Bioinspired small language models in edge systems for bee colony monitoring and control

Abstract

This paper proposes a hybrid IoT architecture based on Generative Artificial Intelligence (Gen-AIoT) for the intelligent monitoring and control of beehives, designed with processing capabilities both at the edge and in the cloud, thus adapting to environments with or without internet connectivity. Through an IoT sensor network, the system collects critical data on environmental parameters and hive conditions, such as temperature, humidity, wind speed, and hive weight, processing them locally at the edge or centrally in the cloud. The architecture incorporates a recommendation system that uses a small language model (SLM) to generate real-time alerts based on data provided by the IoT sensors. This system implements two distinct SLM models, Phi-3.5 and Tinyllama, enabling hardware performance measurement and optimizing efficiency for edge processing. To establish optimal environmental ranges, the recommendation system uses bio-inspired algorithms, such as ant colony optimization, genetic algorithms, and bee swarm algorithms. Additionally, LSTM neural networks are included to predict honey production and plan hive placement based on climate and weight projections, allowing for precise and personalized adjustments. This dual processing capability (edge and cloud) reduces the need for human intervention, optimizes hive inspection times, and minimizes false positives in monitoring, making it especially beneficial for large-scale beekeeping, where weekly inspection times can exceed 50 h. With this architecture, inspection time is reduced by 80%, significantly improving efficiency in hive management and promoting sustainable practices for bee conservation through intelligent agriculture.