This　work　aims　at　a　novel　approach　to　estimate　the　root-pass　penetration　towards　its　feedback　control,　in　which　the　real　penetration　is　measured　by　the　backside　bead　width.　The　major　challenge　is　that　it　happens　under　the　workpiece　and　likely　cannot　be　directly　observable.　The　dynamic　evolution　of　the　weld　pool　surface　has　been　analysed　to　design　an　active　vision　method　monitoring　the　pool　surface,　yet　fundamentally　correlated　to　the　unobservable　penetration.　The　designed　convolutional　neural　network　model　is　trained,　validated,　and　tested　for　recognising　the　weld　penetration　with　satisfactory　accuracy.