摘要
由於環境污染惡化與燃料汽車使用費用日漸增加問題，使車商聚焦在電動車研究及開發。本文使用電流訊號分析作為研究對象，診斷無刷馬達軸承與線圈異常，預防或是減少電動車馬達之故障機率。 本論文乃建立可診斷個無刷直流馬達(Brush DC motor, BLDC)故障的自動辨識系統。研究首先對於馬達軸承(Bearing)與線圈(Winding)異常，需要量測馬達在正常及故障下的電流訊號，然後使用小波多重解析(Multi-resolution analysis, MRA)、本質模態函數(Intrinsic mode function, IMF)與希爾伯特-黃轉換(Hilbert-Huang transform, HHT)等三種訊號處理方法分析馬達電流訊號。然後藉由人工智慧演算法來辨識馬達電流訊號，如倒傳遞類神經網路(Back propagation neural network, BPNN)與K最鄰近分類法(K nearest neighbor algorithm, KNN)。 其次，提出結合K-means聚類演算法與人工蜂群演算法(Artificial bee colony algorithm, ABC)，計算特徵與資料相關來尋找最佳特徵。並且使用UCI資料庫(UC-Irvine machine learning repository)來證實提出之特徵選取可以降低特徵數目。 最後，透過加入雜訊測試以取得辨識系系統之抗噪能力。實驗結果得到，使用希爾伯特-黃轉換分析馬達電流訊號可成功辨識馬達各種故障。此外，處於雜訊環境中，可藉由特徵選取策略減少系統準確率下降。由上述可知，希爾伯特-黃分析對於馬達故障檢測有良好表現，而提出特徵選取提高系統抗雜能力。
並列摘要
Due to the worse air pollution and the use-cost raising of fuel vehicles, the vehicles manufacturers focus the research and development of electric vehicles (EV). This thesis is using analyzing motor current signal as a research to diagnosis the motor bearing and the winding faults to prevent or reduce the motor fault probability of EV. In the thesis, an automatic diagnosis system for a brushless DC motor (BLDC) fault is established. First, considering different motor damage kinds, as bearings and windings, the system measured the motor current signal in normal and fault operating. Then, three signal processing methods, as the wavelet multi-resolution analysis (MRA), the intrinsic mode function (IMF) and the Hilbert-Huang transform (HHT), were used to analyze the motor current signal. Besides, artificial intelligent algorithms, as back-propagation neural network (BPNN) and K-nearest neighbor (KNN), were used to identify motor current signals. Second, an algorithm method combine the K-means clustering algorithm with the artificial bee colony algorithm (ABC) was proposed to select best feature by calculating the correlation between features and data. So that, a example using UCI database to verify feature selection method would be reduce the dimension of features. Finally, a noise test was added to obtain the noise resistance of the classification system. As results, the motor current signal using HHT would correctly identify various motor fault kinds. Otherwise, it would improve the accuracy drop problem of the classification system by feature selection strategy in the noise environment. Summary, the use of HHT for motor faults had better performance and the proposed feature selection method increased the noise tolerance of the classification system.