COMPARATIVE STUDY OF MACHINE LEARNING ALGORITHMS IN THE DIAGNOSTICS OF ELECTROMECHANICAL ACTUATORS FOR SAFETY ENHANCEMENT IN AEROSPACE SYSTEMS

Abstract

The increasing use of aerospace systems highlights the need for developing diagnostic algorithms for electromechanical actuators in control systems. The reliable operation of electromechanical actuators is critical for ensuring the functionality of controlled object, which is essential for safe flight. This paper presents the results of a study on the application of machine learning technologies to develop classification models capable of identifying faults related to gear performance degradation of execution mechanisms. Training and test datasets were generated using a mathematical model of an electromechanical actuator, accounting for faults caused by changes in backlash, Coulomb friction, viscous friction, and key external factors such as temperature, supply voltage, and load. The parameter grid defining technical condition and external conditions was built using optimal space filling and full factorial design methods. The generated datasets were used to investigate a data processing pipeline, including: feature extraction, feature selection (filter and embedded methods) and subsequent building classification model using machine learning techniques. A total of 3990 combinations of data processing methods were generated. Classification models were built using the training dataset (optimal space filling), for which confusion matrices were obtained and F1-scores were calculated on the test dataset (full factorial design). As a result of large-scale computational studies heatmaps were built to visualize the effectiveness of different method combinations, quantitative distributions were developed to assess the relative frequency at which selecting a particular method leads to high classification model accuracy, the best-performing machine learning methods were identified.