It　is　crucial　point　that　accurate　temperature　in　distributed　optical　fiber　temperature　sensor.　In　order　to　solve　the　problem　of　temperature　measurement　error　due　to　weak　Raman　scattering　signal　and　strong　noise　in　system,　a　new　based　on　improved　wavelet　algorithm　is　presented.　On　the　basis　of　the　traditional　modulus　maxima　wavelet　algorithm,　signal　correlation　is　considered　to　improve　the　ability　to　capture　signals　and　noise,　meanwhile,　combined　with　wavelet　decomposition　scale　adaptive　method　to　eliminate　signal　loss　or　noise　not　filtered　due　to　mismatch　scale.　Superiority　of　algorithm　filtering　is　compared　with　others　by　Matlab.　At　last,　the　3km　distributed　optical　fiber　temperature　sensing　system　is　used　for　verification.　Experimental　results　show　that　accuracy　of　temperature　generally　increased　by　0.5233　square.