Feature Extraction and Classification Based on Skewness Clustering Algorithm for Lactating Sow

Abstract:

The lactation period is a critical period for sows to breed their piglets, and the specific voice of lactating sows in this period is the most direct expression of their physiology, emotional health, and maternal ability to care for piglets. The rapid location and accurate identification will be more complex due to a variety of vocalizations during this period. Therefore, the vocalizations of nursing grunt, drinking, feeding and sham chewing were observed, and a fine energy calculation for frequency domain with a power ratio as a vector was carried out. Then, the subband clustering method based on skewness was presented to merge the sub bands without significant characteristics to reduce the number of parameters. Thirdly, the recognizer for sow’s vocalizations was built based on support vector machine(SVM) to calculate the duration of the different types of vocalization. A sound mode of successful nursing was established further within single lactation circle. It is shown that the max power ratio frequency domain of the nursing grunt, the sham chawing, the feeding and the drinking are ranged from 0Hz to 1000Hz, 1000Hz to 1500Hz, 1500Hz to 2500Hz, and 2500Hz to 8000Hz, respectively. The accuracy of the vocalization recognition mode with these four sub bands power ratio frequency as parameters were 100%, 100%, 95.17% and 96.61%, respectively. Compared with the uniformlyspaced subband division and principal component analysis (PCA), the number of features was reduced, and the recognition accuracy was significantly improved in the clustering algorithm based on skewness. Thus, the proposed method could be further applied in the health and maternal ability of sows monitoring realtimely and nonstressly.