Research On Cow Behavior Discrimination Method Based On Multi-feature Fusion

In order to improve our country’s dairy farming technology and reduce farming costs,the labor force can be further liberated.Existing studies have found that dairy cows’ estrus and disease conditions are related to the cow’s exercise volume.The cow’s exercise volume during the estrus period is twice that of the normal cow’s exercise volume.According to this theory,researchers use three-axis acceleration sensors to identify and detect step counting and posture of cows,but the recognition accuracy needs to be further improved.Therefore,this thesis proposes a method to discriminate cow behavior based on multi-feature fusion.This method firstly improves on the hardware sensor.In order to collect more cow movement data,this thesis made a set of nine-axis motion sensor nodes.In order to improve the accuracy of cow behavior recognition,this thesis first preprocesses the collected data,and then designs a multi-label chain GBDT+LR(Gradient Boosting Decision Tree+Logistics Regression)model to learn and predict the processed data.The model designed in this thesis can realize the six posture recognition of standing,resting,feeding,walking,crawling and running.The main work of the thesis is as follows:1.Design and manufacture a set of cow data collection device.The device mainly includes nine-axis motion sensor(three-axis acceleration,three-axis angular velocity,three-axis angle),ESP8266 WIFI module,SD card module,STM32F103RCT6 controller module.The device is used to collect and store cow nine-axis motion data and real-time time.In addition,an Android-based time synchronization recorder software was designed for video data collection of dairy cows.2.Using the time axis as the reference,the collected cow’s motion video data is calibrated for the first time,and the posture label is added.Then the 9-dimensional data is derived to 57 dimensions,and the data is balanced and sampled for imbalanced data.In order to improve the accuracy of the model,this thesis does not directly classify the six poses,but adds a label(still,small movements,large movements)to the second calibration and normalization of the cow’s behavioral poses.Finally,the recursive feature elimination method is used to select the best 9-dimensional feature from the57-dimensional feature.3.Using the data after the above processing,this thesis designs a multi-label chain GBDT+LR model for learning and prediction,and makes a comparative experiment through logistic regression,decision tree and gradient improvement decision tree.The experimental results show that the multi-label chain GBDT+LR model designed in this thesis has an average recognition rate of 93.9% for the six behavioral postures of dairy cows,which is higher than the comparative experimental algorithm.In addition,this thesis finally verifies that the cow’s estrus status is related to the cow’s behavior.