Abstract: In order to realize the underwater auto-welding based on rotating arc sensor and get high accuracy tracking,it is necessary to study the seam offset identification algorithm.First,the wavelet and median filter methods were used to process welding current signals,and then the signal was divided into cycle and normalized.PCA was used to remove the self-correlation of data set and reduce the number of inputs of Nu-SVR.The result showed that the maximum error and mean error of PCA_Nu-SVR was 0.95 mm and 0.65 mm.The precision of PCA_Nu-SVR was as good as Nu-SVR,and better than interval integral method and neural network.The runtime of PCA_Nu-SVR was more than interval integral method,and less than neural network and Nu-SVR.