Discerning pig screams in production environments

Abstract

Pig vocalisations convey information about their current state of health and welfare. Continuously monitoring these vocalisations can provide useful information for the farmer. For instance, pig screams can indicate stressful situations. When monitoring screams, other sounds can interfere with scream detection. Therefore, identifying screams from other sounds is essential. The objective of this study was to understand which sound features define a scream. Therefore, a method to detect screams based on sound features with physical meaning and explicit rules was developed. To achieve this, 7 hours of labelled data from 24 pigs was used. The developed detection method attained 72% sensitivity, 91% specificity and 83% precision. As a result, the detection method showed that screams contain the following features discerning them from other sounds: a formant structure, adequate power, high frequency content, sufficient variability and duration.

Fig 1. Ground plan of the pig compartment.

Each pen had six animals, one feeder and one drinker. One microphone recorded the sound.

Fig 2. Overview of the classifier elements.

The rectangles represent the four parts. The raw data is transformed into an output that indicates if a scream is present. (CGD = Chirp Group Delay, FFT = Fast Fourier Transform)

Fig 3. The FFT spectrograms and CGD spectrograms.

The upper two figures show the same screams, the lower figures show the same sneeze. The left figures depict the spectrogram made from FFT while the figures on the right side depict the spectrogram made from CGD. The formant structure of a scream is visible (B). These formants are the whiter values in the CGD spectrogram. It is not straightforward to find the same structure in the FFT spectrogram. Because the difference between the formant value and the surrounding values is much bigger relative to the maximum and minimum values of the CGD compared to the FFT

Fig 4. The Mel CGD spectogram.

The left figure (A) shows the CGD spectrogram with Mel frequency resolution. The right figure (B) shows the corresponding mean values of each frequency value from the Mel CGD spectrogram. This figure further shows the straight line fitted trough these mean values.

Fig 5. The resulting classification.

These were the found thresholds for each feature after applying the gini index to all sound events in the traing dataset. Each line resembles on feature with acompagning threshold. When an event’s feature value was above the threshold value, the sound event received one vote. All ten votes were subsequently summed together for one sound event.

Fig 6. The two ROC curves.

They showing the True and False positive Rates (TPR and FPR). The numbers on the plots give the minimal required votes for the training and validation set.