AI-Assisted Adaptive Feed-Roller Control of the Saw-Cylinder Load in Cotton Linter Machines: An Extended Kalman Filter–Sliding-Mode–Neural-Network Framework

The saw cylinder of a cotton linter machine is loaded through a rotating seed-cotton (seed-roll) mass whose density is set by an upstream feed roller; because this density is coupled nonlinearly to the cylinder speed and is not directly measurable in production, conventional fixed-gain feed-rate regulators struggle to reject the load and density disturbances that arise from cotton-grade, trash-content, and moisture variation. This paper develops an AI-assisted adaptive control architecture for the feed roller that (i) reconstructs the unmeasured seed-roll density from the cylinder encoder and the motor-current load signal using an Extended Kalman Filter (EKF), (ii) drives the estimated density to a speed-tracking set-point through a cascade feedback-linearising sliding-mode law, and (iii) compensates residual, imperfectly identified feed-side nonlinearity with an on-line-adapted radial-basis-function (RBF) neural network trained under a Lyapunov-based update law. A fourth-order Runge–Kutta (RK4) simulation of the coupled saw-cylinder/seed-roll dynamics is used to benchmark the proposed EKF-SMC-NN scheme against a classical PID regulator and a linear active-disturbance-rejection controller (ADRC) under combined torque-ripple, sinusoidal, and step feed-density disturbances. A local observability analysis confirms that the (angular-velocity, density) pair remains fully observable across the operating envelope (rank 2, condition numbers 21.6–35.9). Under the nominal disturbance scenario the proposed controller reduces the integral absolute error by 90.1% relative to PID and 64.5% relative to ADRC, the integral squared error by 99.4% and 92.7% respectively, and the steady-state tracking RMSE by 83.9% and 60.8% respectively, while the EKF recovers the hidden density with an RMS error of 0.030 kg/m³. A 40-trial Monte-Carlo sweep with the seed-roll load coefficients perturbed by ±25% confirms the transient-response advantage of the proposed scheme, while also revealing a steady-state sensitivity that motivates gain-scheduling as future work. The results indicate that combining a physically structured EKF observer with a sliding-mode/RBF-NN control law offers a practical route to density-aware, load-adaptive automation of the linter feed process, with the observability and robustness analyses providing design guidance for implementation on programmable-logic-controller (PLC) and variable-frequency-drive (VFD) hardware already used in gin and linter plants.

Перевод пока недоступен