In　this　paper,　the　ideas　of　deep　auto-encoder　(DAE)　and　manifold　learning　are　adopted　to　solve　the　problem　of　structural　condition　diagnosis.　A　scalable　decentralized　end-to-end　unsupervised　structural　condition　diagnostic　framework　is　proposed.　Three　damage　diagnosis　mechanisms　are　clarified.　The　structural　damage　diagnosis　approaches　are　presented　from　the　latent　coding　domain　and　the　time　domain,　respectively.　In　the　latent　coding　domain,　an　undercomplete　DAE　is　established　to　extract　the　distribution　of　the　low-dimensional　manifold　of　the　signal.　On　the　contrary,　in　the　time　domain,　an　overcomplete　DAE　is　adopted　to　extract　the　reconstruction　error　of　the　signal.　Subsequently,　normalized　damage　quantitative　indicators　are　developed　in　the　two　domains.　The　damage　localization　method　is　also　clarified.　The　proposed　method　can　extract　features　directly　from　original　vibration　data　without　the　need　for　additional　signal　preprocessing　techniques.　More　importantly,　the　algorithm　relies　only　on　the　output　signals　and　does　not　require　a　numerical　or　scale　model.　This　framework　can　be　used　to　identify,　locate,　and　quantify　structural　damages.　The　validity　of　the　diagnostic　framework　is　verified　using　a　well-designed　laboratory　benchmark　structure.　A　large-scale　grandstand　structure　is　further　used　to　prove　the　ability　of　the　proposed　method　for　identifying　slight　structural　damage　caused　by　the　loosening　of　joint　bolts.　The　results　clearly　demonstrate　an　elegant　performance　of　the　proposed　damage　detection　algorithm　in　structural　condition　assessment　and　damage　localization.