旋轉機械故障診斷方法之探討

Investigations in Fault Diagnosis Methods for Rotating Machinery

王俊傑(Chun-Chieh Wang)

指導教授 ： 張永鵬 ； 康淵

中原大學/工學院/機械工程學系/博士(2008年)

https://doi.org/10.6840/CYCU.2007.00664

摘要

中文摘要

模糊類神經網路為機電設備故障智能診斷的近代方法，以專家知識或實驗結果定義出故障類型與頻譜特徵之關係，並且以一倍頻分量振幅為正規化基底，利用模糊隸屬度函數描述信號特徵之比強度等級。傳統以If-then規則或類神經網路推論故障類型與振動特徵構成的關聯矩陣，由於故障種類繁多，因此模糊類神經網路架構龐大，導致訓練時間冗長，且有尚未確認為專家知識的故障對應信號特徵，新添的或修正的故障診斷規則又需要重新訓練。
本文提出單一故障類型與對應的頻譜特徵關係形成的關聯向量，將整個關聯矩陣分解成獨立的關聯向量，因此可以獨立訓練及診斷計算，不受大盤知識增大或修正的影響；至於大盤知識只考量ㄧ倍頻之分量振幅強度為模糊化時的正規化基底，診斷結果列成輸出矩陣，每行為關聯向量之輸出，各行輸出由故障類型為輸出的行向量之元素，全部行向量同一列元素構成關聯故障類型診斷結果之可信度列向量。
類神經網路的訓練運算中，誤差容易陷入局部最小值，因此本文使用基因演算法，作全域搜尋，取代類神經網路的倒傳遞計算，以MSE(誤差均方根)全域極小值為目標，求解得到模糊類神經網路的最佳權值，改善了類神經網路MSE陷入局部最小值的缺點，因而獲得準確的診斷結果。
本文以監測機械運轉狀況之信號量測系統得到振動位移或加速度，以時域分析、振動軌跡分析、頻譜分析、頻瀑分析、全息分析等信號分析中提取振動信號特徵，用來作為故障診斷專家系統之輸入；這些振動特徵與故障類型的推論，根據前人研究及專家知識結合而成的關聯矩陣，訓練成模糊類神經網路的專家系統，但是仍有許多未知之故障及其振動特徵以及修正的過去知識，因此，本文以自回歸分析振動信號作為新增故障或修正關聯向量之診斷知識。
對於機器故障運轉的振動信號特徵不明確時，或是針對新機器，以FFT得到的頻譜為準，將正常運轉中的機器，取得振動信號的時間歷程，以足夠的階數作時間序列分析，得到的AR係數，求根及功率比得到功率比AR譜，得到正常狀況之AR譜，訓練成模糊類神經網路之正常運轉關聯向量。該機器經久使用發生故障時，故障類型與其信號特徵，是否包含於專家診斷系統之關聯矩陣中雖無法預知，將此機械正常的功率比AR譜之AR係數存成關聯向量，故障之振動信號與原始正常狀態之AR係數經此模糊類神經網路診斷結果，立可判斷是否正常，再以人為判斷故障類型，進一步改成頻譜的故障關聯向量，增加為診斷專家系統之一行。
足夠的階數才能得到正確的AR係數，本文以FFT的頻譜為準，來決定AR係數之階數，所需階數非常大，隨頻譜複雜度而增加，因此，運算時間非常冗長，而且AR譜的精度以FFT頻譜的精度為極限，無法超過FFT精度，因此，在頻譜關聯矩陣中已知的故障類型，本文認為不需要使用AR係數作為故障診斷的必要或充分步驟，而僅作為故障關聯矩陣之輔助及補充之用。
本文使用P積分觀察法作為旋轉機械故障診斷的方法，此方法是以轉子運動軌跡與原點之間的距離作為積分函數，利用積分區間的改變與P積分值的關係，對旋轉機械運轉時所產生的非線性系統響應進行判斷。
對於監督式類神經網路是利用學習訓練樣本的方式，將輸入及輸出樣本之關係儲存於網路權值之中；診斷時，將待測樣本進行正向計算即可獲得診斷結果。此缺點在訓練樣本不足的情況下或測試樣本不再訓練樣本範圍內時，容易造成錯誤的診斷結果。本文應用時域振動信號之統計參數並結合Bayesian網路作為齒輪故障診斷，並與模糊類神經網路及機率類神經網路方法相互比較，探討在訓練樣本不足或無故障樣本的情況下，Bayesian網路診斷結果之優越性。

關鍵字

基因演算法 ； 機率類神經網路 ； 關聯向量 ； 故障診斷 ； 模糊類神經網路 ； 關聯矩陣 ； P積分觀察法 ； Bayesian網路 ； 自回歸係數

並列摘要

ABSTRACT

The fuzzy neural networks is a modern method of the intelligence fault diagnosis for mechatronic equipments. The relationships between faults and symptoms in frequency spectrum were defined by expert’s knowledge and experiment results. The amplitude of rotary frequency was used for normalization and the level of signal symptom ratio was described by using fuzzy membership functions. The relationships between faults and vibration symptoms were inferred by using If-then rules or neural networks. Due to the type of faults are numerous, the structure of fuzzy neural networks is large and leads to spending more training time. The new faults and symptoms were added, the modified diagnosis rules were retraining.
This study proposed the relation vector which is the relationship between single fault and symptoms in frequency spectrum. The whole relation matrix can be decomposed to several independent relation vectors and trained and computed independently, and did not influenced by whole knowledge increase and correction. Due to the whole knowledge use amplitude of rotary frequency for normalization, the diagnosis results display output matrix. Each column is output of relation vector and the element of column vector is output of faults. The row of all column vectors composed the certainty row vector of diagnosis results for relation faults.
Because the error fell into local minimum easily in training of neural network, in this study, using genetic algorithm for globe search and replacing the backward computation of neural network. On the basis of MSE(mean square error) is globe minimum, solving the optimal weightings of neural network and improving the shortcomings of MSE of neural network fell into local minimum and obtaining the superior diagnosis results.
The vibration displacement or acceleration obtained from signals measurement system in mechanical monitoring. The symptoms in vibration signals extracted as the inputs of fault diagnosis expert system by using time domain analysis, orbit analysis, frequency spectrum analysis, waterfall spectrum analysis, whole spectrum analysis and so on. The inference of symptoms and faults based on the relationships matrix which composed of research and expert knowledge, the expert system was trained by using fuzzy neural networks. Due to the more unknown faults, symptoms in vibration signals and knowledge, in this study, the vibration signals analyzed by using auto regression and been the diagnosis knowledge of new faults or modified the relation vectors.
The symptoms in vibration signals are not definite or mew machine, the frequency spectrum is correct by using FFT. The vibration signals in time domain of normal machine, the time series analysis by enough order and obtain the AR coefficients. The normal spectrum of AR obtained with roots and power ratio, and trained to be relation vector of normal operation by using fuzzy neural networks. Although it does not known the relation matrix of expert diagnosis system includes the faults and symptoms when machine used for a long time and faults happened, the AR coefficient in normal machine save in relation vectors. The AR coefficient of vibration signals between fault and normal conditions can obtained diagnosis results by using fuzzy neural networks and determined machine is normal or not immediately. The type of fault was recognized by experts and correct as frequency spectrum of relation vector and add to the new row in diagnosis expert system.
The correct AR coefficient obtained with enough order, in this study, based on the frequency spectrum of FFT to decide the order of AR coefficient. Because the order is large and complication of frequency spectrum increased, the computation time need long. And the precision of AR spectrum limit within the precision of FFT frequency spectrum and can not exceed. Thus, in this study, it does not necessary for using AR coefficient to fault diagnosis in known frequency spectrum relation matrix, and used to assistance and supplement in fault relation matrix.
This study use P integration observation method to diagnose rotor systems using vibration responses. An integral is defined by integrating the distance of trajectories and origin in phase plane. The responses of a nonlinear system for rotating machinery are used to identify the relationship between the P integration value and the integrated interval.
The relation between input and output samples save in network weightings by training samples in supervise neural network. The diagnosis results by forward computation with test sample. When the training samples of neural networks are not enough, it is easy to get the wrong diagnosis results. This study utilizes the statistical parameters of vibration signals in time domain and combines the Bayesian networks in gear fault diagnosis and compared with fuzzy neural networks and probability neural networks. Also, the superior diagnosis ability investigated with Bayesian networks when training samples are not enough or have not fault samples.

並列關鍵字

Fault diagnosis ； Probability neural networks ； Autoregressive coefficients ； Bayesian networks ； Relation matrix ； P integration observation method ； Relation vector ； Fuzzy neural network ； Genetic algorithm

參考文獻

[1] Dyer, D. and Stewart, R. M., “Detection of Rolling Element Bearing Damage by Statistical Vibration Analysis,” Journal of Mechanical Design, Vol. 100, pp. 229-235, 1978.

[2] Heng, R. B. W. and Nor, M. J. M., “Statistical Analysis of Sound and Vibration Signals for Monitoring Rolling Element Bearing Condition,” Applied Acoustics, Vol. 53, No. 1-3, pp. 211-226, 1998.

[5] Randall, R. B., “A New Method of Modeling Gear Faults,” ASME J. Mech. Des., Vol. 104, No. 2, pp. 259-267, 1982.

[7] Zhongming, Y. E. and Bin, W. U., “A Review on Induction Motor Online Fault Diagnosis,” Power Electronics and Motion Control Conference, Vol. 3, pp. 1353-1358, 2000.

[8] Filippetti, F., Franceschini, G., Tassoni, C. and Vas, P., “Recent Developments of Induction Motor Drives Fault Diagnosis Using AI Technique,” IEEE Transactions on Industrial Electronics, Vol. 47, No. 5, pp. 994-1004, 2000.