Structural　condition　perception　is　a　crucial　step　in　contemporary　structural　health　monitoring.　Sensor　malfunction　and　under-sensing　seriously　hamper　the　performance　of　the　structural　health　monitoring　system.　This　paper　proposes　a　novel　structural　dynamic　response　reconstruction　and　virtual　sensing　approach　for　structural　health　monitoring　using　a　sequence-to-sequence　modeling　framework　with　a　soft　attention　mechanism　from　the　perspective　of　sequence　data　generation.　This　framework　explicitly　utilizes　the　potential　spatiotemporal　correlation　in　sequence　data　and　promotes　the　efficient　flow　of　information　in　the　network,　thereby　significantly　improving　the　reconstruction　performance.　In　addition,　a　reconstruction　error　estimation　and　uncertainty　quantification　method　based　on　signal　complexity　characterized　by　entropy　is　also　developed.　The　effectiveness　and　robustness　of　the　proposed　method　are　verified　based　on　the　vibration　signals　of　a　footbridge　measured　onsite　under　low-amplitude　ambient　excitation.　Finally,　the　applicability　of　this　method　in　the　scenarios　of　modal　identification　and　sensor　validation　is　demonstrated.