Research On Feeding Algorithm For Pregnant Sows Based On Recurrent Neural Network

The pig farming industry is the pillar industry of China’s aquaculture economy.In recent years,China’s aquaculture industry has steadily advanced from mechanization and scale to automation and intelligence,and has achieved significant development.However,most farms have non-standard feeding management methods and unscientific feeding plans,which directly affect the economic benefits of the farms.The feeding and management of pregnant sows is crucial for the production efficiency of breeding farms.In China,most methods such as fuzzy control and expert systems are used to predict the feeding amount of pregnant sows.However,these methods rely on expert knowledge and extensive experience,and potential information and complex interference factors can lead to system uncertainty,with a large range of system prediction errors.In order to improve the management level of pregnant sows in the breeding farm,achieve precise feeding of pregnant sows,improve the litter yield and healthy litter size of pregnant sows,and save feed,this article builds a feeding volume prediction model based on the Gated Recurrent Unit(GRU),develops a feeding management system,and achieves precise feeding of pregnancy.The main work of this article includes the following aspects:(1)Establish a predictive model for the feeding volume of pregnant sows based on recurrent neural networks.In view of the characteristics of multiple factors affecting feeding amount,this paper screened the factors based on expert experience and reading literature,constructed a data set based on the gestation date,backfat thickness,parity and feeding amount of sows,built a GRU model,adjusted the network structure optimization model,selected the optimal model through the evaluation index results,and compared it with convolutional neural network(CNN),recurrent neural network(RNN),and short-term memory(LSTM)models,Compared with RNN and LSTM,the mean square error(MSE)of GRU decreased by 14.89% and 3.92%,respectively,indicating that the GRU model can effectively predict the feeding amount of pregnant sows.(2)Propose an improved model for predicting the feeding volume of pregnant sows.This article analyzes the factors that affect feeding volume.In response to the low weight of neural network parameter updates due to parity characteristics,principal component analysis(PCA)and CNN were used to reduce the dimensionality of the dataset.The parity information was fused with pregnancy date and backfat thickness,and PCA-GRU and CNNGRU models were constructed.The network structure was adjusted to select the best prediction model.The results showed that compared to traditional GRU models,The MSE of the CNN-GRU model was reduced by about 5.23%,while the MSE of the PCA-GRU model was reduced by about 3.91%.Experiments have shown that using the method of reducing data dimensions can effectively improve the predictive performance of the model.Moreover,by reducing the dimensions of the dataset,the number of model parameters can be reduced and training speed can be improved.(3)Design and construction of a feeding management system for pregnant sows.This article designs and develops a system structure based on the functional requirements of the breeding farm.A feeding management system is built based on the Django framework,connected to the My SQL database,and implemented functions such as employee information and permission management,pregnancy sow information management,pregnancy report and transfer report,abnormal feeding alarm and ear tag loss alarm,statistics of pregnancy sow entry times,and feed consumption statistics.