Logo image
Back
An efficient deep attention framework for multi-rotor UAVs fault diagnosis
Journal article   Open access   Peer reviewed

An efficient deep attention framework for multi-rotor UAVs fault diagnosis

Md. Najmul Mowla, Davood Asadi, Mohammad Khaneghaei and Ferdous Sohel
Applied mathematics and computation, Vol.530, 130170
2026
DOI: https://doi.org/10.1016/j.amc.2026.130170
pdf
Published5.99 MBDownloadView
Open Access CC BY V4.0

Abstract

Deep learning High-fidelity simulation data Improved attention mechanism Multirotor UAV fault
Motor faults in multirotor unmanned aerial vehicles (UAVs) jeopardize flight integrity and mission safety; practical operations require real-time detection with reliable severity estimation. Existing diagnostic frameworks struggle with nonlinear, noise-prone telemetry, often lack robust temporal modeling, and underperform at severity classification while being too heavy or unstable for onboard use. We introduce an enhanced multi-head self-attention mechanism-convolutional-LSTM network (EMHSAM-CLN), a compact deep attention framework that couples a convolutional-recurrent backbone with an enhanced multi-head self-attention mechanism for stable, noise-robust sequence modeling and efficient onboard inference. We develop a high-fidelity six-degrees-of-freedom (6-DoF) simulation dataset with graded motor faults (5-40% thrust reduction), realistic sensor noise (e.g., gyroscope: 0.01 deg/s; accelerometer: 0.4×10−3 g), and mission phases spanning takeoff, hover, waypoint tracking, and landing. On nine classes (No-Fault plus eight severity levels), EMHSAM-CLN attains 98.83% accuracy with macro-precision/recall/F1 near 98% and area under the receiver operating characteristic curve (AUC) up to 1.00, while remaining compact (0.0578 M parameters; 0.220 MB, 32-bit floating point). It outperforms strong attention baselines (+4.28 percentage points (pp) versus a convolutional-block attention variant; +6.15 pp versus a standard multi-head self-attention (MHSA) variant) and classical machine learning (+6.77 pp versus extreme gradient boosting (XGBoost)). Ablations attribute an absolute accuracy gain of 4.43 pp to the entire architecture relative to the best non-attention variant. These results indicate that EMHSAM-CLN provides a severity-aware and computationally compact framework for intelligent UAV health monitoring. Real-time feasibility is supported by the reported complexity and CPU-based runtime profiling; embedded SWaP validation (latency, peak memory, and power) remains a key direction for future work.

Details

Metrics

1 File views/ downloads
2 Record Views
Logo image