Motor　is　one　of　the　most　frequently　used　machines　in　industry.　Ensuring　the　reliability　of　motor　and　timely　identifying　the　fault　type　of　motor　can　guarantee　the　properly　working　and　prevent　the　great　loss.　In　this　paper,　a　turn-to-turn　short　circuit　of　motor　stator　and　unbalance　power　supply　fault　diagnosis　system　based　on　Deep　Sparse　Auto-Encoder　and　Softmax　Classifier　is　proposed.　The　selection　of　neural　network　parameters　and　their　influence　are　systematic　given.　A　new　method　where　dropout　rates　of　Sparse　Auto-Encoder　are　various　in　each　layer　is　adopted　to　increase　the　stability　and　convergence　speed　of　training　process.　Finally,　the　proposed　system　is　applied　to　an　experiment　on　a　motor　in　laboratory.　The　conclusion　shows　the　ability　to　identify　the　fault　type　of　motor　at　the　continuous　state　that　the　accuracy　reaches　100%,　when　only　the　data　from　motor　at　discrete　state　point　are　used　in　training,　which　makes　the　system　extensible　and　promising.　©　2018　IEEE.