Blind source separation (BSS) is one of the fundamental techniques for resolving meaningful features in functional magnetic resonance imaging (fMRI). BSS methods based on unsupervised shallow models (e.g., restricted Boltzmann machine, RBM) have improved fMRI BSS compared to conventional matrix factorization models (e.g., independent component analysis (ICA)). In machine learning field, it is widely accepted that deeper models (e.g., deep belief network, DBN) are more powerful in latent feature learning and data representation. Thus, in this paper we propose a BSS model based on DBN with two hidden layers of RBM. In addition, we apply the model to fMRI time series for BSS instead of fMRI volumes as proposed in previous studies, such that the parameter searching space is significantly pruned and large-scale training samples of fMRI time series are available. Our experimental results on an fMRI dataset acquired with a movie stimulus showed that the proposed model is capable of identifying not only latent components related to distinct brain networks, but also the ones related to functional interactions across different networks.