Abstract: The recognition of defections is still a difficulty in non-destructive testing field. A new method for recognition of weld defects based on multi-probe source data fusion was proposed in this paper,which improved the reliability of detection and accuracy of defection recognition. Several welds,containing defects of hole,slag and crack,lack of penetration and lack of fusion were respectively inspected by two probes which possessed different angles of incidence. Then the ultrasonic signal features of defect echo were extracted. Finally,an intellectualized pattern classifier with two-level feature fusion and decision fusion was developed to realize the defect recognition with data fusion. BP neural network was selected as the classifier of feature fusion to obtain the basic probability function of each probe and probability value of each type of defect. Then D-S evidential theory was used to combine the probability function of each probe and to carry out the defect recognition. The results show that the multi-probe information could be effectively fused,and the recognition rate of weld defect was improved.