Adaptive Profiling Control for Tea-Plucking Machines via Spatiotemporal Perception Reconstruction and Feedforward–Feedback Anti-Windup Control

Mechanized harvesting of bulk tea in unstructured hilly tea gardens requires accurate dynamic profiling of irregular canopies at a constant cutting depth to avoid harvesting coarse and old leaves, missing tender shoots, and reducing yield and quality. This task is complicated by high-frequency vibrations from crawler travel; perception disturbances caused by impacts with hard, thick branches; and actuator saturation and control chattering when abrupt chassis-posture changes occur under motor torque constraints. To address these challenges, this study proposes a cooperative control architecture that integrates spatiotemporal perception reconstruction with adaptive anti-windup control. First, a global inverse-kinematics mapping model that fuses an inertial measurement unit (IMU) with an absolute encoder is developed. The model decouples chassis motion from local terrain measurements and establishes a unified spatial reference for constant-depth actuator profiling. Second, a spatiotemporal perception reconstruction mechanism is designed for composite multifrequency noise. In the amplitude domain, a robust outlier-rejection kernel adaptively suppresses non-Gaussian transient pulses; in the temporal domain, environmental background-noise estimation and abrupt-change activation dynamically adjust the filtering bandwidth and reduce phase lag in absolute elevation reconstruction during terrain steps. Finally, a nonsingular fast terminal sliding-mode control (NFTSMC) law incorporating active feedforward and anti-windup compensation is developed to address the servo motor torque limit. The law uses chassis pitch angular velocity for feedforward compensation and combines a continuous boundary layer with an auxiliary dynamic system to correct the error manifold, thereby reducing integral windup and high-frequency servo chattering caused by torque saturation. Comparative simulations show that, under extreme multifrequency composite conditions, the global root-mean-square (RMS) tracking error is 41.1% and 35.8% lower than those of Kalman filter-proportional-integral-derivative control (KF-PID) and standard moving-horizon-estimation sliding-mode control (SMHE-SMC), respectively. Field evaluations across 30 measurement areas showed that under the Challenging Scenario (combined slope, speed, and density variations), the proposed architecture maintained mean missed-plucking and loss rates of 0.46% and 0.43%, respectively. ANOVA confirmed a significant main effect of the control strategy on the missed-plucking rate (p < 0.001, ηp2>0.9). Additionally, a structured component-wise ablation study was conducted to evaluate the specific contributions of the core modules. The results indicated that the omission of mechanisms such as inverse-kinematics mapping or anti-windup dynamics notably degraded the overall tracking accuracy by inducing uncompensated spatial drift and saturation-induced errors. These results indicate the potential of the proposed architecture for closed-loop profiling control in intelligent agricultural machinery.