Poultry production which plays a key role in ensuring food security, faces challenges related to
increasing its productivity and improving animal welfare. The chicken hatching process which
still requires significant manual labour is technologically outdated, leading to higher mortality
rates. This master's thesis presents the development of a sensor prototype that utilises machine
learning on an edge device to detect chicken hatching in an industrial incubator.
We used sound as the primary source of information, enabling the machine learning model to
detect the external pipping and hatching phase and to track the hatching percentage over the
last 72 hours of the process. Audio signals were collected in both laboratory and industrial
incubators, and the data was labelled before being added to the dataset. Subsequently, we
developed a preprocessor that removes noise from the signals, generates a Mel spectrogram,
and extracts relevant features.
In the next step we addressed sound classification. We developed a convolutional neural
network (CNN) model using Imagimob Studio which classifies input data into four target
categories: external pipping, hatched phase, 1% hatched, and 80% hatched.
We used the Infineon Evaluation Kit equipped with two microphones to implement our
algorithm along with the optimised convolutional neural network model. To protect the circuit,
we designed and 3D printed the enclosure and mounted it inside the incubator. Additionally,
we developed an application with a user interface that provides real-time information on the
hatching process and enables easy monitoring.
Evaluation results in a laboratory environment confirmed the prototype’s success in achieving
the set objectives, demonstrating its potential for industrial implementation in future
developments.
|
