Monitoring and responding to complaints generated by consumers are critical for enterprises and governments. In this paper we propose a multimodal and bi-layer fusion framework to identify microblogs posted on SinaWeibo, the largest social media platform in China. We design an effective and simple method for feature fusion. Based bag-of words model, we represent text and images of a post of microblogs as a unified representation. Using multimodal features, we propose a bi-layer fusion framework. Experiments on two datasets of 5000 microblogs demonstrate our method performs better than text-based methods. The weighted F-measure reaches 0.904, and F-measure of complaint class is 37.07% higher than the baseline.