Feature extraction and classification of laying hens' vocalization and mechanical noise

Abstract: Vocalizations of farm animals may accompany particular states of animals’ mood or emotion. Based on these vocalizations, we can judge animals’ current needs and impaired welfare, so they may be regarded as indicators of animals’ state of welfare. However, the noise made by different mechanical systems in the commercial poultry house can interfere with the detection of laying hens’ vocalization. The purpose of this study is to analyze and classify vocalizations of laying hens and mechanical noises. The analysis and classification is based on time-domain and frequency-domain characteristics of the signal. Vocalization in the egg laying process and song are two kinds of typical laying hens’ vocalizations. Mechanical sources of noise on the farm mainly include the ventilation system, manure-removal systems, egg-collection systems, and feeding systems. The power spectral density and sub-band power ratio of laying hens’ vocalizations and mechanical noises were extracted by using a sound analysis system based on the program development environment LabVIEW. A J48 decision tree algorithm was used to classify and identify laying hens’ vocalization and mechanical noise on the data-mining platform Weka. The results showed that the frequency ranges of vocalization associated with the egg-laying process and singing were mainly distributed within 400-2 500 Hz, the frequency ranges of ventilation-system noise and feeding system noise were mainly distributed below 1 500 Hz, the frequency ranges of manure-removal system noise and egg-collection system noise were located within 100-3 000 Hz, which was wider than the frequency ranges of other sounds. The max power ratios of vocalization in the egg-laying process and singing were (83.4±9.9)% and (76.7±18.8)%, which were within the frequency range >689-1 378 Hz;. The power ratios of vocalization in the egg laying process and singing were higher than that of mechanical noises in the frequency range >689-1 378 Hz. The maximum power ratios of ventilation-system and feeding-system noise were 68.1±2.1% and 74.5±9.7%, respectively, which were within the frequency range 0-689 Hz. The power ratios of ventilation-system and feeding-system noise were higher than that of others in the frequency range 0-689 Hz. The power ratio of manure-removal system and egg-collection system noise were relatively uniform; the maximum power ratios were just 37.2±4.1% and 40.9±3.4%, respectively, and were within the frequency range >1 378-2 756 Hz. The power ratios of manure-removal system and egg-collection system noise were higher than that of others in the frequency range >1 378-2 756 Hz. The sound recognition model based on sub-band power ratio had an average correct classification rate of 93.4%. Further, 92.5% of vocalizations associated with the egg laying process and 85.9% of songs were correctly identified, and the correct classification ratios of ventilation system, feeding system, manure-removal system, and egg-collection system noise were 97.7%, 96.2%, 97.8%, and 94.4%, which were higher than that of laying hens’ vocalizations. This method, based on sub-band power ratios, effectively recognizes and discriminates noise from different sources, which can provide a reference for detecting vocalizations of animals within the complex noise environment on the commercial farm.