Rapid　identification　of　the　rock　mass　condition　at　the　tunnel　face　is　a　key　problem　for　TBM　operating　parameters　optimization　and　subsequent　tunnel　support　measures　selection.　The　vibration　induced　by　the　rock　breaking　contains　essential　information　for　evaluating　the　tunnel　face　condition.　However,　conventional　vibration-based　methods　face　difficulties　in　continuously　obtaining　vibration　records　for　long　tunnel　sections.　Additionally,　there＇s　a　lack　of　TBM　cutterhead　vibration　monitoring,　and　they　heavily　depend　on　expertise　and　prior　knowledge.　In　this　study,　an　end-to-end　deep　learning　(DL)　method　was　developed　for　rock　mass　class　identification　of　TBM　tunnel　working　faces　based　on　the　measurement　of　TBM　cutterhead　vibration　signals,　including　cutterhead　vibration　signal　measurement,　signal　preprocessing,　model　training　and　optimization,　and　application　verification.　The　model　combines　the　advantages　of　1DCNN,　BiLSTM,　and　self-attention　mechanisms,　where　the　structural　innovation　of　1DCNN　inspired　by　Inception　v2　for　multi-scale　feature　extraction.　Which　can　automatically　extract　the　spatial　and　temporal　domain　features　in　the　signals　to　promptly　identify　the　rock　mass　class　at　the　working　face　without　stopping　the　normal　tunneling　process.　The　accuracy　on　the　test　set　is　95.89%　compared　to　85.34%　for　a　traditional　ML　model,　and　it　has　better　performance　than　other　DL　model　architectures.　The　model　underwent　validation　during　subsequent　TBM　tunneling　within　the　same　project,　successfully　proving　its　practical　reliability.