[Objectives]　The　accurate　identification　of　critical　road　segments　is　crucial　for　effective　traffic　management　across　entire　road　networks.　While　significant　progress　has　been　made　in　identifying　critical　road　segments,　existing　methods　often　fail　to　identify　relatively　critical　road　segments　in　local　areas　with　lower　traffic　flow,　particularly　in　large-scale　networks　such　as　city-level　road　systems.　[Methods]　To　address　this　limitation,　this　study　proposes　a　two-stage　feature　learning　method　based　on　the　dynamic　and　static　embeddings　of　road　segments　to　identify　critical　road　segments　in　large-scale　networks.　The　proposed　method　consists　of　several　key　steps.　First,　travel　routes　are　extracted　from　mobile　positioning　data　to　construct　a　comprehensive　traffic　corpus,　which　serves　as　the　foundation　for　further　analysis.　Next,　a　two-stage　feature　learning　process　is　conducted:　(1)　Static　embeddings　are　extracted　for　each　road　segment　to　capture　their　inherent,　unchanging　characteristics.　These　embeddings　are　clustered　to　identify　initial　cluster　centers,　which　serve　as　preliminary　indicators　of　critical　road　segments.　(2)　Dynamic　embeddings　are　then　extracted　for　each　road　segment　and　processed　using　attention　pooling,　which　emphasizes　the　most　relevant　aspects　of　the　traffic　data.　These　pooled　feature　vectors　are　subjected　to　differentiable　clustering,　a　technique　that　optimize　the　clustering　process　through　a　loss　function.　The　model　iteratively　adjusts　until　the　loss　value　converges,　signaling　optimal　clustering.　Upon　convergence,　the　static　and　dynamic　features　are　fused　to　generate　comprehensive　feature　representations　for　each　road　segment.　These　fused　features　are　clustered　again　to　identify　the　final　cluster　centers,　which　represent　the　critical　road　segments　within　the　network.　To　validate　the　proposed　method,　a　traffic　corpus　is　constructed　by　using　mobile　positioning　data　from　the　Third　Ring　Road　area　of　Fuzhou　City.　[Results]　An　identification　experiment　and　comparative　analysis　of　critical　road　segments　are　conducted　using　this　road　network　as　a　case　study.　The　results　show　that　the　proposed　method　effectively　identifies　critical　road　segments　in　large-scale　road　networks　and　relatively　critical　segments　in　local　areas.　[Conclusions]　Furthermore,　compared　to　existing　methods,　this　method　achieves　superior　performance　across　various　evaluation　metrics,　indicating　that　the　identified　set　of　critical　road　segments　is　more　reasonable　and　practical.　©　2025　Science　Press.　All　rights　reserved.