LB-MPC drum-speed control strategy for a pepper harvester with feed-rate feedforward and torque-observation-based correction

• An LB-MPC method was developed for drum speed regulation under variable feeding conditions. • Vision-based feedrate feedforward and torque-observed BLM correction were jointly integrated. • GPR-based residual learning and chance constraints were introduced to improve torque safety. • The proposed method achieved real-time implementation in simulation and control experiments. • Its effectiveness was validated by 180 bench tests and field trials with near-zero torque exceedance. To mitigate drum-speed excursions and overload/impact risk caused by feed-rate fluctuations under strongly time-varying and stochastic fruit load, a learning-based Model Predictive Control is proposed with dual objectives of banded speed stability and torque safety. Firstly, vision-based short-horizon feed-rate prediction is converted into a drum-side equivalent-load prior and injected into nominal Model Predictive Control as a feedforward term, enabling earlier power matching and mitigating speed oscillations caused by incoming-flow uncertainty. Secondly, measured drive torque is treated as an online observation of equivalent load, and a Bayesian Linear Model is used to recursively learn the correction mapping between the vision prior and torque observation while providing uncertainty quantification. Finally, Gaussian Process Regression is adopted to learn the residual of nominal torque prediction online, where the residual mean corrects torque forecasts and the residual variance tightens chance constraints to suppress torque excursions at a prescribed confidence level. Comparative simulations against nominal Model Predictive Control and robust Model Predictive Control show that learning-based Model Predictive Control reduces speed RMSE by 66.1%, markedly decreases constraint violations, and completes online optimization within 18.4 ms. Bench tests with 180 repeated runs on real pepper plants and a physical drivetrain further demonstrate a 51.8% reduction in speed RMSE, a 67.1% shorter settling time after step disturbances, and zero torque violations. Field trials confirm deployability under practical operating conditions: at travel speeds of 0.6–0.8 m/s, band compliance remains around 96% and torque exceedance events are nearly eliminated. Results indicate that learning-based Model Predictive Control delivers reliable banded speed regulation and torque-safe operation under strong feeding variability and model uncertainty, offering a practical control route for safety-constrained intelligent agricultural machinery.