Comparison of different de-noising methods in vocalization environment of laying hens including fan noise

Abstract: Vocalization, which plays an important role in the communication of many animals, could be regarded as an easy way to evaluate the current needs of animals and their state of impaired welfare by non-invasive and continuous monitoring. Modern techniques of sound analysis have provided tools for the quantitative description and statistical analysis of animal vocalization. However, a farmer always regulates the temperature and exchanges fresh air through mechanical ventilation in a poultry house, so that there is often low-frequency fan noise which can complicate the time-frequency features of poultry sound. In order to extract useful sound signals from the laying hens' sounds including fan noise, a digitized sound acquisition platform was built to record the pure sounds of laying hens (Hy-Line Variety Brown), the simple fan noise, and the mixed sounds of them both separately. All data processing and analysis were completed in LabVIEW. First, the time-frequency features of the pure sounds of laying hens, the simple fan noise, and the mixed sounds were analyzed in turn. Then the analog signals of different signal to noise ratios (SNR) were constructed by using the clean sounds of laying hens and fan noise. Finally, the effects of removing fan noise from the analog sound and field sound of laying hens were compared by 2 filters (Infinite Impulse Response, IIR, and Finite Impulse Response, FIR) and 3 wavelet de-noising methods (Stein's Unbiased Risk Estimation, SURE, Universal and Minimax). The results showed that although the characteristic parameters of laying hens' vocalization at different growth stages and in different rearing environments were disparate, the main frequency range of the sound of laying hens in the laying period was about 400-2 500 Hz, while the frequency range of fan noise was under 600 Hz. Root mean square errors (RMSE) of the filters under different SNRs were larger than that of the wavelet de-noising methods. The RMSE of the IIR filter was less than that of the FIR filter, which means that the band pass filter of IIR had a better filtering effect than that of FIR. The RMSE of the wavelet de-noising method based on SURE was at a minimum in three wavelet de-noising methods, which means that the wavelet de-noising method using SURE for threshold selection yielded better de-noising performance than that of using the others for threshold selection.