Aiming at the problems of large plowing load and low measurement accuracy, a six dimensional force sensor of radial beam type was designed on the basis of classical cross beam structure, which could measure force and moment at the same time. The sensor structure was optimized by simulation method, and the dimension length, width and height of strain beam were determined to be 9mm,10mm and 6mm, respectively. The strain capacity of the sensor structure under load was analyzed, and the position of the strain gauge patch was determined. Based on the calibration data, the improved XGBoost (extreme gradient boosting) machine learning network was used to decouple the force signal. The improved XGBoost model achieved R2P (determination coefficient of test set) of 0.9804, 0.9418, 0.9434, 0.9868, 0.9969, and 0.9822 in six loading modes of force and torque in X, Y and Z directions, respectively. The prediction performance was good, avoiding getting stuck in local optimal solutions. And then compared with the conventional network, the R2P and MAEP (average absolute error of test set) of the improved XGBoost model in the six dimensional force loading direction were significantly better than that of the random forest model and the traditional multiple linear regression. Compared with the traditional multiple linear regression method, the R2P of the six dimensional loading force/moment was increased by 22.57%, 20.99%, 23.32%, 26.27%, 26.05% and 18.72%, respectively. Machine learning based decoupling algorithms could significantly reduce the impact of coupling errors and improve the measurement accuracy of sensors and provide technical support for optimizing agricultural machinery.