The Cattle Collar Device And The Classification Of Cattle’s Individual Behavior

The behavior of cattle is an important manifestation of their health status.Accurate and efficient monitoring and analysis of cattle behavior can help understand their physiological,health,and welfare status.Taking cattle as the research object,and a cattle collar based on multi-sensor fusion was designed.The collar can collect four series of data,including cattle’s motion acceleration,angular velocity,body surface temperature,and position information.At the same time,it can also collect three audio data,including the cow’s normal call,the call during oestrus,and the swallowing sound.Then,pre processing such as enhancement,denoising,and normalization is performed on the collected data.Finally,SVM,KNN,RF machine learning algorithms and Deep Conv LSTM deep learning method were used to classify the four behaviors of cattle:lying,standing,walking,and eating.A two-layer,two-way GRU network was used to classify the cattle’s vocalizations(normal,estrous)and swallowing sounds.The main work of this article is as follows:(1)In order to monitor the behavior and physical signs of cattle,a cattle collar based on multi-sensor fusion was designed.Firstly,device selection and overall system design are conducted based on functional requirements.Then,the hardware circuits of each module of the system are designed,including data acquisition modules and a data storage module.Based on the hardware design,the software functions of each module are written,including communication driver design,acquisition program design of posture data,body surface temperature,and audio data,as well as positioning program design based on Bluetooth Received Signal Strength Indicator.Finally,from the perspective of comfort and simplicity of wearing a cattle collar,auxiliary wear design is carried out to achieve a complete cow collar development.(2)In order to verify the overall function realization of the collar and the accuracy of data collection,the collar is tested and data preprocessing is performed.Firstly,the least square method is used to fit the path loss function in a cattle farm environment,and then a positioning test is conducted,with an average error of about 1m;Using binary curve fitting and 3-dimensional polynomial surface fitting method considering time,the correction of body surface temperature and rectal temperature of cattle was achieved,~2=0.9854;At the same time,the obtained posture information and audio information are analyzed.Finally,relevant preprocessing such as enhancement,denoising,and standardization were performed on the collected data,resulting in a total of 260084 sequence data and 258 audio data,providing a reliable data basis for recognition and classification of cattle behaviors。(3)Using the data collected from the cattle collar,recognition and classification of cattle behavior are realized.Firstly,the manual feature extraction method is used to extract features from cattle behavior data,including three aspects of statistics,spectral statistics,and information entropy.Filtering and dimensionality reduction of redundant features using variance filtering,chi-square filtering,and PCA;Finally,three machine learning methods,SVM,KNN,and RF,were used to classify the four daily behaviors of cattle:walking,standing,lying,and eating.RF had the highest accuracy rate,reaching 99.59%.Then,using the Deep Conv LSTM deep learning model,we compared and analyzed the classification performance of different data sets(whether location information is integrated or not)for cattle’s daily behavior.The results showed that the integration of acceleration sensors and location information can more accurately monitor cattle’s behavioral activities.Finally,the recognition of cattle voice based on a two-layer,two-way GRU network is introduced.The F1score of swallowing voice is the highest,reaching 93.87%;Under normal circumstances,the lowest F1 score for a cow’s call is 87.66%.