類神經網路與田口法於引擎故障聲訊辨識之應用

An Application of Artificial Neural Network and Taguchi’s Method on Engine Fault Noise Recognition

隨著現代科技進步，今日汽車引擎之故障檢修皆以電腦診斷方式進行，惟此方式對於因機件異常所產生的噪音聲訊卻無法判斷。本論文旨在以類神經網路結合田口法來建構一套有效率的引擎故障聲訊辨識系統，以期在一般引擎故障檢修範圍之外，能夠及早發現其它的機件異常現象，以協助維修人員正確辨識故障來源。本研究係以倒傳遞類神經網路結合田口法為研究架構，透過類神經網路學習來辨識引擎在不同狀況下所產生之故障聲訊。在聲訊正規化部分，採用OTA及TPM兩種不同正規化處理方式。建構類神經網路模式時選擇其輸入層單元數、隱藏層單元數、學習速率及學習循環次數等四項控制因子，並採用田口法直交表的配置，在不同水準條件下進行實驗，配合訊雜比、變異數分析及F檢定之計算，分析控制因子對網路辨識率的影響程度，以獲得最佳的參數組合。研究結果顯示引擎故障聲訊最佳辨識率為77.1％。

Due to the improvement of modern technology, the detecting of vehicle engine’s faults can be achieved by computer. However, computer can’t detect the noise resulting from the abnormalities of mechanical parts. The objective of this research is to combine artificial neural network and Taguchi’s method to construct an engine fault recognition system. It is expected to detect the abnormalities of engine before it breakdowns, and to assist mechanical technicians to identify the origin of the problem. This research, constructed by artificial neural network and Taguchi’s method, is to identify different engine fault noises by network learning process. For noise signal normalization, OTA and TPA methods are used. In the construction of artificial neural network model, number of input layer, number of hidden layer, learning rate and learning cycle number are chosen as control factors. Taguchi’s orthogonal array is then used to conduct the calculations in different levels. Signal to noise ratio, analysis of variance, and F-test are applied to analyze the effect of different control factors on the engine fault recognition rate, and to accomplish optimum combination for relevant control factors. The best fault recognition rate achieved in this work is 77.1%.