Determination of optimum feature sets for vibration-based sensor events classification

Chickrin Dmitrii Evgen’evich, Golousov Svyatoslav Vladimirovich, Glavatskiy Nikita Vladimirovich, Ermakov Dmitrii Vladimirovich, Stepanov Andrey Nikolaevich, Kokunin Petr Anatol’evich

The paper presents comparative analysis of machine learning feature extraction methods in relation to the problem of three-class classification using the experimental sample. Input data sampling represents a uniformly discretized sequence of normalized amplitudes of actions received from a vibration sensor. The k-nearest neighbors algorithm is used for the classification method. As a result of the investigation, the optimum feature set and optimum metrics of distance from the viewpoint of minimization of an error of classification are determined; based on the considered features, the level of their positive or negative impact on the classification process is found.

Keywords:classifcation, machine learning, vibrations, k-nearest neighbors algorithm, wavelet, cepstrum, kurtosis, skewness.