In order to realize the real-time on-line detection of grain protein content and record the sampling geographical location information during combine combine-harvester harvest grain, an in-line detection system of grain protein content based on the principle of near-infrared spectroscopy was developed, which was mainly composed of near-infrared spectral sensor module, spiral sampling and conveying mechanism, control module, GPS/Beidou positioning module, industrial display integrator, etc. When the grain combine-harvester near-infrared spectral protein content in-line detection system was working, when the grain discharged by the combine-harvester grain outlet was through the spiral sampling and conveying mechanism, the stepper motor of the sampling mechanism was controlled by the controller according to the detection rate requirements and intermittent grain transmission, the controller system also controls the near-infrared spectral sensor to sample the spectral when the stepper motor stops turning, and the data such as the grain near-infrared spectrum and the positioning signal of GPS/Beidou positioning module were transmitted to host computer by RS485. The control and data processing analysis software of near-infrared sensor and sampling mechanism was compiled, and the grain protein, sampling location information, etc. were displayed and saved in real time after the grain protein prediction model. In order to verify the performance of grain protein content prediction model and online detection system, indoor calibration and field system dynamic testing were carried out, and the decision coefficient of wheat protein content prediction model was 0.865, the absolute error range was -0.96 to 1.22, and the relative error range was -7.30% to 9.53%, the root mean square error of prediction(RMSEP) was 0.638, the decision coefficient of the rice protein content prediction model was 0.853, the absolute error range was -0.60 to 1.00, the relative error range was -8.47% to 9.71%, and the RMSEP was 0.516. The results of the system dynamic field test shows that the maximum relative error of wheat protein content was -6.69%, the maximum error of rice protein content was -8.02%, the system was not significantly affected by sampling and analysis interval, and the system stability and detection accuracy meet the need of grain protein online detection in the field, which provides a scientific basis for precision agricultural operation.