At present, most breeding farms rely on artificial observation to identify sows' estrus behavior. Many breeders have neglected the sow's estrus performance and it is difficult to master the sow's scientific scientific timing of breeding, which leads to poor breeding time or conception. The reduction of the rate greatly affects the normal reproduction rate of sows. In order to solve the above problems, a parallel KNN algorithm for likelihood estimation is designed in combination with the sow estrus detection system. The algorithm first uses the maximum likelihood estimation method to process the collected body temperature, activity and feed intake data, and solves the problem of incomplete data due to problems such as mechanical failure. The processed data is then used to perform data anomaly detection using a parallel KNN algorithm, which improves the efficiency of data operation detection. Under the same experimental simulation environment, comparing this algorithm with the traditional KNN algorithm takes 31% time, and the overall accuracy rate is increased by 3.37%.