With the rapid development of online We Media, it is not easy to monitor and guide massive public opinion information. The massive forwarding, emotional polarization, and mixed rumors of blog posts are highly likely to cause online public opinion events. A Weibo public opinion monitoring system based on bidirectional encoder representation from transformers (BERT) is proposed to support the full lifecycle guidance for public opinion. After Weibo text data is crawled from blog posts of important informants according to the lifecycle stages, and preprocessed for data cleaning and text vectorization, an emotional tendency classification and rumor detection model with the way of ensemble learning is proposed with the latent Dirichlet allocation (LDA) model and three BERT-based combining model. For extracting emotional tendency information efficiently, the LDA model is applied to extract topics on emotional polarity in blog posts. In order to address the issue of long text data caused by the lifting of restrictions on short posts on Weibo, three BERT-based model are ensembled together to achieve better performance. The experiment test demonstrates that the evolution of online public opinion can be described efficiently by this system.