Audio Analysis: A Path to Monitoring the Well-Being of Birds in Confined Housing
Walk into any broiler (chickens raised for their meat) growout house and you will hear an interesting chorus. An array of chirps, clucks, and squawks abound. Researchers at the Georgia Tech Research Institute (GTRI) are discovering that the unique sounds actually tell a lot about the birds’ well-being. Using their machine learning-based Growout Monitoring System, the team has characterized when birds are under stress due to sickness or adverse environmental conditions inside the house.
“It is well-known that environmental conditions during broiler growout can affect the performance of the birds,” explains Dr. Wayne Daley, GTRI principal research engineer and project director. “The goal of the research has been to monitor various audio characteristics of the birds to determine the flock’s health status and well-being based on environmental conditions.”
Early detection of adverse conditions or sickness can mean quicker intervention, saving dollars and birds, notes Daley. In other words, vocalizations can equal warning signs, and more importantly, understanding what they mean could help to improve production efficiency and bird well-being.
The system includes a computer data collection station and interconnected recording microphones. It uses machine learning algorithms to analyze the audio data (bird sounds, also known as vocalizations) and then trains the system to characterize those vocalizations. The techniques use Mel Frequency Cepstral Coefficients (MFCCs are typically used in speech analysis) along with other statistical and spectral features to extract and track significant occurrences from the audio stream. One-minute segments from this stream are then processed in a Vocalization Processing Testbed (VPT) that houses the algorithms that process the data.
Working with colleagues at the University of Georgia (UGA), Daley’s team has recorded thousands of hours of audio from birds grown in experimental conditions.
In several studies conducted at UGA’s Poultry Research Farm, data was collected under normal and stressed (temperature increased 10 degrees above normal) growing conditions. Analyses of the data showed that it is possible to detect a change in the vocalizations of the birds due to a change in temperature. Similarly, two separate studies at UGA’s Poultry Diagnostic and Research Center investigated the correlations of vocalizations to two common broiler diseases: Infectious Bronchitis and Laryngotracheitis (LT). In both experiments, algorithms identified vocalizations that correlate with the progression of the diseases.
Recently, the team performed a study on breeders (chickens that are the parent stock for broiler chickens) being raised under different feeding regimens (i.e., birds fed every day vs. birds fed every other day) to determine if differences in vocalizations would be apparent between the feeding regimens. Preliminary data indicates that the rates of the vocalizations appear to be different on the off days.
“Our research to date shows that it is possible to determine the conditions of birds based on their vocalizations,” says Daley. “The question now being explored is how best to implement a system that uses this information to support farm managers in optimizing their production with animal well-being as one of the significant drivers.”
To address that question, researchers are refining the VPT and developing tools to make the system easier to use for practical field operation.
“In our current mode, we would have to detect every type of occurrence. This could be extremely difficult as you move from house to house and environment to environment. The approach now being explored is to detect anomalies,” explains Daley.
Simply put, the system learns to describe the acceptable environment and then looks for significant deviations that would signal a change that should be addressed. Daley says these could be described as novel or anomalous events. The team hopes these anomalous occurrences could be used to create a descriptor of well-being that could constantly be provided to the growout house manager for action when the need arises.
Researchers are also investigating the use of low-cost computers like the Raspberry Pi to acquire and analyze audio data, and have developed a smart phone app on the Android platform. The app is called SCAR (Sick Chicken Audio Recorder) and can be used to easily record audio samples in minutes, hours, or days.
“It is now possible to visualize an ecosystem that allows for overall confined housing management where data is collected and processed to determine the overall conditions of the birds,” says Daley.
Some of this data processing could be done remotely, reducing the power and computational capability needed locally while also ensuring that the data is being processed by the latest algorithms and classifiers.
“Successful implementation would provide the farm manager and integrator with significant control over the production process including almost 24/7 monitoring of the environment. This should enable a quick response to environmental or health issues based on a measure of the actual condition of the birds,” adds Daley.