Algorithm for Diagnostics of an Aircraft Electromechanical Drive Using a Fully Connected Neural Network of the Autoencoder Type

This paper proposes an algorithm for analyzing the technical condition of an electromechanical actuator deflecting the control surface of an unmanned aerial vehicle using machine learning models designed for anomaly detection. These models address the well-known novelty detection problem in observed data, which is interpreted as a fault in the monitored system. Computational experiments were conducted using data generated by simulating the actuator operation. Faults were considered. The procedure for splitting data into training and test sets is described. A fully connected neural network with an autoencoder architecture was selected to solve the anomaly detection problem. The paper discusses the extraction of diagnostic features and presents a methodology for deriving them from time series. The efficiency of the proposed diagnostic algorithm is estimated on various test sets, and the results are analyzed.