Partially Observable Markov Decision Processes (POMDP) Framework for Decision-Making Under Uncertainty in EHD Printing Using Image Based Monitoring System

Abstract A Partially Observable Markov Decision Processes (POMDP) framework has been developed for Electrohydrodynamic (EHD) printing to facilitate decisionmaking under uncertainty and automate adjustments in additive manufacturing. The POMDP system utilizes its reinforcement learning structure to optimize decisions by interacting with the environment. It processes feedback from a high-speed camera imaging system that monitors the printed line width (in-pixels) during printing which represents the state space. Image processing and unsupervised learning algorithms including K-means, Gaussian mixture models (GMM), and Agglomerative clustering are implemented to analyze this state space and create the transition model. This state space is partially observed and linked to the recorded highspeed images, through the observational model. As the difference between the first and last groups of the region of interest (ROI) grayscale profile of 200 images received from the stream images represents the observational space of the POMDP. Then the observed space is discretized to develop a finite set of observations. This model facilitates the dynamic adjustment of printing parameters such as plotting speed and standoff distance, ensuring optimal printing quality within the desired specification. It applies different solution methods and compares their performance to select the best for the EHD settings. Results show that the point-based value iteration method is the best approximation solution that can be used in this POMDP EHD settings framework with moderate complexity compared to other solution approaches.