在馬達旋轉故障診斷的研究議題上,類神經網路識別法因具有較佳的圖形辨識能力而受到歡迎。由於訓練資料無法有效的使類神經網路學習正常頻譜與故障頻譜的差異性,造成對於類神經網路於馬達旋轉故障的研究文獻當中,只能單針對某種故障類型進行診斷,或者是無法有效的將正常馬達與故障馬達振動訊號辨識。因此,本論文提出一種新的振動訊號正規化方式,改善以往文獻上無法將正常馬達與故障馬達振動訊號有效分類以及正規化方式無一定基準的缺點。對於類神經網路演算機制,本論文將結合終端引點與改良式動態結構類神經網路,以增加類神經網路收斂的速度以及鑑別的精確性。除此之外,考慮馬達於工廠當中實際運轉的情況,在週遭低頻雜訊的干擾之下容易造成馬達振動頻譜存有暫時性異常,進而使得類神經網路將訊號識別為異常情況而引發誤警報。因此,本論文引入統計製程管制當中的管制圖技術,搭配西方電器公司定義的區間法則,藉由連續的訊號測量與診斷進行決策。最後,本論文利用測試機台產生故障訊號,並藉此驗證所提出正規化方式搭配終端引點改良式動態結構類神經網路與管制圖結合之診斷系統的有效性。由實驗結果得知,所提出方法的確能夠有效的鑑別出馬達旋轉故障。
In the research of diagnosis of motor rotary faults, the neural network is widely welcomed because it has better capability of characteristic discrimination. Since the training data can’t effectively train the neural network to distinguish the spectrum of a health motor from that of a faulty motor, the conventional neural network based motor fault diagnosis system only classifies a certain type of fault, or even can’t effectively discriminate the healthy motor from the fault vibration signals. Therefore, this thesis proposes a new normalization method to pre-process the vibration signals and thus resolves the aforementioned problem. This method also provides a standardized normalization process. As to the neural network algorithms, this thesis combines the terminal attractor technique and the improved dynamic structure neural network to increase the speed of convergence and the identification accuracy. In addition, the motor operating in the factory often suffers from the low-frequency noise, and a temporary fault signature may appear in the spectrum of the measured vibration signal. In such circumstances, a false alarm occurs by applying the neural network diagnosis method. Therefore, this thesis introduces the control chart technique in statistical process control together with Western electric rules, which can rule out the temporary fault signatures while monitoring the vibration signals. Finally, a test platform is used to generate the faulty motor vibration signals and perform vibration experiments. From the experimental results, it is seen that the proposed method can effectively classifies the fault characteristics.