this work is to build an analytical framework that allows for online droplet anomaly detection from droplet videos. There are several challenges that have not been addressed before: (1) the features of the droplet videos need to be extracted in an online fashion; and (2) there is no well-defined baseline to support the droplet anomalies detection, due to the complex process dynamics. Here, we propose a novel online framework to efficiently detect the anomalies from process streaming videos. In particular, we extend the multivariate Bayesian online change detection (BOCD) framework to high-dimensional data (i.e., tensor data of droplet videos) by leveraging online tensor factorization (OTF). OTF decomposes the streaming data into non-temporal and temporal low-dimensional factorization matrices. The non-temporal factorization matrices are deployed to extract the frame-specific temporal factorization matrix within a user-defined sliding window. Subsequently, the temporal factorization matrix for each frame is monitored with BOCD, which accurately detects anomalies in the streaming data. The proposed framework is demonstrated by detecting droplet anomalies from streaming data in IJP, showing excellent accuracy and efficiency.