A High-precision method for detecting rolling bearing faultis in unmanned aerial vehicle based on improved 1DCNN-Informer model

• By combining Fast Fourier Transform (FFT) and Variational Mode Decomposition (VMD) for time–frequency domain feature extraction, multi-scale features within fault signals can be effectively revealed. • Introduced a dual-branch feature extraction network that separately models local time-domain features using 1DCNN and global temporal dependencies with Informer, improving fault detection performance. • Feature enhancement is achieved in the encoder layer of the Informer by integrating the Squeeze-and-Excitation Network (SENet). • A multi-head attention mechanism is employed to fuse features from different branches, enhancing the hierarchical structure and diversity of the integrated features. With increasing unmanned aerial vehicle (UAV) integration and system complexity, motor bearing failures have become more frequent due to long-term high-load operation. Effective vibration feature extraction and an improved classification model are essential for accurate and automated fault diagnosis of UAV motor bearings. This paper presents a novel fault diagnosis method based on a fused 1DCNN-Informer with MATT architecture. The proposed approach integrates signal preprocessing using Fast Fourier Transform (FFT) and Variational Mode Decomposition (VMD), dual-branch feature extraction through One-Dimensional Convolutional Neural Network (1DCNN) and Informer networks, and feature fusion via a multi-head attention (MATT) mechanism to enhance diagnostic accuracy and model robustness. Specifically, FFT and VMD are jointly employed to extract multi-scale time–frequency features, effectively capturing subtle variations in the signals. Subsequently, a dual-branch network processes the signal in parallel, where the 1DCNN branch focuses on local temporal features, and the Informer branch models long-range dependencies. These complementary branches enable comprehensive feature representation. Finally, the MATT module adaptively fuses the extracted features by assigning dynamic weights, thereby improving sensitivity to key fault characteristics. Simulation results show that, under the same preprocessing conditions, it outperforms CNN-LSTM, TimesNet, Autoformer, and the original Informer. The model achieves 99.99% classification accuracy. Experiments confirm its effectiveness in diagnosing UAV motor bearing faults, showing strong practical value.