機械故障診斷之非穩態信號特徵擷取技術

Feature Extraction Techniques for Nonstationary Signal in Fault Diagnosis of Machinery

張耕與(Kang-Yu Chang)

指導教授 ： 張永鵬 ； 康淵

中原大學/工學院/機械工程研究所/碩士(2011年)

https://doi.org/10.6840/cycu201101080

摘要

以往機械故障診斷環境在轉速固定或穩態狀態下，如核電廠主汽機所使用的泵、馬達等關鍵設備，應用頻譜分析方法(如傅立業轉換，Fast Fourier Transform, FFT)提取信號之頻譜特徵，作為辨識故障成因之根據。但對於瞬間轉速變化大的機械設備，如風力發電機齒輪箱或3C組裝用之機械手臂等，所產生的信號模式大部分為非穩態信號，其信號特徵會隨分析時間長度而平均化，無法被突顯，導致在故障診斷或辨識上之困難。
因此，本文利用短時傅立業轉換(Short-Time Fourier Transform，STFT)及小波轉換(Wavelet Transform，WT)結合轉速頻率階次化方法，擷取非穩態狀態信號之階次特徵。由於，時頻階次譜之信號特徵不因轉速變化而改變，其對應之特徵信號亦將固定，可有效作為機械設備在非穩態狀態下之故障辨識。
利用此時頻階次譜，本文分別使用倒傳遞類神經網路(Back propagation neural network，BPNN)及自組織映射圖(Self-Organizing Map, SOM)，做為故障診斷之方法。然而，時頻階次譜特徵資料量龐大，導致訓練時間冗長，因此，結合主成分分析法(Principal Components Analysis，PCA)保留時頻階次譜之主要成份，達到資料量降維、縮短類神經網路之訓練時間及提高正確率之效果。

關鍵字

倒傳遞類神經網路 ； 自組織映射圖 ； 非穩態信號 ； 短時傅立業轉換 ； 小波轉換 ； 時頻階次譜 ； 主成份分析法

並列摘要

In the past, fault diagnosis of machinery in the stationary rotational speed, such as the steam turbine in nuclear power factory owners use the pump, motor, etc. Often used spectrum analysis methods (such as: Fast Fourier Transform) to extract the signal's features, in order to observe the spectrum response of structure system. However, for rotational speed instantaneous change big mechanical equipment, such as wind turbine gear box or 3C assembly robot arm, etc. The resulting signal mode most of the nonstationary signals. If using Fourier transform analysis of nonstationary signal, the signal's features will be averaged by time length, therefore can not highlight the signal features, Result in fault diagnosis can not effectively identify the causes of failure.
In this paper, using Short Time Fourier Transform and Wavelet Transform ,etc. other analysis methods, analysis the unusual signal of the gear and rotor at nonstationary rotational speed, then combined then speed frequency and Time Frequency spectrum obtain the Time-Frequency Order spectrum. Because, the feature of Time-Frequency Order spectrum do not change due to rotational speed changes, and corresponding of the feature of signal will be fixed, to observe nonstationary signal of the feature of order amplitude change.
By using Time Frequency Order spectrum, In this paper, using Back propagation neural network and Self Organizing Map, as the fault diagnosis method. However, the features of Time Frequency Order spectrum have a large amount of data, resulting in lengthy training time, therefore, combined with Principal Components Analysis retention principal components of the Time Frequency Order spectrum, achieves the material quantity dimensionality reduction, reduction time of neural network training and enhances effect of the accuracy.

並列關鍵字

Nonstationary signal ； Self Organizing Map. ； Back propagation neural network ； Principal Components Analysis ； Short-Time Fourier Transform ； Wavelet Transform ； Time Frequency Order spectrum

參考文獻

2. G. Dennis, “Theory of communications”, Journal of the Institution of Electrical Engineers, vol. 93, No.4 pp. 429–457, 1946.

3. F. Al-Badourl, L. Chedect, and M. Suna, “Non-stationary vibration signal analysis of rotating machinery via time-frequency and wavelet techniques,” Information Sciences Signal Processing and their Applications (ISSPA), pp.21-24, 2010.

4. Lin J., Zuo M.J., Gearbox fault diagnosis using adaptive wavelet filter, Journal of Mechanical Systems and Signal Processing 17 (6)(2003) 1259–1269.

5. Zheng H., Li Z., Chen X., Gear fault diagnosis based on continuous wavelet transform, Journal of Mechanical System and Signal Processing 16 (2002) 447–457.

6. Lin J., Zhang A., Fault feature separation using wavelet-ICA filter, International Journal of NDT&E (2005) 421–428.