As the characteristic parameters of a multi-sensing keyhole reflect effectively the welding quality of high power lasers, this paper researches the extraction method of keyhole characteristic information and establishes a prediction model for welding formation. By taking a high power disk laser to weld 304 austenitic stainless steel plates for an example, a near-infrared high-speed camera and an X-ray vision imaging system were used to capture the molten images in welding processing and to obtain the keyhole region by image processing. The invariant moment characteristics were extracted from near-infrared visual images by the moment method, meanwhile the keyhole area and ordinate value of the keyhole forefront were calculated as the characteristic parameters. Depth and entropy of the keyhole were extracted from X-ray visual images. In different laser powers, the keyhole characteristics were obtained and three BP (Back Propagation) neural network models were set up through feature fusion of all the characteristic parameters. The relationship between the keyhole formation, welding condition and welding state was explored and the on-line monitoring for welding process was implemented. Experimental results show that the average absolute value of relative errors between predictive and measured values of weld width and penetration are 0.18 mm and 0.57 mm, respectively through fusion analysis and principal component analysis on characteristic parameters of two sensors, and they have been reduced by about 0.03 mm and 0.31 mm as compared with that of a single sensor. The proposed method can be applied to monitoring high-power disk laser welding quality in real time.