Combustion　line　is　one　of　the　key　parameters　to　represent　the　combustion　stability　of　municipal　solid　waste　incineration　(MSWI)　processes.　A　complete　flame　image　template　library　is　the　foundation　for　achieving　combustion　line　quantification　from　a　detection　perspective　instead　of　relying　on　operational　experts　to　manually　observe　the　flame,　and　improving　the　intelligence　level　of　MSWI　process　control　through　real-time　feedback.　Aiming　at　the　problem　of　missing　combustion　line　flame　image　under　extreme　combustion　state,　a　generation　method　based　on　fusion　of　mechanism　knowledge　and　adversarial　technology　is　proposed.　First,　based　on　the　mechanism　mapping　relationship　between　the　three-dimensional　spatial　position　in　the　incinerator　and　the　image　pixels,　the　extreme　abnormal　flame　samples　are　analyzed.　The　pseudo-label　samples　are　obtained　by　shifting,　stitching　and　combining　the　pixel　points　of　normal　sample.　Then,　the　candidate　extreme　abnormal　flames　are　obtained　by　using　cycle　generative　adversarial　networks　(CycleGAN).　Finally,　a　new　two-level　evaluation　strategy　is　designed　by　combing　the　optimal　model　parameters　evaluation　based　on　Fréchet　Inception　Distance　(FID)　and　the　final　extreme　abnormal　flame　image　selection　according　to　the　pseudo　label.　The　experimental　results　based　on　a　MSWI　plant　of　Beijing　show　that　the　combustion　line　can　be　categorized　into　three　ranges:　51%　to　73.6%　as　normal,　47%　to　51%　and　73.6%　to　100%　as　abnormal,　and　0%　to　47%　as　extreme　abnormal..　When　the　second　level　evaluation　threshold　is　set　to　0.4,　the　proportion　of　the　satisfied　extreme　abnormal　flame　image　generated　by　the　proposed　method　is　85.7%,　which　is　better　than　the　traditional　evaluation　method.　©2024　Chin.Soc.for　Elec.Eng.