本研究重點在於建立一套對齒輪故障檢測的方法與系統雛型,分別使用快速傅立葉、小波分析、經驗模態分解法等訊號處理技術來分析齒輪振動故障訊號,並建立齒輪故障對時域與頻域的特徵關係後,結合類神經網路中的倒傳遞神經網路來判斷故障類型。本文先以虛擬實驗方式分別產生無異常與數種齒輪故障之時域波形,經訊號處理後提取特徵參數之特徵值,輸入類神經網路進行訓練,並以輸出結果來探討上述訊號處理技術結合類神經網路的效果與可行性。最後以實例來佐證三種訊號處理法結合類神經網路來診斷齒輪故障類型具有一定的準確率與可靠度。結果顯示,三種方法皆可診斷出正確的齒輪故障類型。
This study concentrates on how to establish a set of gear fault detection methods and prototypes by using several kinds of signal processing techniques. FFT (Fast Fourier Transform), wavelet analysis, and EMD (Empirical Mode Decomposition) are applied to deal with vibration signal of the gear faults. The relationship among the gear faults, time domain, and frequency domain is built up, and combine with BPN (Back-propagation Neural Network) of ANNs (Artificial Neural Networks) to diagnose the type of gears faults. Firstly, the signals of normal and several types of gear faults are generated in the virtual experiment, and then extract characteristic values of feature parameters by using signal processing and train the data by ANN. At last, some examples are proposed to support the methods and prove that the diagnostic capability of ANN is accuracy to diagnose the real type of gear faults. The results showed that the three methods of signals processing combined with ANN can predict the fault of gears correctly.