Online　monitoring　and　diagnosis　of　welding　quality　is　essential　for　intelligent　welding　manufacturing.　The　recognition　performance　of　penetration　for　aluminum　alloy　in　gas　tungsten　arc　welding　(GTAW)　still　needs　to　be　improved　to　meet　the　strict　industry　demands.　This　paper　proposed　a　novel　recognition　method,　time-frequency　image　based　convolution　neural　network　(TF-CNN),　for　GTAW　penetration　recognition.　Time-frequency　images　were　calculated　from　arc　sound　signals　using　short　time　Fourier　transform　and　applied　to　analyze　the　non-stationarity　of　arc　sound.　The　logarithm　of　time-frequency　image　was　taken　to　construct　the　appropriate　input　matrix　of　CNN,　which　was　optimized　to　improve　its　recognition　performance,　including　the　activation　function,　learning　rate　and　architecture　of　network.　The　experimental　results　show　that　the　proposed　TF-CNN　achieved　an　excellent　recognition　performance　with　98.2%　recognition　accuracy　and　0.21　accuracy　variance　for　GTAW　seam　penetration　recognition　and　outperformed　the　traditional　methods.　This　paper　provides　some　guidance　for　the　application　of　CNN　to　other　monitoring　signals　of　intelligent　manufacturing.