Scientific methodology for sound labelling in relation to precision livestock farming

The analysis of animal sounds and vocalizations allows the interpretation of stress, behaviour or disease pattern in a non invasive way, and it might be used to assess welfare, management as well as health status and social adaptation (e.g. hierarchies ). Farm sanitary level especially is crucial for modern livestock farming where high animal density affect health. Monitoring animal health is vital for the sustainable management of a farm but it’s often difficult to apply punctually. The results of a low monitoring are increased disease rate, increased pharmacologic costs and decrease in quality and quantity of productions. In Italians intensive farming realities the main issue is given by respiratory pathologies , the second in incidence and prevalence after enteric diseases. Coughing is one of the most frequent presenting symptoms of many of them and it is a sudden air explosion in the airways followed by a characteristic sound. Being cough one of the body's defence mechanisms against respiratory infections, it can be a sign of disorder or infection of the respiratory system. It has been identified as an index for over 100 diseases and experienced physicians, in human medicine, can identify an infection based on the cough sound. But the importance of coughing as a means of prognosis refers also to animals: it has been shown that pig vocalisation is directly related to pain and classification of such sounds has been attempted; it is also common practice by veterinarians to assess cough sounds in pig houses for diagnostic purposes. However, this last approach cannot be used as a continuous monitoring and early warning for infections in pig houses since it is time limited and subjective. In this regard, there have been attempts to identify the characteristics of coughing in animals. More objective and automated detection of respiratory diseases in pig houses should be possible by on line sound analysis of cough monitoring. To develop automatic algorithms for pig cough recognition, experiments and well-labelled cough data are needed. The arguments exposed in this thesis are: (1) the attained results in cough recognition and classifications, (2) definition of methodology to label cough data and sound data in general in field condition, (3) the achieved results for sound source localization by sound analysis, (4) the steps toward the creation of an algorithm for automatic sounds recognition and localization, (5) the creation of a “labelling tool”. The first issue suppose that changes in cough character may have a considerable value in identifying the mechanisms of airway pathology present in respiratory diseases. The cough sounds gives information about the patophysiological mechanisms of coughing by indicating the structural nature of the tissues during pathology that leads to certain patterns of cough. Signal processing allowed to identify the peculiar features, in time and frequency domain, of different sound groups. Comparisons were based on the Root Mean Square (RMS) normalized pressure, the peak frequency (frequency with maximal energy content) and the duration of the signals (in s). The second argument developed is sound labelling which is fundamental for the steps of sound analysis, modelling and for the creation of the algorithm. Labelling is a manual procedure, based on acoustic analysis combined with visual spectral analysis, which is used to extract cough sounds from the entire recordings. In this thesis labelling is done manually, offline by the operator to extrapolate only those sounds that the visual observation of the spectrogram and the auditive confirmation classified as cough attacks. In this way we have obtained a database of various coughs sounds. For the third issue on sound localization an algorithm based on the “Time Difference of Arrival” (TDOA) of the signal in 7 microphones is proposed. It is shown that this computationally efficient method provides acceptable accuracy levels for the specific application. In the forth point of this discussion the time domain characteristics of sound signals have been investigated to evaluate their value to automatically classify sick coughs. The instantaneous energy of the signal is used to detect and extract individual sounds from a continuous recording and their duration is used as a pre–classifier. The sick cough identification algorithm is based on Auto Regression (AR) analysis of the signal and has been shown that the AR parameters of sick pig cough signals form a separate cluster from that of other sounds. The proposed localization algorithm is based on the time difference of arrival (TDOA) between the sound as received by 7 microphones. This technique is the first application for combined online cough recognition and localization presented in the relevant literature and can pose as a starting point for further research. The last subject of this thesis is the result of the effort to study sound characteristics and particularly cough sound features and it is concretised in a pc program called Labelling tool capable of automatic sound extraction from continuous recordings. Subsequently to the results obtained the bioacoustic approach may be directly applied to intensive farming systems where its efficiency will be tested along with the effect it can have on animal welfare, faster animal treatment, reduction of antibiotics and therefore pig health. By applying the system in the stables farmers will be provided of automatic cough counting and coughs positions visualization which will automatically monitor the spread of respiratory diseases and eventually contribute to early diagnosis of the disease bringing to antibiotics use reduction by means of selective, early treatment.