Application of Neural Network for Modeling the Coal Gleaning Performance of the FGX Dry Separator

摘要

The FGX Dry Separator is a gravity concentrator that has been recently developed to remove high density rocks from run-of-mine coal samples. A recently completed FGX optimization study by the authors at Southern Illinois University developed multiple regression models to predict the coal cleaning performance of the FGX Dry Separator as a function of the key process variables of a 10 tph FGX prototype unit. The objective of the present study was to develop an Artificial Neural Network (ANN) model to predict the coal cleaning performance of the FGX Dry Separator and to compare the utility of the ANN-based model with that of the regression model developed in the previous study. The process variables that are known to affect the performance obtained from a FGX Dry Separator include feeder frequency, longitudinal deck angle, lateral deck angle, deck vibration frequency, product splitter position, tailings splitter position, fluidization airflow rate and baffle plate height. The ANN-based model developed in this study utilized all eight process variables as input parameters. The output parameters included ten different process responses: yield to clean coal, combined yield to clean coal and middling, combined ash rejection to middling and refuse, ash rejection to refuse stream only, separation efficiency for clean coal, separation efficiency for clean coal and middling, combustible recovery to clean coal stream, combined combustible recovery to clean coal and middling streams, and clean coal ash and refuse ash contents. All data obtained from 72 tests conducted applying the FGX Dry Separator using a specific coal sample was utilized for developing and validating the ANN-based model. The ANN-based model was trained using the results obtained from several randomly selected tests conducted during the abovementioned optimization study. The back-propagation method was utilized following a non-linear steepest descent type of algorithm to optimize the weights to minimize the final error. Finally, the ANN-based models were validated using the remaining test results. The comparative results indicate that the ANN can also be used as an alternative approach for modeling the performance of the density-based separators like FGX Dry Separator.