Abstract: The behavior state of sheep can directly reflect its health status and physiological stage. In order to realize automatic behavior recognition of sheep, the paper constructed a wearable behavior monitoring device based on acceleration sensor, taking captive small tail Han sheep as the experimental object. The nine-axis posture sensor with MPU9250 as the core was used to collect the behavior information of the sheep when they were still, walking and eating, and the sensors were deployed in the neck, back near the front leg, front leg and back leg respectively. For the collected data, noise reduction and dimension reduction were used for preprocessing, and k-means and SVM were used for classification and recognition respectively. The k-means clustering algorithm has the highest accuracy rate of 79.34% for cervical behavior recognition, and the SVM support vector machine algorithm had the highest accuracy rate of 92.63% for cervical behavior recognition. The experimental results showed that the overall accuracy of SVM method for sheep behavior recognition was higher than that of k-means, and the recognition efficiency of sensors located in sheep neck was better than other parts under different recognition methods. The results of this study have important practical significance for the construction of automatic sheep behavior detection system.