Aiming　at　the　problem　that　linear　data　reduction　algorithm　is　difficult　to　deal　with　data　with　nonlinear　structure,　this　paper　proposes　a　new　algorithm　for　facial　expression　feature　extraction　based　on　manifold　decomposition　algorithm.　The　algorithm　utilizes　the　characteristic　of　local　linearity　of　nonlinear　manifolds.　Through　classical　principal　component　analysis　The　local　linear　patches　of　nonlinear　manifold　structures　are　reduced　in　dimension.　The　local　PCA　representation　can　be　obtained　by　local　dimension　reduction,　and　then　the　local　coordinates　are　aligned　by　the　coordinate　arrangement　technique,　so　that　the　low　dimensional　coordinates　of　the　whole　manifold　can　be　obtained.　The　simulation　results　show　that　the　local　linear　dimensionality　reduction　algorithm　of　nonlinear　manifold　decomposition　is　superior　to　other　classical　manifold　learning　algorithms　in　the　recognition　accuracy　when　applied　to　facial　expression　feature　extraction.　©　2019　IEEE.