Abstract: In order to track fillet weld with small bending angle and detect endpoint of weld seam, the welding current was filtered by spatial median filtering, average filtering, threshold processing and linear filtering depending on the type of noise in welding current, and finally most of noises were removed. Tracking of fillet weld with small bending angle was realized through connecting the points corresponding to average vector of two types of model by line, considering slope of straight line as deviation, regarding deviation in the last three times as input of perceptron for multiple pattern classes, and taking the rate of horizontal slider stretch as output. It was found that the robot arrived at the endpoint of weld seam if three consecutive deviation values were larger than 1. Tracking experiment of fillet weld with small bending angle was conducted, the endpoint of weld seam was detected, and drain hole experiment was carried out. The results show that the algorithm improved the precision of tracking fillet weld with small bending angle. The accurate recognition of endpoint of fillet weld and through drain hole was realized, which was conducive to the welding automation.