Abstract
BACKGROUND: Centralized machine learning (ML) models often face challenges related to data privacy, messaging latency and limited internet access in rural areas. To address these issues, a federated learning (FL) based architecture for real-time detection of health and stress in cattle using sensors is deployed. This model enables edge devices such as smart collars, wearable sensors and cameras, to collaboratively learn a global model without sharing raw data. OBJECTIVES: To create a FL-enabled architecture of real-time health and stress detection in cattle with distributed smart farming devices, and to deal with the issue of data privacy, latency and poor internet in rural conditions. METHODOLOGY: In the proposed system, FL is used to allow edge devices, including smart collars, wearable sensors and cameras to jointly train a global model without exchange of raw data. There is the utilization of multimodal time-series data, such as temperature, heart rate, motion trajectories and environmental conditions. The architecture combines LSTM and CNNs to find anomaly and behaviour patterns. Measurement of performance is compared with centralized ML models with real or simulated livestock data. RESULTS: The results of an experiment prove that the FL-based model has better performance than centralized and baseline federated models and this has an accuracy of 93.1%. The model demonstrates better convergence properties and saves a lot in the transmission of the raw data. It is applicable in cattle, early stress detection, illness and abnormal behaviour. CONCLUSION: The possibility of the use of federated AI systems provides an accurate solution to secure, privacy-preserving and efficient livestock monitoring. The suggested solution can contribute to the improved economic performance and animal welfare in the future since it promotes sustainable smart agriculture and can be integrated with systems like managing irrigation in the future.