- 學位論文
基於企業職級演算法及灰狼演算法之軸承故障診斷
Bearing Fault Detection Using the Grey Wolf Algorithm and Heap-based Optimizer
摘要
本研究提出基於特徵選取(feature selection)方法應用於馬達軸承故障診斷模型，此模型共可分為三個階段，分別為特徵擷取(feature extraction)、特徵選取及分類器(classifier)。在特徵擷取階段，本研究結合多重解析度分析(multiresolution analysis, MRA)及快速傅立葉轉換(fast Fourier transform, FFT)分析馬達軸承原始訊號並從中擷取特徵。在特徵選取階段，本研究所提出之特徵選取方法企業職級二元灰狼演算法(heap-based binary grey wolf optimization, HBBGWO)選取由特徵擷取階段所提取之特徵，此方法結合灰狼演算法(grey wolf optimizer)及企業職級演算法(heap-based optimizer, HBO)，藉由刪除不必要特徵可提升模型辨識率(accuracy)並減少模型運算成本(computation cost)，因此提升模型效率。在分類器階段，本研究採用支撐向量機(support vector machine, SVM)及線性鑑別分析(linear discriminant analysis, LDA)作為模型分類器，並比較兩者之辨識率。為獨立驗證所提出特徵選取方法之效率，本研究以加州大學爾灣分校(University of California at Irvine, UCI)所提供之公開數據集對所提出之特徵選取方法進行驗證。本研究另採用三種馬達原始訊號驗證模型效率，分別為實驗室量測之故障感應電動機電流訊號、凱斯西儲大學(Case Western Reserve University, CWRU)所提供之馬達震動訊號公開數據集及Machinery Failure Prevention Technology (MFPT)所提供之馬達震動訊號公開數據集。
並列摘要
In this study, a feature selection approach applied in motor bearing fault diagnosis model was proposed. The model can be separated into three stages, feature extraction, feature selection and classifier. In the stage of feature extraction, multiresolution analysis (MRA) and fast Fourier transform (FFT) were adopted to extract features from motor raw signal. In the stage of feature selection, a proposed feature selection approach heap-based binary grey wolf optimization (HBBGWO) was applied, which hybrid grey wolf optimizer (GWO) and heap-based optimizer (HBO). To increase the classification accuracy and reduce the computation time, the feature selection approach can eliminate the redundant features, which lead to the increasing the model efficiency. In the stage of classifier, support vector machine (SVM) and linear discriminant analysis (LDA) were utilized, the accuracy of two methods were also be compared. On the other hand, three different kinds of motor raw signal were adopted to validate the diagnosis model, respectively Case Western Reserve University (CWRU) benchmark motor vibration dataset, Machinery Failure Prevention Technology (MFPT) benchmark motor vibration dataset and motor current signal from laboratory. To verify the proposed feature selection approach independently, University of California at Irvine (UCI) benchmark dataset is also applied in this study.