INTELLIGENT SYSTEM FOR CRITICAL STATE DETECTION OF MULTIROTOR UAVs BASED ON RECURRENT NEURAL NETWORKS

Abstract

Relevance. Multirotor unmanned aerial vehicles are rapidly expanding their scope of application in logistics, infrastructure monitoring, remote sensing, and the defence sector. The growing autonomy of flight missions imposes increasingly stringent requirements on the reliability of onboard systems and the quality of automatic platform state diagnostic algorithms. Standard threshold-based protection mechanisms in ArduPilot and PX4 autopilots are unable to recognise complex multi-sensor anomalies that manifest only through interactions among several channels over a given time interval.

Problem statement. Most existing UAV diagnostic approaches address a binary task (normal / anomaly) or are limited to a single failure type without considering the temporal structure of telemetry sequences. The task of simultaneously classifying five platform state categories in real time on a real multi-sensor dataset remains insufficiently explored, which determines the relevance of this study.

Proposed solution. This paper presents the development and evaluation of an automated critical state detection system for multirotor UAVs based on multi-sensor flight telemetry analysis using deep learning. The proposed approach employs a single-layer LSTM recurrent neural network with 64 hidden units that processes sliding windows of 20 steps over 17 time-series parameters of the inertial measurement unit, magnetometer, and flight controller. A comparative analysis of anomaly detection methods — threshold rules, SVM, Random Forest, MLP, and 1-D CNN — demonstrates the superiority of the LSTM architecture for telemetry time-series classification. The paper details the hyperparameter selection process, data preprocessing, sliding-window construction, and model quality evaluation on a stratified test subset.

Results and conclusions. The model was trained on a dataset of 12 253 MAVLink telemetry records labelled across five state classes: normal flight, GPS anomaly, IMU anomaly, engine failure, and RC link loss. The overall classification accuracy on the test set reaches 99 %, with a weighted F1-score of 0.99, outperforming methods that do not account for temporal structure (SVM: 83–87 %). The GPS Anomaly, IMU Anomaly, and Engine Anomaly classes are classified with F1 = 1.00 owing to distinct signatures in their respective sensor channels. The RC Anomaly class is the most challenging (recall = 0.87) due to a limited number of training examples and the partial resemblance of the initial link-loss phase to normal flight. Processing latency for a single window on a CPU is under 2 ms, enabling real-time operation without specialised hardware. The developed demonstration module confirmed an anomaly detection latency of 1.2–1.8 s from failure onset, which is sufficient for autopilot emergency procedure activation.