Overcoming　the　challenges　of　low　prediction　accuracy　and　poor　generalization　ability　in　autonomous　vehicle　crash　analysis　based　on　high-dimensional　small　sample　data　is　a　significant　challenge.　By　integrating　interpretable　machine　learning,　data　augmentation,　and　two-layer　stacking　algorithms,　it　is　possible　to　accurately　identify　the　contributing　factors　of　crashes　and　enhance　the　performance　of　the　model.　The　research　findings　indicate　several　significant　insights:　a)　Although　autonomous　vehicles　are　effective　in　reducing　crashes　caused　by　illegal　driver　operation,　they　still　pose　a　potential　risk　under　extreme　working　conditions.　These　include　ambiguous　road　markings,　severe　weather　conditions,　and　untrained　traffic　scenarios.　b)　Two-layer　stacking　integrates　the　heterogeneity　of　the　base　algorithms,　resulting　in　improved　prediction　accuracy　for　different　injury　classes　compared　to　existing　models.　c)　The　feature　crosses　algorithm　combines　the　contributing　factors　that　have　strong　coupling　effects,　maximizing　the　prediction　accuracy　and　generalization　ability　of　the　two-layer　stacking　model.　The　research　results　are　expected　to　strongly　support　the　development　of　the　emerging　technology　of　autonomous　vehicles,　and　provide　an　important　reference　for　crash　analysis　and　risk　prevention　and　control　policy　design.　©　2024　The　Authors.