Quality prediction model of pellet feed basing on BP network using PSO parameters optimization method
Chen Xiao,Wang Hongying,Kong Dandan,Yue Yan,Fang Peng,Lü Fang
DOI: https://doi.org/10.11975/j.issn.1002-6819.2016.14.041
2016-01-01
Abstract:For a large number of advantages such as better palatability, high return, and avoiding automatic grading, pellet has been one of the major application forms of animal feed. Aiming at the fluctuation of pellet feed quality due to the change of diet content and processing parameters, a prediction model was proposed in this article to provide assistance for quality and cost control in feed industry. In this research, back propagation neural network (BPNN) was designed as the core for the proposed model considering its advantages such as unique abilities of self-organizing, self-learning and self-adaptation. For the purpose of improvement of data utilization efficiency, mean impact value (MIV) method was combined in this model as data preprocessing technique for its concise and rapid feature in data processing. As for several crucial structural parameters within BPNN model, particle swarm optimization (PSO) algorithm was applied for better performance since its advantages of simple structure, easy realization, fast search speed, etc. Meanwhile, these 2 methods were chosen to cooperate with BPNN algorithm for their preferable collaborating properties among various mathematical models according to literatures. Based on index system of inputs and outputs which met practical requirement of industry, the dataset for the model included diet content and processing parameters as input and powder content, productivity and pellet durability index (PDI) as output, which were collected from actual feed production processing in a feed mill in Beijing in the period of March-December, 2015. And the model structures including neuron number of hidden layers and proportion of dataset for different purposes were established by pre-test method/literature experience. After trained and tested by collected data, the MIV-PSO-BPNN prediction model was established and showed fairly good performance in following aspects. Analyzing the fitting optimal linear regression curve of predicted value produced from the proposed model and true value from the collected data, it showed that there was a significant positive relationship between these 2 values with determination coefficient of 0.9692, 0.9421 and 0.9465 for powder content, productivity and PDI, respectively. Meanwhile, the fitting curve presented a highly coincident relation with the line of y=x, which meant predicted value was very close to true value. Also, high prediction performance of MIV-PSO-BPNN prediction model was verified based on mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE). In order to compare predict function, the multi-linear regression method and basic BPNN model were established with the same dataset for training and test. Results showed that MIV-PSO-BP prediction model possessed a fairly high promotion in prediction function compared to multi-linear regression method for its unique adaption quality to non-linear properties of feed processing which was attributed to the advantage of BPNN model as the core. Particularly, the average falling range of MAE, MAPE and RMSE value for production rate, PDI and powder content reached about 90.77%, 86.36% and 77.85%, respectively. On the other hand, compared to basic BPNN model, the average optimization range of error index of powder content, PDI and production rate for MIV-PSO-BP prediction model reached 43.85%, 68.93%and 58.08% respectively. So, the MIV-PSO-BPNN prediction model possessed a better forecast stability, which was indicated by the changing tendency of relative error for each output and its curve being more stable and smooth while there was several significant peaks and beats in the other. Statistically, the average decreasing of extreme value difference of relative error for 3 outputs reached 91.46%. The proposition of this model provides a novel thought for feed industry as well as the theoretical reference and practical tool for production and quality control of pellet feed with high efficiency and low consumption. However, there are several problems that need to be solved. Parts of parameters such as initial threshold and weight of BPNN as well as random value in PSO updating process are determined randomly, which may cause precision dissatisfaction of users and requires more time for model run. In following research, targeted measure should be taken for initial parameter assignment to make this model provide better service for relative practitioners.