以綜合感測器為基礎之小風扇損壞初期快速檢測系統

Abstract

近年全球製造業逐漸走向智慧化、自動化。產線自動化方面，檢測之自動化也是關鍵的一環。長期以來風扇製造廠皆以傳統之人耳聽取異音作為瑕疵檢測的指標，證實問題重重。本研究使用隨產線、檢測自動化而愈趨重要之綜合感測器(Sensor Fusion)技術，透過分析聲音與振動兩種訊號提高檢測之穩定性與準確性。在考量量產環境所需匹配之檢測速度，本研究僅以啟動段3秒之訊號作為分析目標，以建立一個有效快速之風扇損壞初期診斷流程。 為提升檢測速度與流暢度，本研究選用傳統之傅立葉轉換與方均根值作為分析聲音訊號與振動訊號之演算法，在執行訊號處理後進行特徵擷取，以解決因綜合感測器而造成之資料維度過高的問題。本論文最後利用擷取之特徵建立支持向量機與決策樹兩種機器學習模型，並以經專業聽音員判斷是否符合出廠標準為樣本的分類作為標籤。使用36個樣本進行訓練，並針對新的9個樣本進行模型之測試，結果發現無論是支持向量機或是決策樹模型準確率皆達100 %。證實兩種模型皆務實可行。

On the trend of developing smart manufacturing and digitalization factory, the automation of both production line and quality inspection has become an essential element nowadays. In comparison with conventional fan manufacturing factories which make use of diagnostic inspections by human senses, in this research an approach named multi-sensor data fusion to achieve more accurate and reliable results by using two kinds of sensors such as accelerometer and microphone is utilized. In considering practical application requirement, signals sampling for 3 seconds is set, and RMS and FFT analytical methods are employed in this study. Feature extraction is then followed to solve the problem of high data dimensions due to multi-sensor data fusion. Two kinds of machine learning models, SVM and decision tree, are applied using the labeled samples that had been classified by the professional fan quality controllers. Using 36 samples for training as well as other 9 samples for testing, and the support vector machine and decision tree model are found accurate in making correct diagnosis, which validates the model of the diagnosis system proposed in this thesis.