Accurately,　apace　and　intelligently　identifying　the　diverse　faults　of　rotating　machines　is　of　great　significance.　However,　high　diagnostic　accuracy　is　usually　accompanied　by　lower　model　efficiency.　To　address　this,　a　light　intelligent　diagnosis　model　based　on　improved　Online　Dictionary　Learning　(ODL)　sample-making　and　simplified　Convolutional　Neural　Network　(CNN)　is　proposed.　Within　the　sampling　time,　ODL　based　on　Orthogonal　Matching　Pursuit　(OMP)　is　used　to　select　time-domain　multi-channel　signals　to　make　RGB　samples,　which　results　in　samples　with　smaller　size　and　stronger　features.　Benefiting　from　the　high-quality　samples,　the　CNN　model　is　simplified,　only　small-scale　one-dimensional　convolution　kernels　that　undertake　different　tasks　and　global　average　pooling　(GAP)　layer　are　used,　which　greatly　improve　diagnostic　efficiency　of　the　network　while　ensuring　diagnostic　accuracy.　Three　different　fault　diagnosis　cases　of　rotating　machine　suggest　that　the　proposed　model　has　high　diagnostic　accuracy　along　with　high　efficiency.