Methods of Intelligent Data Analysis in Hive State Assessment Problem

A new method to solve the topical problem of classifying the state of a beehive for the presence or absence of a queen bee in it is given. The efficiency of such classification based on the parameters of bee sound (buzzing) records using a number of modern machine learning methods is estimated. Bee sounds are described by features based on the averaged signal energy in mel-spectral frequency bands (mel-spectral and mel-cepstral coefficients) and their first time derivatives. For the experiments, an original corpus of records of bee sounds collected at an apiary in an automated mode using smartphones installed in the hives was prepared and annotated. The set of machine learning methods to be tested included logistic regression, support vector method, gradient boosting, and a statistically weighted syndrome method developed by the authors. As a result, the methods of support vectors and logistic regression are experimentally shown to allow for correct classification of the current state of the hive up to high accuracy (more than 99%), with the absence or presence of a queen bee determined by the parameters of bee buzzing. The informativeness of mel-spectral coefficients turned out to vary significantly depending on their center frequencies. The maximum ability to distinguish hives with a queen bee from hives without it is observed for the coefficients corresponding to frequencies in the range 70–130 Hz. As the frequency of mel-spectral coefficients grows, their informativeness starts to decrease gradually, so that for the indicators corresponding to frequencies from 1 kHz and above the distinguishing ability for the problem involved is rather low and does not change significantly with the increase of the frequency of indicators. The obtained results prove that machine learning methods and mel-spectral features are promising to be used to solve the problem of automated noninvasive determination of the presence or absence of a queen bee in the hive and that algorithms and technical tools can be developed for remote automated monitoring of the hive state.