In　this　letter,　we　propose　the　B-GHF　framework,　an　end-to-end　collision　state　inference　method　based　on　a　Bayesian　framework　that　does　not　rely　on　external　force/torque　(F/T)　sensors　in　the　human-robot　collaboration　(HRC)　environment.　This　method　integrates　GMM　for　probabilistic　object　position　and　error　analysis,　HMP　for　temporal　collision　state　evolution,　and　BNN　for　observational　uncertainties.　Dynamic　collision　state　assessment　and　decision　uses　multi-joint　state-weighted　integration　and　recursive　Bayesian　updates.　The　experimental　results　show　that　B-GHF　achieves　a　detection　success　rate　of　98.36%　and　an　average　detection　time　of　8.34　ms,　significantly　outperforming　both　a　state-of-the-art　(SOTA)　learning-based　method　(MCD-CNN)　and　a　classic　model-based　approach　(MO-ID)　in　terms　of　accuracy,　speed,　and　robustness.