Fault Detection in Unbalanced Data of Brushed DC Motor Worm and Helical Gears for Automotive Applications Based on CNN-LSTM-Attention

Abstract

Direct Current motors (DC motors) and their worm-helical gear pairs are critical to the NVH performance and driving safety of new energy vehicles. To address the lack of dedicated fault datasets and the severe sample imbalance in existing research, this paper establishes a test bench to collect a real-world vibration dataset under typical operating conditions. A 1,000-point sliding window preprocessing strategy is proposed to mitigate sample imbalance while preserving complete gear meshing characteristics. Furthermore, a hybrid fault detection model integrating CNN, LSTM, and attention mechanisms is designed to effectively extract local impact features, capture temporal dependencies, and adaptively weight fault-sensitive data. Experimental results show that the proposed model achieves 99.23% accuracy and a 99.18% F1 score, with a fault recall rate exceeding the baseline by over 6%, validating the superiority of this approach.