The objective of Smart Manufacturing is to improve productivity and competitiveness in industry, based on in-process data. It requires reliable, explainable and understandable models such as Bayesian networks for performing tasks like condition prediction. In this context, a Bayesian network can be classically learned in a supervised, unsupervised way or a semi-supervised way. Here, we are interested in how to perform the learning when the ground truth isn't included in the learning data but is observable indirectly in another related dataset. This paper introduces a fully unsupervised variation of co-training that allows to include this second dataset, with two learning strategies (split and recursive). In our experiments, we propose one simple probabilistic graphical model used for predicting the state of a machine tool from results given by several sensors, and illustrate our unsupervised cotraining strategies first with benchmarks available from the UCI repository, for which 4 out of 5 datasets have best results with the recursive strategy. Finally, the recursive strategy was validated by McNemar's test as being the best strategy on a real industrial dataset.