The　traditional　method　of　analyzing　the　brain　network　of　depression　based　on　sub-frequency　bands　only　considers　the　classification　effect　of　single-frequency　sub-band　signals,　and　does　not　explore　the　relationship　between　different　time　bands.　Analyzing　the　correlations　between　each　brain　functional　segment　in　isolation　comes　from　global　information　missing　from　building　brain　networks.　Therefore,　this　paper　builds　a　brain　function　network　based　on　five　common　EEG　bands,　introduces　a　graph　attention　network　to　extract　the　global　information　of　the　depression　brain　network,　recalculates　the　correlation　between　brain　function　nodes,　and　extracts　more　refined　brain　functions　Connect　the　information,　and　build　an　autonomous　encoder　to　reduce　the　dimensionality　and　feature　enhancement　of　the　sparse　brain　functional　connection　matrix,　reduce　a　small　number　of　residual　information,　and　fuse　the　brain　networks　of　the　five　video　bands　to　obtain　the　brain　functional　connection　features　and　brain　features　that　include　all　video　band　features.　Classification　of　functional　node　features.　It　is　verified　on　the　public　data　set　MODMA,　and　the　highest　recognition　accuracy　rate　is　98.42%.