Research On Individual Identification Method Of Dairy Cows Based On Activity Data And Deep Learning

Accurate and effective individual identification of dairy cows is the basis to realize the scale,informationization and refinement of animal husbandry.The traditional cow individual identification method needs manual identification,which has high labor intensity and low recognition efficiency.The image acquisition condition of non-contact cow individual recognition technology is complex,and only the static external features of dairy cows are used for recognition.In order to solve the above problems,a cow individual identification method based on cow activity data and deep learning is proposed in this dissertation.In this method,the dynamic activity is collected,the dynamic activity is extracted by Bidirectional Long Shortterm Memory Network(Bi LSTM)model,and the Self-Attention mechanism is introduced to distribute the weight of dairy cow individual temporal correlation.The purpose of this dissertation is to realize the accurate and rapid recognition of multiple cows at a long distance.In this dissertation,firstly,the author analyzes the advantages and disadvantages of common dairy cow individual identification techniques,then the dynamic attributes of dairy cow activity are mined,and the activity data of different dairy cows in different states are analyzed from the perspectives of time domain and frequency domain.The aim is to find the individual differences of dairy cows based on activity data.Secondly,the wearing mode and working principle of the activity collector are introduced,and the dynamic activity data form and pretreatment process are expounded.Thirdly,according to the characteristics of dynamic activity changes over time,the cow individual identification algorithm is determined,and the relevant principles of the algorithm,network framework,evaluation indicators and parameter testing are introduced.Finally,it is tested in different aspects: the effect of recognition method,the effect of recognition model,the comparison of data granularity,the comparison of sample number and the comparison of sample duration.The test results show that in the effect test of the recognition method,the activity data and pretreatment method used in this paper are better,and can show high degree of distinction in the individual identification of dairy cows;In the identification model effect test,the cow individual identification model based on Bi LSTM＿Attention in this paper is compared with the LSTM＿Attention model and RNN model under the same test conditions,the model used in this paper has the highest accuracy,the lowest consumption of time and the best performance;In the data granularity test,when the granularity of the test data set containing ten cattle for240 hours is divided into 60 minutes,the model recognition accuracy is 91.8% and the training time is 98.7 seconds,which can improve the recognition efficiency;In the sample number test,with the increase of the number of samples,the recognition accuracy will gradually decrease;In the sample duration test,recognition accuracy increases with the increase of sample duration.According to the above test results,the cow individual recognition algorithm based on dynamic activity data and Bidirectional Long Short-term Memory Network under Self-Attention mechanism can quickly achieve accurate and effective identification of dairy cow individuals.It provides new ideas and methods for the research direction of cow individual identification based on dynamic activity and the Live Mortgage Field,and contributes a reasonable scheme to promote the development of large-scale,digital and intelligent pasture farming and the implementation of Rural Revitalization Strategy.