Owing　to　the　increasing　number　of　cybersecurity　threats　targeting　industrial　control　systems　(ICSs),　intrusion　response　systems　(IRSs)　have　become　essential.　However,　the　current　IRSs　exhibit　several　limitations,　such　as　neglecting　physical　domain　security　policies　and　relying　significantly　on　expert　input.　While　deep　reinforcement　learning　(DRL)　methods　yield　superior　outcomes,　they　suffer　from　low　interpretability　and　unreliability.　This　study　introduces　an　interpretable　cross-layer　intrusion　response　system　(ICL-IRS),　which　is　a　decision-tree-based　IRS.　It　offers　a　robust　understanding　of　cyberattacks　and　industrial　control　logic　specific　to　ICSs.　ICL-IRS　employs　a　DRL　model,　tailored　to　the　characteristics　of　physical　process　control,　to　refine　policies.　It　then　scrutinizes　the　optimized　intrusion　response　policy　and　generates　decision　trees.　Our　experimental　results　reveal　a　21%　enhancement　in　the　success　rate　of　the　proposed　ICL-IRS　over　competing　methods.　The　effectiveness　of　ICL-IRS　was　further　validated　through　a　case　study　on　a　simulated　process-control　system.