The　classification　ambiguity　caused　by　the　traditional　auto-encoding　network　that　only　focuses　on　minimizing　reconstruction　error　in　process　of　feature　automatic　extraction.　A　novel　discriminative　auto-encoding　network　is　hence　proposed.　A　full　connection　layer　with　Softmax　classifier　is　linked　to　the　hidden　layer　of　the　auto-encoding　network,　and　the　cross-entropy　of　its　output　and　label　is　added　to　the　original　loss　function.　The　discriminative　auto-encoding　network　is　trained　with　the　objective　of　minimizing　the　composite　loss　function.　The　hidden　layer　of　upper　discriminative　auto-encoding　network　is　taken　as　the　input　of　the　next　network　and　is　stacked　in　their　order　to　form　a　discriminative　auto-encoding　network.　The　stacked　auto-encoding　network　before　and　after　improvement　are　used　to　extract　fault　features　of　rolling　bearings　automatically,　and　diagnosis　tests　are　carried　out　on　a　constant　speed　dataset　and　a　variable-speed　and　multi-load　dataset　in　laboratory,　respectively.　For　all　quantitative　calculation　results　of　the　extracted　features,　the　discriminative　network　can　reduce　the　intra-class　distance　by　8.26%　and　increase　the　inter-class　distance　by　23.02%　at　least.　Three　common　classifiers　are　used　to　classify　the　bearing　faults　respectively　with　the　features　extracted　by　two　networks,　and　42　manually　extracted　features　are　also　comparatively　analyzed.　The　classification　results　show　that　for　the　constant　speed　dataset,　the　recognition　accuracy　by　using　the　features　extracted　by　discriminative　network　and　manual　method　is　similar,　which　is　higher　than　that　by　using　the　features　extracted　by　the　traditional　auto-encoding　network.　However,　for　the　variable-speed　and　multi-load　dataset,　the　features　extracted　by　discriminative　network　are　superior　to　the　other　two　types　of　features.　The　quantitative　calculation　and　fault　diagnosis　indicate　that　the　discriminative　network　has　a　good　ability　of　automatic　feature　extraction　and　is　independent　on　working　conditions.　©　2019,　Editorial　Office　of　Journal　of　Xi＇an　Jiaotong　University.　All　right　reserved.