A Neural Network Approach to Predict Quality of Pellet Feed Basing on Materiel and Process Parameters

Abstract. With the development of feed industry, the proper way to reducing production cost while keeping quality gained more attention from producer. While processing of pellet feed as main part of formula feed has been confirmed as an extremely complicated system due to varieties of hydrothermal conversion as well as stress fluctuation which effected by lots of factors. Meantime, the comprehensive relationships in the process including nonlinear relation between these factors bring difficulties for exact quantitative mathematical model to describe the relationship between these effected parameters and production quality. Therefore, if a particular tool can be applied to forecast production quality in an certain reliability based on effected parameters before actual production, it will assist producer to make targeted adjustment of these effected parameters to promote production quality in time which avoiding the risk of loss of benefits. As a typical Black Box Approach, model established by Artificial Neural Network (ANN) is based on inside relation between inputs and outputs which overcome the limitations of traditional mathematical model facing the complexity of feed processing, avoiding interference of human factors in maximum. For this reason, the Back-Propagation algorithm was selected to modeling the processing of pellet feed and achieves quality prediction on account of effected parameters. Basing on livestock feed, an indicator system consisted by inputs (processing parameters and diet character) as well as outputs (Pellet Durability Index and pellet hardness) was established. Meantime, the ANN model was actualized applying a toolbox in the MATLAB software using the Levenberg-Marquardt algorithm which shows high stability and fast convergence rate. After a preliminary experiment, the structure of the model was designed as four layers include input layer, output layer and two hidden layers with 7 and 8 neurons accordingly. By analyzing results of the testing, it can be concluded that the utilization of ANN in predicting the product quality of pellet feed was feasible with relative ideal results. The application of this method is capable of saving time and money costing for proper processing parameters search as well as product quality fluctuation in a certain degree, which is meaningful to the profit growth of enterprise and the rational use of social resources.