In　view　of　the　urban　water　supply　pipeline　leak　detection,　the　method　of　leak　acoustic　signal　recognition　is　studied.　The　features　of　time-domain,　frequency-domain　and　waveform　of　the　leakage　signals　are　analyzed,　20　features　which　can　be　used　to　characterize　the　leakage　signal　are　extracted.　Based　on　the　features,　the　BP　neural　network　identification　system　for　leakage　acoustic　signal　is　constructed.　The　influences　of　the　neural　network　structure　(the　number　of　hidden　nodes,　transfer　function,　learning　rate)　and　the　number　and　type　of　the　input　parameters　on　the　leakage　signal　recognition　performance　are　studied,　the　best　structure　and　input　parameters　of　the　neural　network　are　optimized.　Based　on　the　above　research,　the　optimized　neural　network　was　used　to　cross-train　and　identify　the　leak　signal　of　the　laboratory　and　water　supply　pipelines.　The　overall　recognition　rate　reaches　92.5%.　The　results　show　that　the　neural　network　system　based　on　the　leakage　features　has　high　reliability　and　universality,　which　can　be　well　recognition　the　leakage　signals　under　different　scenarios.　The　research　work　has　done　a　useful　exploration　to　solve　the　leakage　signal　identification　under　different　working　conditions.　©　2016,　Science　Press.　All　right　reserved.