Bayesian network (BN), as a preliminary framework for representing and inferring uncertain knowledge, is widely used in social network, knowledge graph, medical diagnosis, etc. The centric computing task of BN-based analysis, diagnosis, and decision-support in specific fields includes multiple probabilistic inferences. However, the high time complexity is doomed on the same BN by using the traditional inference methods, due to the several intermediate results of probability calculations that cannot be shared and reused among different inferences. Therefore, to improve the overall efficiency of multiple inferences on the same BN, this study proposes the method of BN embedding and corresponding probabilistic inferences. First, by incorporating the idea of graph embedding, the study proposes a BN embedding method based on the autoencoder and attention mechanism by transforming BN into the point mutual information matrix to preserve the directed a cyclic graph and conditional probability parameters simultaneously. Specifically, each coding layer of the autoencoder generates node embedding by using the correlation between a node and its neighbors (parent and child nodes) to preserve the probabilistic dependencies. Then, the method for probabilistic inferences to measure the joint probability by using the distance between embedding vectors is proposed. Experimental results show that the proposed method outperforms other state-of-the-art methods in efficiency, achieving accurate results of probabilistic inferences.