In　this　paper,　a　self-organizing　modular　neural　network　(SO-MNN)　is　proposed　for　nonlinear　system　modeling,　in　which　the　modular　structure　and　sub-networks　are　optimized　to　improve　the　modeling　accuracy　and　efficiency.　First,　the　modular　structure　is　constructed　by　seeking　the　maximum　modularity　degree　of　the　whole　neural　network　and　the　optimal　hub　center　in　each　sub-network.　Simultaneously,　the　task　is　divided　into　several　sub-tasks.　Then,　according　to　the　assigned　sub-tasks,　subnetworks　are　constructed　based　on　an　incremental　radial　basis　function　(RBF)　neural　network,　whose　performance　can　be　guaranteed　by　a　structure　growing　mechanism　and　an　adaptive　second-order　learning　algorithm.　Finally,　during　the　testing　or　application　processes,　a　winner-take-all　strategy　is　used　to　integrated　all　the　sub-networks.　The　effectiveness　of　the　proposed　methodology　is　verified　by　two　benchmark　problems　and　a　real-world　application.