The　flame　state　recognition　during　the　combustion　stage　of　municipal　solid　waste　is　an　important　component　of　intelligent　optimization　control　in　the　process　of　municipal　solid　waste　incineration　(MSWI).　A　single　feature　quantity　cannot　fully　represent　all　combustion　states.　Using　too　many　feature　quantities　will　consume　a　lot　of　computer　resources,　and　make　timeliness　worse.　In　response　to　the　above　issues,　this　article　first　based　on　the　experience　of　domain　experts　and　research　achievements　in　related　fields,　divides　the　flame　combustion　state　into　four　categories　based　on　global　information:　normal　burning,　Channelling　burning,　partial　burning,　and　Smothering;　Then　the　flame　image　is　de-fogged　and　denoted　using　pre-processing　means　such　as　artificial　multi-　exposure　image　fusion　de-fogging　algorithm,　feature　normalization,　trap　filtering,　median　filtering,　etc.,　to　obtain　a　clear　flame　image;　Next,　multiple　features　such　as　brightness,　flame,　color,　and　principal　components　are　extracted　from　the　flame　image　to　represent　the　image　from　multiple　views,　and　the　multi-view　features　are　reduced　based　on　mutual　information;　Finally,　a　combustion　state　recognition　model　is　constructed　based　on　reduced　image　features　and　deep　forest　models　©　2024　IEEE.