隨著感應馬達在各行各業中的數量越來越多且容量也越來越大，如今感應馬達在各種系統中扮演著重要的角色，其感應馬達的可靠性和安全性將直接地影響到整個系統，因此對感應馬達進行故障診斷的研究也越來越受關注，故障診斷系統由三個通用步驟組成：數據收集、特徵擷取和故障分類，其中特徵擷取極大地影響故障診斷準確性。傳統機器學習在數據要進入訓練故障診斷模型之前，必須依賴專家所提取的特徵，而深度學習則是透過深度學習網路自動提取並學習原始數據特徵的有效方法。
本論文即採用卷積神經網路來進行訓練及測試，整個訓練包含三個步驟:訊號前處理、特徵擷取及特徵分類，並且透過五種案例(健康狀態、定子匝間短路、轉子斷條、軸承外環損傷及不對心)所量測的電氣、振動及混合訊號進行分析，並同時進行時域、頻域以及負載狀態為半載、滿載的分析。為了驗證其分類準確率，此研究對其學習到的特徵進行特徵可視化，此方法將多維資料降至二維，透過特徵在二維平面之分布，即可驗證分類準確率。最後得到的重要結果為：1) 卷積神經網路可以透過自行學習有效分辨出五種馬達狀態；2) 當負載變動時，採用振動的頻域訊號相較於電氣、振動時域訊號及電氣頻域訊號的變動影響較小；3) 同時輸入電氣及振動頻域訊號作為混合訊號，使卷積神經網路同時學習兩者之特徵，準確率高達100%。

Induction motors have had performance-capacity improvements and an increasing range of applications, playing a major role in various systems. Their reliability and safety directly influence the performance of an entire system. Therefore, induction motor fault diagnosis has garnered increasing research attention. Fault diagnosis proceeds in three general steps: data collection, feature extraction, and fault classification. Among these steps, feature extraction is especially influential on diagnosis accuracy. Conventional machine learning relies on expert-provided feature data before a fault diagnosis model receives data training, whereas deep learning enables a diagnosis model to automatically acquire data from a deep learning network for effective feature learning from raw data.
This study employed a convolutional neural network (CNN) to train and test a fault diagnosis model. The training was completed in three steps: signal preprocessing, feature extraction, and feature classification. Under five cases of system faults (normal status, short circuit between stator turns, broken rotor bars, and damages to and misalignment of the bearing outer ring) the electrical, vibration, and mixed signals of the motors were measured for analysis at different time and frequency domains and at half or full load. To verify the classification accuracy of the diagnosis model, the features learned by the model were visualized. The multidimensional data were converted to 2D data, and the planar distribution of the feature data was analyzed to verify the model’s classification accuracy. The results were as follows: (1) the CNN effectively identified all five types of motor statuses through automated learning; (2) load changes had less effects on the frequency-domain vibration signals than on the frequency-domain electrical signal and time-domain electrical and vibration signals; (3) when the frequency-domain electrical and vibration signals were simultaneously input as mixed signals, the CNNs simultaneously learned the features of both signals to yield 100% accuracy.

[1] 葛世偉、李明遠。馬達軸承故障線上診斷技術研究。工程科技與教育學刊，第12卷，第1期，第001～007頁。
[2] 人工智慧發展簡史，2019年6月。Available: https://futurecity.cw.com.tw/article/743
[3] Haibo He, Jun Yan, Ping Xie, and Guoqian Jiang, “Multiscale Convolutional Neural Networks for Fault Diagnosis of Wind Turbine Gearbox,” IEEE Transactions on Industrial Electronics, Vol. 66, No. 4, April 2019.
[4] Wenjun Sun, Rui Zhao, Ruqiang Yan, “Convolutional Discriminative Feature Learning for Induction Motor Fault Diagnosis,” IEEE Transactions on Industrial Informatics, Vol. 13, No. 3, June 2017.
[5] Siyu Shao, Ruqiang Yan, and Yadong Lu, “DCNN-based Multi-signal Induction Motor Fault Diagnosis,” IEEE Transactions on Instrumentation and Measurement, Vol. 69, No. 6, June 2020.
[6] Long Wen , Xinyu Li , Liang Gao, and Yuyan Zhang, “A New Convolutional Neural Network-Based Data-Driven Fault Diagnosis Method,” IEEE Transactions on Industrial Electronics, Vol. 65, No. 7, July 2018.
[7] Singh GK, Al Kazzaz SAS, “Induction Machine Drive Condition Monitoring and Diagnostic Research—A Survey,” Electric Power System Research, Vol. 64, No. 2, pp. 145–158, 2003.
[8] W. S. Gongora, H. V. D Silva, A. Goedtel, W. F. Godoy, and S. A. O. da Silva, “Neural approach for bearing fault detection in three phase induction motors,” IEEE International Symposium on Diagnostics for Electric Machines, Power Electronics and Drives (SDEMPED), 2013, pp. 566-572.
[9] Agam G., Gurmeet Singh, and V.N.A. Naikan, “Effective combination of motor fault diagnosis techniques,” 2018 International Conference on Power, Instrumentation, Control and Computing (PICC), Thrissur, India, Jan, 2018.
[10] 徐子權，「電氣與振動檢測法於馬達狀態評估之研究」，碩士論文，國立台灣科技大學，2014年。
[11] 曾思憲，「感應馬達主動型狀態估測系統之研發」，碩士論文，國立台灣科技大學，2018年。
[12] 林芝以，「運用電氣訊號於感應馬達狀態監測與故障診斷之研究」，碩士論文，台灣科技大學電機工程學系，2016年。
[13] 快速傅立葉變換(FFT)，2016年10月。Available: https://blog.csdn.net/shenziheng1/article/details/52891807
[14] Eduardo D. Piedad, Jr，「基於頻率遞歸圖之卷積類神經網路應用於馬達故障診斷」，碩士論文，國立台灣科技大學，2019年。
[15] 卷積神經網路。Available: https://zh.wikipedia.org/wiki/%E5%8D%B7%E7%A7%AF%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C
[16] 初探卷積神經網路，chtseng，2017年9月。Available: https://chtseng.wordpress.com/2017/09/12/%E5%88%9D%E6%8E%A2%E5%8D%B7%E7%A9%8D%E7%A5%9E%E7%B6%93%E7%B6%B2%E8%B7%AF/
[17] 卷積神經網路的運作原理，Jimmy Lin。Available: https://brohrer.mcknote.com/zh-Hant/how_machine_learning_works/how_convolutional_neural_networks_work.html
[18] Hong-Chan Chang *, Yu-Ming Jheng, Cheng-Chien Kuo and Jian-Zhong Feng, “Intelligent Fault Diagnosis System for Rotating Machinery using Convolutional Neural Network,” MDPI energies, July 2020.
[19] Transfer Learning轉移學習，陳明佐。Available: https://medium.com/%E6%88%91%E5%B0%B1%E5%95%8F%E4%B8%80%E5%8F%A5-%E6%80%8E%E9%BA%BC%E5%AF%AB/transfer-learning-%E8%BD%89%E7%A7%BB%E5%AD%B8%E7%BF%92-4538e6e2ffe4