The　hybrid　algorithm　which　combines　improved　genetic　algorithm　(GA)　with　error　back-propagation　algorithm　(BP)　is　used　to　train　artificial　neural　network.　The　defects　of　conventional　BP　algorithm,　i.e.,　easy　to　fall　into　local　minimum,　slow　convergence　speed　of　the　weight　value　of　learning　network,　and　that　of　GA,　i.e.,　the　training　speed　is　too　slow　when　GA　is　used　to　train　the　neural　network　effectively　improved　by　itself,　are　effectively　improved　by　the　hybrid　algorithm.　The　application　of　the　hybrid　algorithm　to　power　transformer　fault　diagnosis　is　simulated,　the　results　show　that　the　hybrid　algorithm　possesses　faster　convergence　speed　and　higher　calculation　accuracy.