Online Health Management is required in autonomous systems and can be implemented with Bayesian Networks. However, probability tables are usually partially available and a so online learning technique is required. In this paper we present a new implementation of EM for Bayesian Networks based on an AC and partial computation. Compared to a classical implementation based on Junction Tree, our results show a speedup about 50 for a limited impact on accuracy. It demonstrates the possibility to run online diagnosis with learning capabilities that are necessary for real-life applications.