Local Peaks: a Novel Machine Learning Feature for Queen Bee Presence Detection

Nowadays, Artificial Intelligence (AI) and Machine Learning (ML) techniques are increasingly used to address complex classification problems, where defining an explicit algorithm capable of producing a reliable solution is not straightforward. Typical applications range from image and audio recognition to environmental monitoring and speech analysis. The performance of a model depends on both the chosen algorithm and parameter configuration, and is assessed through metrics that reveal its strengths and limitations. A key advantage is the separation between training and inference: while training is computationally intensive and can be done offline, inference is lightweight and suitable for IoT devices, enabling local data processing with lower latency and higher energy efficiency without relying on external servers. The main contribution of this thesis is the study and evaluation of different machine learning algorithms, such as Neural Networks and Support Vector Machines (SVM), for the classification of audio recordings collected inside beehives, distinguishing between queen and non-queen bee sounds. In particular, a novel type of input feature, named Local Peaks, has been introduced. These features are based on the detection of local maxima within specific frequency ranges that are characteristic of queen bee sounds, and are designed to provide more discriminative information to the classifiers. Alongside standard feature sets, this new representation has been tested to assess its impact on the accuracy and robustness of the models. Moreover, an additional dataset has been integrated into the experiments, complementing two open dataset, enriching the training material and enabling a deeper analysis of the generalization capabilities of the algorithms. These two open datasets have been extensively studied in the literature, showing promising and effective results across various ML applications. Several experiments were conducted to evaluate and compare the performance of the proposed features and models, with the ultimate goal of identifying the most effective approach for queen bee detection. The obtained results highlight the potential of combining conventional ML algorithms with tailored feature engineering strategies, offering a promising direction for the development of compact, accurate, and energy-efficient systems for real-world applications in precision beekeeping