Classification　problem　is　important　for　big　data　processing,　and　deep　learning　method　named　deep　belief　network　(DBN)　is　successfully　applied　into　classification.　But　traditional　DBN　is　an　unsupervised　learning　method,　which　leads　to　a　gap　between　extracted　features　and　concrete　tasks.　In　this　paper,　a　semi-supervised　DBN　(SSDBN)　based　on　semi-supervised　restricted　Boltzmann　machine　(SSRBM)　is　proposed　to　shorten　the　gap　and　improve　the　accuracy　of　classification.　Firstly,　through　introducing　relevance　constraint,　supervised　information　is　equivalently　integrated　into　the　learning　process　of　restricted　Boltzmann　machine.　Secondly,　SSDBN-based　model　is　constructed　to　improve　the　accuracy　of　classification　problem.　Finally,　the　proposed　SSDBN　is　validated　with　hand-written　digits　classification　standard　dataset　MNIST,　and　experimental　results　show　that　SSDBN　outperforms　traditional　DBN　and　other　models　with　respect　to　classification.　(C)　2017,　IFAC　(International　Federation　of　Automatic　Control)　Hosting　by　Elsevier　Ltd.　All　rights　reserved.