Due to bad work environment and network transmission failure, it is easy to generate abnormal sensory data in livestock breeding Internet of things system. In order to ensure the quality of sensory data, according to the characteristics of sensory data flow such as periodicity, temporality, infinity, etc., a method was proposed based on sliding window and support vector machines regression (SW-SVR) for livestock breeding Internet of things abnormal sensory data detection in real time. Firstly, the sliding window size was decided according to the characteristic period and sampling frequency of data flow from livestock breeding Internet of things system, and the history data within sliding window was selected as the input value of prediction model. Then, the sensor estimated measurement value at certain moment in livestock breeding Internet of things system was predicted by using SVR model. Finally, the prediction interval (PI) was calculated, and the abnormal sensory data was identified if the sensor actual measurement data fell out of the PI. The abnormal data would be replaced by the predictive data. The abnormal sensory data detection method was tested by data flow from real livestock breeding Internet of things system. Experiment results showed that the mean absolute percent error value of prediction with window size calculated by the sliding window method was 0.1884. The correct detection rate of abnormal data based on SVR model with radial basis function kernel (RBF kernel) achieved 98%, which had higher accuracy compared with BP neural network (BPNN) method. Abnormal data can be effectively detected and treated in livestock breeding Internet of things system.