Using time series and lstm neural networks for identification of hive states

Bees are an essential part of our ecosystem and economy. Given the threats facing bee populations, developing effective methods to monitor and conserve them is critically impor-tant. Traditional methods of monitoring bee colonies, such as visual inspections, hive weight monitoring, and acoustic analysis, have their limitations. Visual inspections are subjective, time-consuming, and can disrupt colony life. The aim of this study is to develop and evaluate methods employing time-series analysis and LSTM neural network architectures for the iden-tification of various beehive states. The research focuses specifically on utilizing time-series data pertaining to hive weight and temperature, supplemented by temporal information (hour and month), to train models capable of distinguishing these different conditions. Data for training includes relative hive weight/temperature and time/month details, sourced from dataset which contained crucial state/event labels. The preparation process in-volves handling data instability and gaps using metadata, aggregating the data (into 15-min or 4-hour intervals), applying interpolation and smoothing, performing feature engineering (like calculating relative weight change), segmenting data into time windows (6 hours for point states, 96 hours for global states), balancing the classes, and splitting the data into training and testing sets. Two LSTM models were developed. The first model identified point states (swarming, feeding, honey removal) using features like weight change, month, and hour, achieving 96% overall accuracy. The second model classified global states (normal function, colony death, spring warm-up, honey collection) using features such as weight change, temperature ratio, and month, reaching 92% overall accuracy. These results confirm the effectiveness of utilizing LSTMs combined with time-series sensor data for the reliable identification of diverse beehive states.