An mmWave radar-based mass flow sensor using machine learning towards a peanut yield monitor

A millimeter-wave FMCW radar-based mass flow sensor was developed and evaluated to monitor peanut mass flow rate towards development of a peanut yield monitor. The radar sensor components were placed outside a customized plastic duct in the pneumatic conveyor of a peanut combine. Two systems to simulate the mass flow conditions during harvest were built: one for research-scale mass flow rate using a retrofitted 2-row combine blower and one for commercial-scale mass flow rate using a modified 6-row combine. The ground truth for mass flow rate was obtained and the radar sensor was used to acquire range-velocity image time-series. Two datasets were generated using a sliding window techniqueand several machine learning models (i.e., linear regression, k-neighbor regressor, support vector regressor, random forest regressor, and multi-layer perceptron) were trained to predict peanut mass flow rate from radar data. After evaluation of 5-fold cross validation, k-neighbor regression achived the best performance for the research-scale combine system with an RMSE of 0.14 kg/s, a sMAPE of 15 %, and an R2 value of 0.85, while random forest regression achieved the best performance for the commericial-scale combine system with an RMSE of 0.52 kg/s, a sMAPE of 10 %, and an R2 value of 0.71. Moreover, the sensor can potentially provide the combine operator with information about the velocity of the peanuts to adjust the air pressure of the pneumatic conveyor to reduce undesired peanut shelling. While the peanut mass flow prediction results are promising, further field investigation is necessary to evaluate the effects of noise caused by combine movement, foreign materials, peanut varieties, moisture content, and soil type.