摘要
為因應智慧化、自動化製造的趨勢,以及全球製造業普遍缺工的困境,產線及檢測自動化為首要的目標之一。有別於一般風扇製造廠以異音做為檢測之指標,本研究針對冷卻風扇的振動訊號進行分析,除傳統的方均根值、傅立葉轉換外,也使用評估訊號之複雜度的多尺度熵做為訊號處理方式,以找出初期損壞之風扇於振動訊號的特徵。
由於風扇在運轉時會受到流場的影響,將造成其振動型態為循環穩態
(Cyclostationary)之訊號,若使用傳統的傅立葉轉換則有其限制,本研究利用多尺度熵將風扇振動訊號進行不同尺度之粗粒化,使得訊號受到紊流影響之程度降低,再計算其亂度值,發現結果之多尺度熵曲線無論在合格或不合格樣本間的重複性皆相當高,因此將其做為樣本之特徵。
最後利用特徵建立類神經網路之模型,以經專業聽音員判斷是否符合出廠標準為樣本的分類。第一個模型使用 36
個樣本進行訓練,並針對新的 9 個樣本進行模型之測試,準確率達 100
%;第二個模型針對重複性實驗的結果建立,確認模型是否可預測出同樣的結果,使用兩次的結果建立模型,以第三次的實驗結果進行預測,驗證之準確率達
88.9 %,達成以振動訊號對風扇之品質進行辨識的目的。
關鍵字:風扇檢測、多尺度熵、類神經網路、迴轉機械
Abstract
Aiming at long term smart
manufacturing goal, in solving the global problem of skilled
labor shortage, production line and quality testing automation
is one of the primary high agenda issues. In this research
vibrational signal is used for fan QC as compared with
conventional manufacturing factories which apply abnormal sound
detection as index. Apart from traditional analytical tools RMS
value and FFT, multiscale entropy, which estimates the
complexity of signals is also adopted in the present study.
Cooling fans will
be affected by the turbulence air flow while it is operating,
making its vibrational signal in a type of ‘Cyclostationary’.
Traditional FFT has its limitations on analyzing cyclostationary
signals. In this research, multiscale entropy is adapted to make
the signal coarse-grained, decreasing the effect of turbulence.
The multiscale entropy curves of the sample are found to have
good repeatability, and also it gives the characteristic of the
sample quality.
A neural network
model was developed in this research. The labeled samples that
had been classified by the professional fan quality controllers
were used to train the model. The first model obtained using 36
fans as the training samples, and the validation has been made
with 9 new samples, with the accuracy 100 %. Repeated
experiments were also carried out for further observation. The
second model obtained was used to validate the result of the
third repeated experiment, with the accuracy of 88.9 %. The
approach has found to be succeessful in classifying fan quality
by its vibrational signal.
Key words: Fan quality diagnosis,
multiscale entropy, neural network, rotary machine
|