The Study On Estimation Method Of Forage Intake Of Tibetan Sheep Based On Acoustic Analysis

The south bank of Qinghai Lake in Hainan Prefecture is located in the northeast of the Qinghai-Tibet Plateau.It is not only an important base for Tibetan sheep production,but also has an extremely important ecological barrier.In recent years,overgrazing has led to grassland degradation and land desertification in some areas.Studying the balance of grass and livestock,determining a reasonable scale for carrying animals,alleviating the contradiction between human and grass,and conducive to the transformation and upgrading of the local animal husbandry to modern and ecological animal husbandry,and achieving the sustainable development of animal husbandry.Zoning rotation can improve the utilization rate of pasture,and improve the livestock production per unit area of grassland by increasing the stocking rate.It shows great advantages under the condition of high stocking rate.The monitoring of Tibetan sheep’s feed intake is the key to the area-based rotation grazing,which reflects the content of the Tibetan sheep’s nutrients from the grassland,which can provide data support for the development of area-based rotation grazing programs.Under grazing conditions.Tibetan sheep have a wide range of activities.It is difficult to monitor their feeding and rumination behaviors by using general methods such as regional forage sampling difference method,chewing behavior observation simulation method,and indigestible indicator labeling method.In addition.the above methods will affect the normal behavior of Tibetan sheep during the monitoring process.This paper has designed and implemented a set of Tibetan sheep feed intake estimation system.Using sound signal processing technology to identify the bite forage sound signals,ingestion chewing sound signals,ruminant regurgitation sound signals,and ruminant chewing sound signals during the feeding process of Tibetan sheep,extracting behavioral measurement signals(BMS)and acoustic measurement signals(AMS)from the audio signals of ingestion chewing.Based on the Pearson product moment correlation coefficient,an optimized Pearson metric signal(PMS)was constructed,and a random forest regression and elastic network method was used to establish a Tibetan sheep feed intake estimation model based on AMS,BMS,AMS&BMS,and PMS.The Web client is used as a display platform to display the change curve of the daily feed intake of Tibetan sheep,so that the sheep farm management personnel can discover the change of Tibetan sheep feed intake in a timely manner through the system,and adjust it in time according to the Tibetan sheep grass feed and the livestock load in each division of the pasture Rotate pastures to avoid overgrazing.The main work completed in this paper is as follows:(1)Design the Tibetan sheep feeding audio collection system and complete the construction of the Tibetan sheep herbivorous audio data set,use Wi-Fi technology as the communication solution for audio collection and transmission,and transfer the original audio when the Tibetan sheep feeding behavior occurs to the PC,The PC end saves the Tibetan sheep feeding audio to achieve real-time and stable acquisition of Tibetan sheep feeding audio.The actual herbivorous amount corresponding to each feeding behavior is weighed and recorded by a dedicated tester with accurate recording accuracy to 0.1g;(2)Audio preprocessing and feature extraction of Tibetan sheep feeding.In this paper,the LSA-MMSE speech enhancement algorithm was used to denoise each piece of feeding audio,and the endpoint detection algorithm of the double threshold method was used to automatically intercept the non-silent fragments in the feeding audio,and the non-silent fragments were manually marked as Ingestion Bite(IB),Ingestion Chewing(IC),Rumination Regurgitation(RR),Rumination Chewing(RC),Noise(N).Wavelet transform was introduced on the basis of traditional Mel cepstrum coefficients to extract wavelet Mel cepstral coefficients(W＿MFCC)of five types of non-silent audio clips,and PCA dimensionality reduction was used(the dimensionality-reduced audio feature coefficients are recorded as PW＿MFCC)to reduce the computational cost of constructing an audio classifier based on PW＿MFCC in the later period.(3)A voice recognition classifier based on Long Short-Term Memory Model Network(LSTM)for Tibetan sheep and sheep grazing behavior was constructed.The data sets of biting pasture,chewing pasture,ruminant regurgitation,ruminating chewing,and other non-silent sounds were divided into training and test sets.MFCC,W＿MFCC,and PW＿MFCC characteristic coefficients.The feature coefficients of the training set data are used to train the sound behavior classifiers based on MFCC,W＿MFCC and PW＿MFCC respectively.The test set data is used to test the performance of each classifier,and F1-Score is introduced as the model quality evaluation standard.The results show that when PW＿MFCC is used as the feature input,the LSTM model gets the highest score value of 94.62%in the training set;when W＿MFCC and MFCC are used as the feature input,the LSTM model scores in the training set are 93.89%and 88.35%,based on PW＿MFCC The feature LSTM model has the best classification effect.(4)Study the estimation model of Tibetan sheep feed intake.Extract the herbivorous chewing audio segment recognized by the sound behaviorrelated sound classifier,extract the number of chewing(C),total duration of chewing sound(TC),average chewing time(TCc),average intensity(VdB),The average sound intensity(V),the total energy flux density of chewing(E),the energy flux density per unit chewing(Ec)seven variables.The least squares method was used to predict the feed intake of univariate.For individual univariate prediction,parameter C performs best,and R~2 in the test set can reach 76.47%.The seven variables C,TC,TCc,VdB,V,E,and Ec construct three combined variables,namely AMS,BMS,AMS&BMS.Among them,AMS was expressed as {VdB,V,E,Ec},BMS was expressed as ｛C,TC,TCc},and the combined variable AMS&BMS was expressed as {C,TC,TCc,VdB,V E,Ec｝,according to Pearson The product moment correlation coefficient selects three variables that have a high correlation with the feed intake variable to form the PMS variable,and the PMS was expressed as {C,TC,E}.Random forests and elastic networks were used to construct multivariable feed intake estimation models.The model verification results showed that the R~2 score of the elastic network model with AMS&BMS as input was 87.75%,the R~2 score of the random forest prediction model with PMS as input was 86.2%,and the elastic network prediction model with PMS as input had the highest R~2 score of 88.52%;(5)Designed and built a software platform for Tibetan sheep feed intake estimation system,based on Tibetan sheep chewing sound monitoring algorithm and feed intake estimation model,built a database based on the monitoring system requirements,and stored and displayed Tibetan sheep feed intake data,Use python language to realize the design of middleware,in order to facilitate the administrator to view,design the web front-end display page,extract the feed intake forecast data from the back-end database and display it in the form of a chart,the system completes the monitoring of Tibetan sheep chew,Feed intake estimation,Tibetan sheep health warning,etc.,provides a modern solution for regional rotation grazing and Tibetan sheep health monitoring.