Zheng Lihui, Fan Xu, Kang Zongshan, Jin Xinjun, Zheng Wenchao, Fang Xiaofen
Faculty of Mechanical and Electrical Engineering, Quzhou College of Technology, Quzhou, Zhejiang, 324000, China.
School of Mechanical and Electrical Engineering, Aksu Industry Polytechnic College, Aksu, Xinjiang, 842000, China.
Sci Rep. 2025 Jul 1;15(1):21478. doi: 10.1038/s41598-025-07546-w.
With the rapid development of new energy vehicle technology, electric drive systems play a crucial role in the modern automotive industry. Ensuring the efficient and stable operation as well as reliability of electric drive systems has become a critical task. In order to prevent serious faults in the short-term leading to potential accidents, this paper proposes an innovative approach for embedding the Token Merging (ToMe) algorithm into the Vision Transformer (ViT), called the VToMe algorithm and combining it with the Bidirectional Gated Recurrent Unit (BiGRU) network to form the VToMe-BiGRU architecture for electric drive system fault prediction. Specifically, the VToMe algorithm achieves stable detection of medium to long term system faults, while the BiGRU network achieves rapid fault prediction in the short term. The VToMe-BiGRU is an intelligent analysis method applied to automobile workshops, which is closer to the data source for data processing and analysis, alleviates the strong dependence on real-time network transmission, reduces the time consuming and labor-intensive process of manually extracting and analyzing the features, and improves the accuracy and reliability of the fault prediction. The optimized VToMe-BiGRU algorithm combines the Transformer model and the BiGRU network, which effectively captures the critical features in the electric drive system data, thus improving the fault prediction performance. Experimental validation on real-world electric vehicle (EV) maintenance datasets demonstrates outstanding performance of the proposed method. The multi-class fault classification achieves an average accuracy of 93.49% with a 32×32 patch size, outperforming state-of-the-art ViT++ by 0.12% while enhancing inference speed by 28% (32 FPS vs. 25 FPS for ViT++) to balance high precision and real-time efficiency. The short-term prediction yields a root-mean-square error (RMSE) as low as 6.33 and an accuracy (ACC) of 74.7% for complex fault modes such as bearing inner ring fault, surpassing traditional GRU/RNN models by over 20% in prediction accuracy. Moreover, the VToMe algorithm reduces computational complexity by 25% through hierarchical token merging, enabling efficient processing of high-dimensional sensor data without performance degradation. This research establishes a robust framework for real-time diagnosis of EV drive systems, effectively detecting critical faults like battery over-discharge and motor encoder errors with minimized false positives (FP < 5%), enhancing system reliability, reducing maintenance costs, and supporting proactive safety measures in EV applications.
随着新能源汽车技术的快速发展，电驱动系统在现代汽车工业中发挥着至关重要的作用。确保电驱动系统的高效稳定运行以及可靠性已成为一项关键任务。为了防止短期内严重故障导致潜在事故，本文提出了一种创新方法，即将令牌合并（ToMe）算法嵌入视觉Transformer（ViT）中，称为VToMe算法，并将其与双向门控循环单元（BiGRU）网络相结合，形成用于电驱动系统故障预测的VToMe - BiGRU架构。具体而言，VToMe算法实现了对中长期系统故障的稳定检测，而BiGRU网络实现了短期内的快速故障预测。VToMe - BiGRU是一种应用于汽车维修车间的智能分析方法，它更接近数据源进行数据处理和分析，减轻了对实时网络传输的强烈依赖，减少了手动提取和分析特征的耗时费力过程，并提高了故障预测的准确性和可靠性。优化后的VToMe - BiGRU算法结合了Transformer模型和BiGRU网络，有效捕捉了电驱动系统数据中的关键特征，从而提高了故障预测性能。在实际电动汽车（EV）维修数据集上的实验验证表明了该方法的卓越性能。多类故障分类在32×32补丁大小下平均准确率达到93.49%，比最先进的ViT++高出0.12%，同时推理速度提高了28%（ViT++为25 FPS，VToMe - BiGRU为32 FPS），以平衡高精度和实时效率。对于诸如轴承内圈故障等复杂故障模式，短期预测的均方根误差（RMSE）低至6.33，准确率（ACC）为74.7%，预测准确率比传统GRU/RNN模型高出20%以上。此外，VToMe算法通过分层令牌合并将计算复杂度降低了25%，能够在不降低性能的情况下高效处理高维传感器数据。本研究建立了一个用于电动汽车驱动系统实时诊断的强大框架，有效检测电池过放电和电机编码器错误等关键故障，误报率最小化（FP < 5%），提高了系统可靠性，降低了维护成本，并支持电动汽车应用中的主动安全措施。
