It　is　a　challenging　problem　to　extract　periodic　impulses　submerged　in　the　heavy　background　noise　for　fault　diagnosis　of　rotating　machinery.　Thus,　in　this　paper,　we　propose　a　novel　algorithm　named　tunable　Q-factor　wavelet　transform(TQWT)-based　multi-scale　dictionary　learning　for　dealing　with　this　problem.　The　algorithm　exploits　TQWT　to　decompose　the　measured　vibration　signal　into　different　scales,　and　then　it　adopts　K-SVD　which　can　also　be　replaced　with　other　more　efficient　dictionary　learning　algorithm　to　learn　dictionaries　at　different　scales.　Once　done,　it　employs　a　global　maximum　a　posteriori　estimator　and　inverse　TQWT　to　extract　feature　signal.　By　comparison　with　TQWT-denoising　and　K-SVD-denoising,　the　proposed　algorithm　enjoys　two　main　advantages:　1)　the　dictionaries　learnt　by　our　algorithm　have　the　multi-scale　characteristic　which　is　essential　to　deal　with　non-stationary　signal.　2)　the　dictionaries　are　learnt　from　noisy　signals　itself　and　thus　are　adaptive　to　different　types　of　feature　information.　Effectiveness　of　our　proposed　algorithm　is　demonstrated　by　numerical　simulation　and　fault　diagnosis　of　motor　bearing.