Abstract: In order to monitor the behavioral status and physical information of cattle, a multi-sensor fusion cattle collar is designed in this paper. The collar selects a control chip based on Bluetooth 5. 0, and three kinds of sensor devices are connected to realize the collection of the four basic sequence data of cattle behavior posture（acceleration and angular velocity values）, body surface temperature, indoor position, and three audio data of cow calls（normal condition and estrus condition） and swallowing sounds. SVM, KNN and RFC are used to classify the walking, standing, eating and lying behaviors of cattle. Among them, the average accuracy of RFC is the highest, reaching 99. 59%, followed by KNN and SVM, and the accuracy rates are respectively 99. 01% and 85. 23%. The GRU-based deep learning algorithm is used to classify cattle calls and swallowing sounds, and the overall accuracy rate reaches 90. 72%. The fitting correction of the collected body surface temperature and rectal temperature shows that the fitting degree R2 is higher than 0. 9. The results show that the cow collar based on multi-sensor fusion can not only effectively collect traditional sequence data, but also collect audio data synchronously. It can provide multi-dimensional data support for cattle behavior analysis.