啟動馬達為車輛動力系統中關鍵元件，其健康狀態對車輛可靠性具決定性影響。隨著循環經濟與再製技術之推進，啟動馬達之再製成為產業關注重點，惟傳統檢測方式高度依賴拆解與人工判斷，檢測效率低落且具潛在誤判風險，亟需導入即時、非破壞性之健康評估系統。
本研究針對啟動馬達動力傳動模組中最易耗損之碳刷元件，設計一套結合電氣訊號、熱影像之多模態健康預測架構。實驗中透過耐久測試平台蒐集馬達運作期間之電壓、電流與熱影像資料，並以數位孿生模型生成預訓練資料，強化模型於樣本不足條件下之學習穩定性與泛化能力。核心分類模型採用卷積、ViT (Vision Transformer) 編碼器與MMTransformer (Multimodal Transformer) 架構，對碳刷健康狀態進行評估。
本研究除了進行資料處理流程與特徵萃取分析，亦透過消融實驗與多層注意力比較測試，量化各模態與模型結構對分類效能之貢獻。實驗結果顯示，結合熱影像與電氣訊號之多模態模型可顯著提升對磨耗過渡區之辨識能力，並有效降低因單一模態訊號異常所造成之誤判風險。最終模型在碳刷健康狀態分類任務上達成高準確率與穩定性，具備部署於回收檢測現場與嵌入式邊緣裝置之潛力。

The starter motor is vital to a vehicle’s powertrain, but traditional inspections rely on disassembly and manual judgment, leading to inefficiency and potential errors.
This study focuses on carbon brushes—the most wear-prone component—proposing a multimodal health assessment system that fuses electrical signals and thermal images. A digital twin generates pretraining data to improve model robustness with limited samples.
To address the limitations of single-modality diagnostics for carbon brush wear, this study develops a multimodal health assessment model that integrates electrical signal and thermal image analysis. The framework employs a convolutional network for local feature extraction, a Vision Transformer (ViT) encoder for global context modeling, and an MMTransformer for cross-modal fusion, enabling accurate and stable classification of carbon brush health states.

[1]StudentLesson, Engine Starter Motor: Understanding its role in starting your vehicle [Video], 2022. [Online]. Available: https://www.youtube.com/watch?v=IeyLbHELT8Y
[2]C. -P. Yi, Y. J. Lin, P. -J. Ho, W. -D. Chung, P. -H. Chou and S. -C. Yang, "A CUSUM-based adaptive bearing fault features tracking method for RUL estimation," 2023 IEEE Energy Conversion Congress and Exposition (ECCE), Nashville, TN, USA, 2023, pp. 1851-1856
[3]A. Choudhary, D. Goyal and S. S. Letha, "Infrared Thermography-Based Fault Diagnosis of Induction Motor Bearings Using Machine Learning," in IEEE Sensors Journal, vol. 21, no. 2, pp. 1727-1734, 15 Jan.15, 2021
[4]M. Piechocki, T. Pajchrowski, M. Kraft, M. Wolkiewicz, and P. Ewert, “Unraveling induction motor state through thermal imaging and edge processing: A step towards explainable fault diagnosis,” *Eksploatacja i Niezawodność Maintenance and Reliability*, vol. 25, no. 3, 2023
[5]A. Vaswani et al., “Attention is all you need,” in Proc. 31st Int. Conf. Neural Information Processing Systems (NeurIPS), Red Hook, NY, USA: Curran Associates Inc., 2017, pp. 6000–6010.
[6]A. Dosovitskiy et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020.
[7]Y. Moroto, K. Maeda, R. Togo, T. Ogawa, and M. Haseyama, “Multimodal transformer model using time-series data to classify winter road surface conditions,” Sensors, vol. 24, no. 11, p. 3440, 2024
[8]蔡忠翰, “深度學習應用於啟動機動力傳動模組品質檢測,” 碩士論文, 國立成功大學, 台南市, 2025.
[9]The Engineering Mindset, How a car’s electric starter motor works [Video],YouTube.[Online].Available: https://www.youtube.com/watch?v=7eN1gxH6lE4[Accessed: Jul. 14, 2025]
[10]K.K.News,“汽車啟動馬達與起動器的工作原理解析,” *DailyVie, https://kknews.cc/zh-my/car/3xk8aao.html, [Accessed: Jul. 14, 2025].
[11]P. Lorrain and D. R. Corson, Electromagnetic Fields and Waves, 3rd ed. New York: Freeman, 1998.
[12]N. S. Nise, Control Systems Engineering, 7th ed. Hoboken, NJ: John Wiley & Sons, 2019.
[13]C. K. Alexander and M. N. O. Sadiku, Fundamentals of Electric Circuits, 4th ed. New York: McGraw-Hill, 2008.
[14]DAHKEECO.,“Allproductlist,”[Online].Available: https://www.cens.com/cens/html/zh/supplier/supplier_home_22015.html. [Accessed: Jul. 15, 2025].
[15]Python Software Foundation, “Python Logo,” [Online]. Available: https://www.python.org/community/logos/. [Accessed: Jul. 17, 2025].
[16]東元電機股份有限公司,“東元感應電動機使用說明書,”技術文件,2012.[Online].Available:https://file.yzimgs.com/392546/2012061512424275.pdf. [Accessed: Jul. 15, 2025].
[17]新瀚工業股份有限公司, “馬達保護的基本觀念與方式,” Shini.com,[Online].Available:https://www.shini.com/ep_edm/tw/contect_589.html. [Accessed: Jul. 15, 2025].
[18]I. Khalfaoui-Hassani, “Dilated convolution with learnable spacings,” arXiv preprint arXiv:2112.03740v4, May 11, 2023. [Online]. Available: https://arxiv.org/abs/2112.03740
[19]K. O’Shea and R. Nash, “An introduction to convolutional neural networks,” arXiv preprint arXiv:1511.08458, Nov. 2015. [Online]. Available: https://arxiv.org/abs/1511.08458
[20]M.-C. Popescu, V. Balas, L. Petrescu-Popescu, and N. Mastorakis, “Multilayer perceptron and neural networks,” *WSEAS Trans. Circuits Syst.*, vol. 8, pp. –, 2009.
[21]A. Mao, M. Mohri, and Y. Zhong, “Cross‑Entropy Loss Functions: Theoretical Analysis and Applications,” in *Proceedings of the 40th International Conference on Machine Learning (ICML 2023)*, Honolulu, Hawaii, USA, 2023.
[22]D. Berrar, “Cross-Validation,” in Encyclopedia of Bioinformatics and Computational Biology, S. Ranganathan, M. Gribskov, K. Nakai, and C. Schönbach, Eds. Amsterdam, The Netherlands: Elsevier, 2018, vol. 1, pp. 542–545.