Aplicação de aprendizado de máquina na identificação de reações de pragas a compostos semioquímicos

Abstract

This study explores the application of machine learning techniques to identify pest reactions to semiochemical compounds. The analyzed time series consist of insect reaction data obtained through electroantennography (EAG) and responses from a gas chromatography system with flame ionization detector (GC-FID). The developed methodological pipeline includes preprocessing steps such as removal of unusable data, data standardization, and frequency equalization between the series using two distinct approaches: removal of excess data and interpolation. To unify the signals in each experiment, the minimum absolute value between the series was applied, a technique that preserves simultaneous peaks—potentially corresponding to insect reactions to semiochemicals—while reducing isolated peaks considered noise. The unified series were then analyzed using machine learning models for anomaly detection, including K-Means and Isolation Forest. A comparative analysis of the equalization approaches showed that the impact of interpolation was low, with minimal influence on K-Means performance. Both models achieved good results in detecting reactions, with K-Means using sliding windows being the most effective for identifying the responses.