Based　on　the　classical　matching　pursuit　algorithm,　the　composite　dictionary　multi-atoms　matching　algorithm　based　on　structural　characteristics　of　signal　is　proposed,　and　applied　to　bearing　fault　diagnosis　field.　For　the　structural　characteristics　of　the　roller　bearing　fault　signals,　a　composite　dictionary　combining　the　impulse　time-frequency　dictionary　and　the　Fourier　dictionary　is　constituted.　Sparse　decomposition　and　reconstruction　algorithm　of　composite　dictionary　multi-atoms　matching　pursuit　is　researched　and　used　to　extracting　fault　features,　and　the　threshold　de-noising　is　introduced　in　the　reconstruction　algorithm.　The　analysis　results　of　roller　bearing　fault　experiment　signals　and　engineering　signals　indicate　that　the　impact　failure　characteristic　extraction　effect　of　composite　dictionary　multi-atoms　matching　algorithm　based　on　structural　characteristics　is　superior　to　single-atom　matching　algorithm,　and　hard　threshold　is　superior　to　that　without　threshold.　©　2012　Journal　of　Mechanical　Engineering.