Evaluation and Simulation of Tractive and Soil-tool Interaction in Tillage: Case of Hard-setting Ferralsols in North-rift Kenya

Abstract

This study aimed at evaluation and simulation of the mechanisms underlying tractive and soil-tool interface interactions during tillage, in the mechanized region of Northrift, Kenya. The research investigated soil physio-mechanical parameters affecting strip tillage in situ, evaluated their influence on soil-wheel and soil-tooling interactions, and simulated draft, tractive performance, and fuel consumption. Triplicated randomized sites were traversed into five profile pits, referenced for soil testing to establish moisture content (θSoil), cone index (CI), bulk density (γSoil), Plasticity index (IPSoil), angle of internal friction (ϕ), cohesion (c) and shear strength (τShear) at 4 levels of strip tillage depth (Tdeth) i.e 0-10, 10-20, 20-30, and 30-40cm. Before tillage experiments, soil-tire contact patch areas (AStc) of the research tractor (Case IhJXM 90hp) were obtained at 5 tire inflation pressures (Ptire) i.e.110.4, 151.8, 193.2, 234.6 and 275.8kPa) at 4 levels of rear wheel load (Wload-11.3, 11.8, 12.3 and 12.8kN), using image-pixel-color correlation and segmentation tool in MatLab. Theoretical wheel slip (Swheel) and rolling resistance (FRr) were obtained at all the Wloads, and Ptire for no Tdeth. Thereafter, tractive locomotion was engaged using Case IhJXM90hp, dynamometric with auxiliary John Deere 5503 tractor, at all the 5 levels of Ptire, 4 levels of Wload, and at 4 levels of Tdeth with triplications. Fuel consumption (ØFuel) data of the research tractor was remotely relayed from the fuel tank, digitally instrumented with Teltonika FMB920 smartphone telemetry. Obtained tractive forces (FTr) and FRr were averaged to determine net drawbar pull (HDbp) and tractive efficiencies (ηte). Experimental variables (θSoil, CI, γSoil, IPSoil, τShear, AStc, SWheel, FRr, FTr, HDbp, Ƞte, Wload, Ptire and Tdeth) were used in modelling 72 Artificial Neuron Networks (ANNs) for predicting (ØFuel) rate, using 12 neurocomputing algorithms (NCAs) learning on logsig, tan-sig, and purelin activation functions. These included Levenberg-Marquardt, Quasi- Newton, Powell-Beale conjugate gradient, Scaled conjugate gradient, Fletcher-Reeves conjugate gradient, Polak-Ribiére conjugate gradient, One step secant, Bayesian regularization, Resilient backpropagation, Gradient descent, Learning rate gradient descent and Gradient descent with momentum. Their metacognitive robustness in predicting Øfuel was evaluated using a broad criterion of regressed neuron layer input-output functions i.e MSE, R, R2, RMSE, mean absolute error (MAE), mean absolute percent error (MAPE), sum squared error (SSE), T-scatter, coefficients of variation (CV) and prediction accuracy (PA). Simulations revealed that 2nd order pedotransfer function polynomials correlated most of the soil physicomechanical properties with Tdepth. Vertical tire deflection (δv) was exponentially correlated with Wloads but quadratically curvilinear with Ptire, while Wload was linearly quadratic with increasing AStc. Increased Ptire was quadratically curvilinear with reducing δv and AStc, while δv ratios declined with Ptire levels curvilinearly. Although increased Wloads developed higher HDbp, increased Tdepth led to a greater increase in Swheel over and above that emanating from increased Ptire due to high draft demand. Increase in Tdepth, increased Swheel at a higher rate than increased Ptire even at reduced Wloads...