Rotating　machinery　is　widely　applied　in　industrial　fields.　However,　it　generally　operates　under　tough　working　conditions,　which　leads　to　the　weak　fault　features　and　renders　fault　diagnosis　more　difficult.　In　this　case,　an　emerging　method　called　sparse　representation　classification　(SRC)　is　proposed　to　enhance　the　fault　features　and　identify　the　fault　status.　However,　the　typical　SRC　theory　fails　to　consider　the　locality　of　the　test　sample　and　training　sample,　and　the　training　set　generally　contains　much　redundant　information,　which　may　reduce　the　fault　recognition　accuracy.　Moreover,　the　time-shift　deviation　of　vibration　signal　cannot　be　avoided　effectively　using　a　typical　SRC　model.　To　overcome　the　above-mentioned　problems,　a　novel　SRC　model,　i.e.,　weighted　SRC　based　on　dictionary　learning　(DL-WSRC),　is　proposed.　For　the　training　set,　different　fault　signals　are　learned　based　on　the　improved　K-singular　value　decomposition　(K-SVD)　algorithm,　which　can　not　only　adaptively　update　the　whole　training　set　but　also　reduce　redundant　information　so　as　to　enhance　the　sample　fault　features.　For　the　test　sample,　DL-WSRC　selects　an　accurate　time-domain　parameter　using　the　K-means　clustering　algorithm　and　computes　the　weighted　coefficients　according　to　the　parameter　distance　between　the　test　sample　and　the　training　samples.　Then,　it　sparsely　represents　the　test　sample　by　solving　a　weighted　$l_{0}$　-norm　problem.　The　goal　of　weighting　is　to　pay　more　attention　to　the　locality　of　the　sample　so　as　to　improve　the　recognition　accuracy.　Finally,　according　to　the　results　of　sparse　representation,　the　fault　status　can　be　identified　through　the　correlation　analysis,　which　can　effectively　solve　the　time-shift　deviation　problem.　The　effectiveness　of　the　proposed　method　is　validated　by　the　experiments　of　rotating　machinery,　and　the　results　indicate　that　the　proposed　method　realizes　fault　classification　with　a　high　accuracy.