Abstract: In the process of pig breeding, it was difficult to extract and recognize audio features of mixed pigs. A blind source signal separation method based on sparse theory was proposed for underdetermined pigs. The 4 months old and 150 kg sows in good health were selected, and the audio signals obtained by mixing their cries in different states according to different coefficients were used as observation signals. Short-time Fourier transform was used to convert the audio signals. Single source points in the signals were selected by grouping, and AP algorithm of adaptive damping coefficient combined with singular value decomposition was used. The single source points were clustered to estimate the mixing matrix, and the audio signal was reconstructed by optimizing the minimum l_p norm. Two groups of experiments were designed. One group explained the general process of the experiment, and the other group analyzed the performance of the whole separation algorithm by comparison. Similarity coefficient, signal-to-noise ratio and mean square error were used to measure the separation audio quality. The results showed that:1) In the underdetermined pig blind source separation of 3 source signals and 2 observation signals, the similarity coefficient, signal-to-noise ratio and mean square error of the separated audio signals and the corresponding source signals at different time ranged from 0.67-0.92, 7.9-9.7 dB and 0.005-0.08, respectively. From the view of waveform, the separation performance of the algorithm was independent of the length of time and the number of tests, and the results had a certain stability. 2) When the number of source signals and observed signals were 3 and 2, 4 and 2, 4 and 3, 5 and 2, 5 and 3, 5 and 4, the average similarity coefficient, signal-to-noise ratio and mean square error of reconstructed signals and source signals were 0.785-0.957, 7.468-10.347 dB and 0.019-0.092, respectively. After comparative analysis, this research method had certain reliability. 3) When the number of source signals was certain, the more the number of observed signals, the better the index measured, and the better the separated audio quality. In conclusion, this method can effectively separate each source signal component of mixed pig audio signal, which layed a foundation for feature extraction of mixed pig audio in real environment.

Key words: pig audio, signal sparsification, AP clustering, l_p norm, blind source separation