Label-free acoustic monitoring of honeybee swarming: An unsupervised online learning approach

Abstract

Colony losses caused by honeybee swarming remain a major operational chal lenge because departure occurs within minutes, although acoustic changes be gin tens of minutes earlier and could allow timely intervention if detected reli ably. Current detection systems miss this window because the circadian rhythm of individual hive acoustics is not modelled, making it impossible to separate genuine pre-swarming drift from normal day-to-night spectral variation. Re cursive least squares is used to estimate a colony-specific circadian baseline, Mahalanobis distance scoring is applied to quantify deviations, and BOCPD ac cumulates Bayesian evidence of a regime shift, with a threshold derived from warmup data guaranteeing a controlled false alarm rate without labelled record ings and regardless of bee race, season, or microphone placement. A controlled simulation was used for evaluation, yielding 3.5× higher precision than the best label-free baseline, swarm anticipation exceeding 25 minutes before departure, and a false alarm rate held at the nominal 5% target. The low per-frame cost and memory footprint make this algorithm deployable on resource-constrained embedded devices, enabling continuous autonomous hive monitoring without expert supervision.