| Inner Mongolia and its surrounding west area are the main districts of sheep production. However, the traditional grazing manner caused the problems of grassland degradation, waste of forage resources as well as low breeding efficiency, so the intensive and large-scale sheep farming has been appeared and developed in recent years, becoming the present tendency in sheep industry. Due to the high density of sheep in farming house and limit range of sheep’s activity, it will be very easy to lead the stress behaviors and health issue if breeder doesn’t provides a good living environment. Therefore, it is essentially necessary to monitor the behaviors and evaluate the level of welfare for facility-breeding sheep.Animal sound is an important means in communication. Considering that sheep will make different sound in different emergent situations, the sound can not only reflect the status of sheep’s organism or health but also can reflect the response behavior to the environmental sudden change. In recent years, some studies have shown that livestock animals’sound, which has the advantages of non-contact, no stress and real-time in situ, can be used to assess the facility-breeding well-being index. In this dissertation, an sheep sound signal acquisition system was established based on wireless network by which the 5 types of sheep sound (fighting, hunger, cough, biting and search companions) were recorded, and the algorithms of wavelet threshold de-nosing, feature parameters extraction as well as classification models were researched.. The main conclusions are as follows:(1) The wavelet threshold de-nosing method was selected to deal with the noise data from running fans, feeding equipment and other sources. In order to enhance robustness of the de-noising algorithm to suit for the sheep house, the wavelet threshold rule and function were improved. The test results show that the improved de-nosing algorithm has the better performance in processing the sheep sound signals contaminated by fans noise. In addition, the sound analysis and processing program for sheep house was developed in software LABView.(2) The traditional linear prediction coefficient algorithm has the problems of combined peaks and false peaks when used to extract formant of sound signal in sheep house. Hence, the linear prediction coefficient algorithm was improved by assuming sheep vocal tract as a series of several cavities, and 12-dimensional feature parameters were obtained by analyzing the fornmant tracking curves of 5 types of acoustic signals in sheep house.(3) The hybrid parameter, the Mel Frequency Cepstrum Coefficient combined with its first difference, was improved based on Fisher criterion of feature components correlation and feature components contribution to recognition. By using the new hybrid feature parameters a higher recognition rate was achieved, even under low dimension.(4) Through Empirical Mode Decomposition of Hilbert-Huang Transform, the sheep sound signals were decomposed into several Intrinsic Mode Functions, and feature parameters, Sub-band Energy Cepstrum Coefficient of sound signals in sheep house were then calculated by the marginal spectrum energy cepstrum coefficient combined with Mel-scale frequency districts.(5) The hybrid sound signal recognition and classification model was constructed by the Hidden Markov Model with the modeling capability of dynamic time sequence signals, and the Back Propagation neural network model with the strong classification and decision capability. By taking the input of Back Propagation neural network model as the output state cumulative probability of Hidden Markov Model, a higher recognition rate than the single model was obtained. |
PDF Full Text Request