Experimental study on DEM parameters calibration for organic fertilizer by the particle swarm optimization − backpropagation neural networks

Abstract

Abstract In order to calibrate the properties of the organic fertilizer particles, this work employs an integrated strategy that combines simulations, machine vision techniques, and physical experiments. Through physical testing, the fundamental physical characteristics of the organic fertilizer particles were identified. The initial analysis was through the Plackett-Burman test. The parameters that greatly influence the angle of repose are established. The previously identified important variables were optimized by the Central Composite Design test. The regression fitting models of the BP neural network have been developed from the data set derived from the Central Composite Design test results. Genetic algorithms (GA) and particle swarm optimization algorithms (PSO) were used to optimize the BP neural network. The R2MAE and RMSE of the BP, GA − BP, PSO − BP and RSM regression models were compared and analyzed. The results showed that PSO − BP algorithm could achieve better fitting effect, and could construct a prediction model with higher accuracy and less error to analyze the repose angle of the organic fertilizer particles. The PSO − BP algorithm was used to iterate until the individual with the closest fitness was obtained. CORO−p was 0.35, COSO−O was 0.49, COSO−p was 0.29 and CODO−O was 0.38 were the optimal parameter combination.