DIAGNOSIS OF THE POWER PLANT OF UNMANNED AERIAL VEHICHLES BASED ON ELECTRIC DRIVE

Abstract

The article addresses the issue of diagnosing the power plant of unmanned aerial vehicles (UAVs) based on electric drives using intelligent data analysis technologies. In the context of modern challenges caused by military aggression against Ukraine, the relevance of this research is determined by the need to ensure the safe operation of autonomous aerial systems in regions with damaged infrastructure and limited telecommunication functionality. It has been established that the efficiency of UAV operation largely depends on the condition of its power unit, and therefore, on the timeliness and accuracy of diagnosing electric drives during flight.

The main focus is placed on the development of conceptual foundations for constructing a monitoring system of the technical condition of the electric drive using artificial neural networks (ANN) within the TensorFlow environment. The study analyzes current trends in deep learning technologies that enable the recognition of nonlinear dependencies between the operating parameters of the power plant and its diagnostic indicators. The structure of a typical neural network, including input, hidden, and output layers that perform nonlinear transformations of input signals, is described. Optimization algorithms for the training process (including stochastic gradient descent, Adam, RMSProp) and loss functions applied for the classification of the technical condition of electric motors during operation are characterized.

The research substantiates the feasibility of using the TensorFlow Keras API for the construction, training, and testing of diagnostic models for UAV power plants. It is shown that the proposed approach allows for high-accuracy determination of the operational state of electric motors in real time, detection of deviations from nominal operating modes, and prediction of potential failures. Simulation results confirm that the developed model can classify the power unit condition into three categories—operational, conditionally operational, and critical—with an accuracy exceeding 95%.

The process of constructing the neural network model is described in detail, including the selection of hyperparameters, the number of hidden layers and neurons, activation functions (ReLU, Softmax), and training parameters. The importance of hardware acceleration (GPU, TPU) is emphasized for ensuring real-time processing of large volumes of diagnostic data. The results of the analysis of the influence of neural network architecture on the convergence rate of the training algorithm and resistance to overfitting are presented.

It is concluded that the proposed diagnostic method for power plants based on electric drives using artificial neural networks is an effective tool for enhancing the reliability of unmanned aerial vehicle operation, particularly under conditions of limited telecommunication support. The practical implementation of such systems enables autonomous monitoring of the technical state of power elements, timely detection of faults, and prevention of emergency situations during flight. The obtained results have significant scientific and practical potential for further development of integrated diagnostic systems incorporated into UAV autopilot architectures.