In order to accurately detect walnut weight online, a walnut weight online testing system based on multi-sensor information fusion was designed, which was adopted the method of dynamic compensation and calibration to process signal collected from an acceleration sensor and symmetrical weight sensors and used the algorithm of support vector regression (SRV) to predict the walnut weight. Then, 400 walnuts were used to acquire weight data at speed of 0.02, 0.03 and 0.05 m/s respectively based on this system. Meanwhile the data were trained and verified. The better prediction model of walnut weight was determined as SVR model with linear kernel function, and the best test speed was determined as 0.03 m/s. Finally, 200 walnut samples were tested online at the speed of 0.03 m/s. The results showed that r² of linear fitting between the walnut weight prediction and their actual weight was 0.85, and the average absolute error of linear fitting was 1.67 g. The results indicated that the system can accurately online test walnut weight.