Researchers with the Georgia Tech Research Institute (GTRI) continue to explore techniques to enhance the functionality and utility of their Growout Monitoring System. The automated system records and analyzes bird vocalizations to assess their condition during the growout cycle. The goal is to detect potential problems early in the growing cycle, enabling a quick response that not only ensures the well-being of the birds but saves the growout house manager time and money.
“Through some of our early work with partners at the University of Georgia, we learned that it was possible to correlate the birds’ vocalizations with the onset of respiratory diseases and temperature-related and other stressors due to changing environmental conditions within the house,” says Dr. Wayne Daley, GTRI principal research engineer and project director.
Since then research has focused on determining how to advance the system’s functionality in a way that best helps growout managers monitor their houses. Here, Daley says the team is focused on determining measures that might be able to quantitatively describe well-being in the production process. That begins with addressing the complications of operating the system in a full production setting.
To this end, the team recently began experiments in a commercial-size growout house at the University of Arkansas Center of Excellence for Poultry Science. According to Sim Harbert, GTRI senior research engineer, researchers have refined the Growout Monitoring System to be a low-cost extensible networked system using the Ubuntu Linux operating system and commonly available hardware and high-quality microphones. Specifically, it includes four cameras and 10 microphones along with the data acquisition component that is based on Raspberry Pi computers and powered over Ethernet. It is also networked with redundant data storage and provides remote access.
“One key change was that instead of attempting to detect every type of adverse condition, we decided it was more advantageous to employ machine learning techniques where the system is trained to describe acceptable conditions and then learns to detect when something abnormal [called an anomaly] happens,” says Dr. David Anderson, an expert in digital signal processing and professor in the Georgia Tech School of Electrical and Computer Engineering (ECE). Anderson is spearheading the team’s algorithm development work.
One of the difficult tasks in analyzing audio is determining characteristics or descriptors (called features) that can represent or be surrogates for house conditions. Through decades of research, useful features have been developed for common signals such as those used in speech recognition applications, explains Anderson. Much of that work, however, simply does not apply to the chicken growout house environment.
Therefore, the team has employed new techniques for generating useful features based on sparse coding — a method of learning audio descriptors that uses only a minimal number of features at a time. The sparse coding approach tends to learn a dictionary of fundamental types of sounds that are commonly encountered in the environment, for example, bird chirps or cooling fans. The recognition algorithm then tries to match recorded sounds to those in the dictionary. If a particular sound does not match well with those in the dictionary, it can be considered an anomaly. The idea is that recognition of these anomalous sounds could alert the growout house manager that an adverse condition may be occurring that needs to be addressed.
The team is testing this approach by creating conditions in the house that would be of interest to managers to see if the proposed techniques would detect these changes. Examples include hot and cold temperature deviation and water and feed restriction. Using the sparse coding technique, the system characterizes the daily cycle or rhythm of the house. Researchers can then analyze the resultant data to determine day to day during the same hour period how the birds’ vocalizations are changing. By characterizing the cycle that occurs day to day, the team can determine what is a deviation that could potentially affect well-being, explains Anderson.
Early testing results seem to indicate that high temperatures and water restriction seem to have more of an effect than feed or temperature reduction within the limits of the experiments conducted thus far.
“During audio data analysis, we are seeing changes in the bands of frequencies. The chicken voice is very low intensity, and the background sounds are sometimes very high intensity. So, we are exploring signal enhancements and noise removal techniques to work on this challenge,” says Muhammad Rizwan, GTRI graduate research assistant.
Anderson says with continued testing researchers anticipate improving on the algorithms as they learn more about the real world challenges of operating in a full production setting.
“These algorithms could allow us to extract quantitative descriptors of what types of things occurred throughout the lifetime of each flock, which could then be correlated with other metrics of performance and well-being. Descriptors that correlate highly could then be traced back to the specific times and events that activated that descriptor, helping to pinpoint the underlying causes that led to a given outcome,” says Brandon Carroll, a graduate student in ECE who is also working on algorithm development.
“Ultimately, we would like to see integrated control systems for managing animal environments that are directly responsive to the needs of the animals,” adds Daley.
The team’s work thus far, he says, would not have been possible without the collaboration of colleagues at other universities, in particular, the University of Georgia and the University of Arkansas.
“They continue to inform and educate us on the complexities of managing biological systems; it has been challenging but fun,” says Daley.