Inner Mongolia is rich in grassland resources, and sheep industry is the main animal husbandry in the autonomous region. The intelligent identification of sheep grazing behaviors combined with GPS monitoring of grazing path can provide theoretical basis for estimating of feed intake distribution in grazing area, grazing planning and grassland livestock balance. Threeaxis acceleration sensors were used to design a wireless data acquisition system for grazing behaviors of herding sheep, and the system can automatically collect the three-axis acceleration data of grazing behaviors. BP neural network model, full connection deep neural network model and convolution neural network model were established to realize the classification and recognition of feeding, chewing and ruminating behaviors of herding sheep respectively. The experiments were carried out in a semi-desertification grassland in Inner Mongolia, and the natural grazing Mongolian sheep were selected as test objects. The results showed that the average recognition rates of BP neural network model, full connection deep neural network model and convolution neural network model were 83.1%, 89.4% and 93.8%, respectively, and the convolution neural network model had the highest recognition accuracy, and the adaptability and stability of the network model was strong, which can meet the requirements of classification and recognition of sheep grazing behavior. The feeding path of sheep can be monitored by GPS. The research result can provide theoretical basis for ranch managers to formulate grazing system and improve grazing level.