Civil　structures　are　susceptible　to　performance　deterioration　during　their　service　lives.　Therefore,　it　is　essential　to　periodically　evaluate　structural　conditions　for　preventing　potential　catastrophic　failure.　Data-driven　methods　have　been　widely　adopted　for　this　purpose,　which　often　utilize　deep　learning　(DL)　algorithms　that　generally　require　extensive　training　data　considering　various　condition　scenarios.　However,　finite　element　(FE)　simulation　data　are　practically　affected　by　uncertainties　such　as　modeling　errors　and　operational　environmental　variations,　limiting　the　applicability　of　supervised　DL　algorithms.　Due　to　this,　an　unsupervised　learning　framework　has　been　proposed　for　structural　condition　evaluation　by　establishing　a　correlation　model　for　the　response　data　measured　at　different　locations　of　a　healthy　structure　using　bidirectional　long　short-term　memory　(BiLSTM)　networks.　During　the　training　process,　the　optimal　hyperparameters　of　a　BiLSTM　network　is　objectively,　instead　of　＇subjectively＇,　determined　through　Bayesian　optimization　(BO)　without　requiring　labeled　measurement　data,　improving　the　network　generalization　performance.　During　the　testing　process,　response　data　from　healthy　and　unknown　scenarios　are　input　into　the　BO-BiLSTM　network,　and　the　errors　between　the　reconstructed　and　actual　data　are　taken　as　the　latent　features.　Then,　the　feature　similarity　between　the　unknown　scenarios　and　the　healthy　structure　is　calculated　using　the　Wasserstein　distance　as　a　structural　condition　indicator.　The　feasibility　of　the　proposed　method　has　been　validated　using　the　IASC-ASCE　benchmark　frame　and　an　experimental　steel　frame,　demonstrating　that　the　proposed　condition　indicator　is　sensitive　to　structural　damage　and　robust　to　different　noise　levels.　As　structural　degradation　developed,　the　condition　indicators　for　the　two　frames　increased　from　0.090　and　0.583　to　1.182　and　0.825,　respectively.　The　structural　conditions　were　successfully　evaluated　without　the　labeled　measurement　data,　validating　the　engineering　applicability　of　the　proposed　method.