beFarm introduces a neural network based vocalization detector for broiler welfare monitoring
One of the main challenges in the poultry supply chain is the need to transform into highly efficient production systems to ensure increasingly sustainable food production, where animal welfare is fundamental and will increasingly become a differentiating factor in the market.
In this context of Precision Livestock Farming (PLF), it is necessary to establish baselines capable of standardizing processes and guiding farmers and technicians in daily operations, along with tools for data collection and (automated) interpretation of animal health and welfare. It is crucial to promptly identify any stress conditions in the animals, which can depend on external factors that are difficult to foresee; for example, particularly rich feed (high in calories and protein) can cause heat stress in animals, which will temporarily need cooler temperatures than the standard.
Current methods for evaluating animal behavior, welfare, and detecting diseases are based on the collection of subjective, labor-intensive data by costly experts, conducted over a short period (several hours), which is not always representative. This approach leads to delayed interventions and incomplete or incorrect data interpretation, and it is not economically advantageous.
For these reasons, we are working, with excellent feedback, on the implementation of an acoustic detector for broiler vocalizations as part of a broader set of behavior and welfare monitoring tools developed within the beFarm platform. The vocalization detector will be an integrated feature of the smart sensor beFarm Tiresia. It will be based on a convolutional neural network; for training, a labeled library of vocalizations will be created, based on a wide set of broiler audio recordings covering the entire production cycle.
Broilers have the ability to communicate through vocal signals, making sound analysis a useful tool for monitoring their behavior and biological responses to external stimuli. Several university studies have identified four different types of vocalizations: pleasure notes, distress calls, short peeps, and warbles. Distress calls are characterized by repetitive, high-energy vocalizations, while short peeps are identified as low-energy, short-duration vocalizations with decreasing energy. Pleasure notes, on the other hand, are described as low-energy vocal expressions that tend to oscillate upward in pitch, with an ascending frequency and short duration. Warble notes can be ascending or descending in frequency and are characterized by repetitive, arc-like vocalizations with low energy. Studies have also shown that the average peak frequency of vocalization is inversely proportional to the age and weight of the broiler. The average peak frequency on the first day of life is 3.6 kHz, dropping to 1.5 kHz after 36 days.
The vocalization detector will be based on a convolutional neural network; for training, a labeled library of vocalizations will be created, based on a wide set of broiler audio recordings covering the entire production cycle.
In this first phase we installed a professional microphone inside a broiler breeding barn, connected to a high resolution audio recorder, capable of continuously recording all the sounds detected during the entire production cycle.
领英推荐
Therefore we record the vocalizations of the animals but also the surrounding noises caused by the fans, the hot air generators, the augers that introduce the feed, and human activities; noises that will have to be isolated later with powerful software filters.
We also installed a Dome video camera capable of recording a video stream of the area in which we installed the microphone, for the entire breeding cycle. The images will serve as a visual confirmation of the possible stress conditions that we will cause inside the barn, and which we will classify with the collection and analysis of the vocalizations.
At a later stage, we will work on forming a vocalization database for training and validation purposes. This database needs to be as clean as possible, containing sound recordings of each type of vocalization for different ages of broilers. The database development is carried out in three phases:
The actual construction of the broiler vocalization detector, trained on this vocalization database, will be carried out in a second phase, in which a customized neural network will be designed and trained (supervised) to perform automatic classification of broiler sounds, with the ability to generate alerts in the presence of animal stress conditions.
Stay tuned!!