本研究旨在建立一套結合無載電氣端訊號、非接觸式聲音量測與深度學習之啟動馬達品質檢測方法，以配合產線快速全檢之需求。為減少多段加載測試與動力計使用所造成之時間成本，本研究選用無載階段之電流、電壓與轉速作為電氣端特徵，並加入完整檢測流程之聲音訊號。由於聲音量測具有非接觸式特性，可降低感測器安裝時間，較適合應用於快速檢測場域。
在資料處理方面，本研究將電氣端訊號與聲音訊號轉換為二維時頻圖，以利模型擷取時間與頻率特徵。考量實際產線中不良品樣本數量稀少，容易造成資料不平衡，因此導入生成對抗網路進行資料擴增，產生與真實資料分布相近之樣本，作為後續分類模型訓練資料來源。
分類模型採用 CvT-13 架構，並比較不同特徵組合對良品與不良品分類結果之影響。此外，本研究亦建立 ANSYS 啟動馬達聲學模擬模型，輔助確認馬達主要聲音特徵頻率區間，提升深度學習分類結果之物理可解釋性。

This study aims to develop a starter motor quality inspection method integrating no-load electrical signals, non-contact acoustic measurements, and deep learning to meet the demand for rapid full inspection in production lines. To reduce the time required for multi-stage loading tests and the use of a dynamometer, the current, voltage, and rotational speed measured during the no-load stage are selected as electrical features, while the acoustic signal recorded throughout the complete inspection process is incorporated as an additional feature. Owing to the non-contact nature of acoustic measurement, the installation time of sensors can be reduced, making the proposed approach more suitable for rapid inspection applications.
For data preprocessing, the electrical and acoustic signals are transformed into two-dimensional time-frequency representations, enabling the model to extract both temporal and frequency-domain characteristics. Considering that defective samples are relatively scarce in practical production environments, which may result in data imbalance, a Generative Adversarial Network (GAN) is introduced for data augmentation. The GAN is used to generate synthetic samples with distributions like those of real measurements, thereby providing additional training data for the subsequent classification model.
The classification model is constructed based on the CvT-13 architecture, and different combinations of input features are evaluated to investigate their effects on the classification performance of normal and defective starter motors. In addition, an ANSYS-based acoustic simulation model of the starter motor is established to assist in identifying the major characteristic frequency ranges of motor-generated sound. The simulation results provide physical interpretation of the acoustic features and further enhance the interpretability of the deep learning-based classification results.

[1] V. Murugesan, G. Chandramohan, K. M. Senthil, R. Rudramoorthy, K. L. Ashok, K. R. Suresh, D. Basha, and M. K. Vishnu, "An overview of automobile starting system faults and fault diagnosis methods," ARPN Journal of Engineering and Applied Sciences, vol. 7, no. 7, pp. 812–819, 2012.
[2] A. Vijaya, J. P. S. Meisterknecht, L. S. Angreani, and H. Wicaksono, "Advancing sustainability in the automotive sector: A critical analysis of environmental, social, and governance (ESG) performance indicators," Cleaner Environmental Systems, vol. 16, p. 100248, 2025/03/01/ 2025, doi: 10.1016/j.cesys.2024.100248.
[3] M. S. Jamil, P. N. Gunaratne, and H. Tamura, "A Study on Classification of Faulty Motor Sound Using Convolutional Neural Networks," in Proceedings of International Conference on Artificial Life and Robotics, 2024, pp. 918–922, doi: 10.5954/ICAROB.2024.GS1-2.
[4] A. Martyr and D. Rogers, "Dynamometers: The measurement and control of torque, speed, and power," Engine Testing, p. 235, 2021, doi: 10.1016/B978-0-12-821226-4.00009-7.
[5] 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, 機械工程學系, 國立成功大學, 台南市, 2025. [Online]. Available: https://hdl.handle.net/11296/g6m8gg
[6] L. Hongming, K. Zhang, and S. Han, "A comparison of CNN-based transformer fault diagnosis methods based on voiceprint signal," in 2024 IEEE 13th Data Driven Control and Learning Systems Conference (DDCLS), 2024: IEEE, pp. 819–824, doi: 10.1109/DDCLS61622.2024.10606641.
[7] V. Bui, T. L. Pham, H. Nguyen, and Y. M. Jang, "Data augmentation using generative adversarial network for automatic machine fault detection based on vibration signals," Applied Sciences, vol. 11, no. 5, p. 2166, 2021, doi: 10.3390/app11052166.
[8] 鄭爲中, "基於卷積神經網路之馬達噪音舒適度與轉子設計研究," 碩士, 電機工程學系, 國立成功大學, 台南市, 2021. [Online]. Available: https://hdl.handle.net/11296/9e6d23
[9] B. Thornton, "Audio recognition using mel spectrograms and convolution neural networks," Academia: San Francisco, CA, USA, 2019.
[10] T. Zhang, G. Feng, J. Liang, and T. An, "Acoustic scene classification based on Mel spectrogram decomposition and model merging," Applied Acoustics, vol. 182, p. 108258, 2021, doi: 10.1016/j.apacoust.2021.108258.
[11] B. Saritha, M. A. Laskar, A. M. K, R. H. Laskar, and M. Choudhury, "CACRN-Net: A 3D log Mel spectrogram based channel attention convolutional recurrent neural network for few-shot speaker identification," Computers and Electrical Engineering, vol. 115, p. 109100, 2024/04/01/ 2024, doi: 10.1016/j.compeleceng.2024.109100.
[12] L.-H. Wang, X.-P. Zhao, J.-X. Wu, Y.-Y. Xie, and Y.-H. Zhang, "Motor fault diagnosis based on short-time Fourier transform and convolutional neural network," Chinese Journal of Mechanical Engineering, vol. 30, no. 6, pp. 1357–1368, 2017, doi: 10.1007/s10033-017-0190-5.
[13] J. Gui, Z. Sun, Y. Wen, D. Tao, and J. Ye, "A review on generative adversarial networks: Algorithms, theory, and applications," IEEE transactions on knowledge and data engineering, vol. 35, no. 4, pp. 3313–3332, 2021, doi: 10.1109/TKDE.2021.3130191.
[14] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, "Generative adversarial nets," Advances in neural information processing systems, vol. 27, 2014, doi: 10.5555/2969033.2969125.
[15] L. Gonog and Y. Zhou, "A review: generative adversarial networks," in 2019 14th IEEE conference on industrial electronics and applications (ICIEA), 2019: IEEE, pp. 505–510, doi: 10.1109/ICIEA.2019.8833686.
[16] X. Chen, Z. Chen, S. Chen, L. Wang, and W. Zhai, "Unsupervised GAN with fine-tuning: A novel framework for induction motor fault diagnosis in scarcely labeled sample scenarios," IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–11, 2024, doi: 10.1109/TIM.2024.3446655.
[17] H. Wu, B. Xiao, N. Codella, M. Liu, X. Dai, L. Yuan, and L. Zhang, "Cvt: Introducing convolutions to vision transformers," in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 22–31, doi: 10.1109/ICCV48922.2021.00009.
[18] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, and Y. Xu, "A survey on vision transformer," IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 1, pp. 87–110, 2022, doi: 10.1109/TPAMI.2022.3152247.
[19] A. Mao, M. Mohri, and Y. Zhong, "Cross-entropy loss functions: Theoretical analysis and applications," in International conference on Machine learning, 2023: pmlr, pp. 23803–23828, doi: 10.5555/3618408.3619400.
[20] G. Parmar, R. Zhang, and J.-Y. Zhu, "On aliased resizing and surprising subtleties in gan evaluation," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11410–11420, doi: 10.1109/CVPR52688.2022.01112.
[21] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, "Grad-cam: Visual explanations from deep networks via gradient-based localization," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 618–626, doi: 10.1109/ICCV.2017.74.
[22] I. McCowan, "Microphone arrays: A tutorial," Queensland University, Australia, pp. 1–38, 2001.