Research And Application Of Pig Sound State Recognition Based On CNN

At present,the pig breeding industry is increasingly intensive,intelligent,and large-scale,and with the changes in living standards and consumption concepts,people have higher requirements for the quantity and quality of meat products.The health status,reproductive efficiency and welfare levels of livestock are the most important factors in determining the benefits of farming.However,it relies only on manual observation,which is costly and difficult to achieve in large farms.In response to this problem,this paper aims to achieve the status recognition of pig sound through deep learning technology and integrated learning technology to promote welfare farming,improve the health of livestock and the survival rate of young animals.Firstly,this paper analyzes and extracts various characteristic parameters of pig sound signals.This is the basis of pig sound state recognition based on integrated learning method and CNN combined with spectrogram recognition method.On the one hand,the time-frequency domain features of pig sound signals are analyzed and extracted,and the 32-dimensional vector sequence is obtained as the input feature vector of SVM and integrated learning model.On the other hand,the spectrogram characteristics of the sound signal are analyzed and studied.Six different types of spectrograms are tested separately,and the optimal method is used to explore the optimal spectrum generation method for adapting to the CNN structure.Secondly,based on the three integrated learning pig sound recognition models based on the Gradient Boosting algorithm,the Random Forest algorithm and the Extra Trees algorithm,the model is tested with the recognition rate and speed as indicators to compare with the SVM baseline model.The experimental results show that the integrated learning algorithm performs better for all kinds of recognition rates,and the difference between recognition speed and SVM is small.The superiority of applying the integrated learning algorithm to pig voice recognition method is verified.Finally,the differences in the recognition rates of different categories in the integrated learning and SVM models are obvious,and the recognition rate of abnormal state sounds is low.This paper focuses on the design,implementation and optimization of the recognition model of CNN combined with the spectrogram.It includes the analysis and preprocessing of the input spectrum and the performance comparison and selection of the existing CNN models.The network structure and optimization techniques of the selected model MobileNetV2 are introduced and improved the original optimization strategy of the network.In addition,this paper further optimizes the two aspects of the spectral map and network structure adjustment.In this paper,Under the condition of the migration training of ImageNet pre-trained MobileNetV2 model,the spectroscopy with 256-point FFT and 1/2-frame shift is taken as the input feature,and Adam is used as the network optimization algorithm,the recognition rate of the final model is 97.3%.The model has a high recognition rate for each category,and can significantly increase the speed of the model with a small loss of accuracy to meet the needs of practical applications.