傳統的馬達故障診斷系統,是利用單一感測器收集馬達訊號加以分析。然而單一訊號可分析的故障種類有限,因此本論文使用馬達振動與馬達定子電流訊號進行分析,達到增加診斷故障種類的目的。 針對振動訊號分析,本論文使用田口法對神經網路之輸入倍頻做篩選。利用田口直交表特性,以較少實驗次數的組合,找出較不影響神經網路學習的輸入因子,以節省類神經網路學習與訓練的時間。為了避免省略第一與第二倍頻可能對故障種類識別造成影響,本研究對第一與第二倍頻建立一個故障識別指標,並配合動態結構神經網路識別,作為綜合診斷之依據。在電流訊號分析方面,本論文利用馬達定子頻譜特性,當故障發生時偵測故障頻率點較大振幅諧波,進而與健康馬達分析比對,作為轉子斷條故障診斷之依據。 本論文使用MATLAB軟體撰寫訊號處理、田口方法、類神經網路、電流訊號分析等功能模組,然後利用Visual Basic整合所製作人機介面並增加電流分析介面以建構出智慧型馬達故障診斷系統。由實驗結果可知,本論文所研製系統的確具有較佳的類神經網路訓練效率並可研判較多故障種類。
Traditional motor fault diagnosis system uses only one sensor to collect motor signals and then analyze them. However, the fault category which can be analyzed by single signal source is limited. Therefore, this work measures motor vibration signal and motor stator current signal to increase the types of faults. For vibration signal analysis, this thesis applies the Taguchi method to filter out the input neurons of the neural network. By utilizing the characteristics of the Taguchi orthogonal arrays, the fewer combinations of experiments can be used to find out the input factors with less influence of the learning of the neural network and consequently the training time is reduced. In order to avoid that omitting the first and the second harmonic may affect the failure diagnosis, this work establishes an index for the failure identification which is based on the features of the first and second harmonics. Together with the identification results of dynamic structural neural network, the diagnosis can be done. As to the current signal analysis, by detecting the magnitude of harmonic at failure frequency, this thesis uses the spectral characteristics of motor fault to compare with that of a healthy motor for the basis of fault diagnosis. The functional modules of the signal processing, the Taguchi method and neural networks are developed by using Matlab software. To establish the intelligent motor fault diagnosis system, all the functions are integrated and executed by using the friendly graphical user interface written by Visual Basic and an additional interface for the current analysis is added. From the experimental results, it is seen that the developed diagnosis system equips with superior training efficiency of neural network and detects more fault types.