噪音訊號診斷模式在風力發電機葉片表層損傷之應用

Application of Noise Signal Diagnostic Mode for Surface Damage Detection in Wind Turbine Blade

湯耀期(Yao-Chi Tang)

指導教授 ： 王昭男

國立臺灣大學/工學院/工程科學及海洋工程學研究所/博士(2017年)

https://doi.org/10.6342/NTU201704460

摘要

本論文的研究重點在於能夠建立兩種風力發電機葉片損傷的診斷系統，近年來風力發電機組如雨後春筍般地陸續建置，因此對於高單價風力發電機組的損壞預警機制能力的提昇則更為重要，尤其是風力發電機葉片的單價更是佔機組價格極高的比例。目前台灣風力發電機大多分佈在沿海周邊，相對地容易受到海水的侵蝕、飛砂的攻擊及長時間暴露在日曬雨淋的嚴苛環境之中，若無法提供有效率的損傷診斷模式，將大大的降低風力發電機組的使用壽命及金錢上的損失。因此本文的診斷模式，將以實際風機運轉時量測下的結果，進行預警系統的建立，且在量測風機葉片時，不需將風機停機，可以在持續運轉的模式下進行，以貼近更實際運行的診斷狀況。
本文為提供最佳的診斷方式，將以噪音訊號進行時頻(Time-Frequency)域及時域(Time-Domain)分析技術為基礎，提供兩種噪音訊號特徵診斷的預估模式。對於時頻分析將使用短時傅立葉轉換為訊號分析基礎，對訊號進行分析進而得到時頻分析的結果，轉換成邊際頻譜，並利用迴歸分析建立正常標準風機基準，再與其他受測風機比較後，判斷是否異常。為得到較科學且客觀的判斷基準，本文利用接收者操作特徵曲線，界定最佳損傷閾值作為判斷的依據，提供特徵抽取診斷的第一種模式。
另一項方法則是直接使用時間域的訊號作為診斷的模式，首先以碎形理論為基礎，並將時域訊號經計算後得到碎形維度(Fractal Dimensions)的計算結果，並將結果使用統計方式轉換成高斯分佈(Gauss Distribution)，比較正常基準風機與受測風機兩者高斯分佈曲線下的面積交集，以提供葉片損傷時異常的診斷。同時亦利用標準正常風機使用高斯面積自我學習方式得到面積交集的最佳損傷重疊率作為異常與否的診斷依據。並以彰濱工業區實際風機運轉時之狀況，使用兩套分析模式來診斷，並已獲得良好的診斷結果，對於風機葉片是否損傷的預警建立已具有相當的成效。

關鍵字

短時傅立葉轉換 ； 接收者操作特徵曲線 ； 最佳損傷閾值 ； 碎形維度 ； 高斯分佈

並列摘要

This dissertation centered on the establishment of the two diagnostic models of damaged wind turbine blades. Wind turbine generators in recent years have mushroomed; the improvement in the early warning of damaged wind turbine blades is, therefore, of great significance, in view of premium prices of wind turbine blades. Currently, wind turbine generators in Taiwan are mostly located along the coasts, where the erosion of seawater and sand and the long-term exposure to the harsh environment increase overall wear and tear of wind turbine blades. If effective diagnostic models of damaged wind turbine blades are not available, the operating life of wind turbine generators will highly be shortened and the loss of money will be great. Therefore, this study proposed two diagnostic models of early warning of damaged wind turbine blades based on the results measured as wind turbine blades are in operation, which further ensured the diagnoses would reflect the actual operation of wind turbine blades.
This study, based on Time-Frequency analysis and Time-Domain analysis, proposed two estimation models of noise signal characteristic diagnoses. The first model is on the basis of Time-Frequency analysis. Noise signals of normal wind turbine blades were first analyzed based on STFT and then the results of Time-Frequency analysis were converted into marginal spectrums. Regression analysis was next adopted to analyze the marginal spectrums; the obtained results thus served as the criterion to judge whether a given wind turbine blade is abnormal. To ensure the criterion is valid, this study utilized Receiver Operating Characteristic Curve and Area Under Curve to determine the optimal threshold for feature extraction diagnosis.
The second estimation model is on the basis of Time-Domain analysis. Firstly, the time domain of a normal wind turbine blade was measured and then calculated based on the fractal dimension theory. The results of fractal dimensions were next converted into Gauss Distribution. The overlaps of the Gauss Distribution of a normal wind turbine blade and that of a given one served as an indicator to determine whether the given wind turbine blade is abnormal. Further, the overlaps of the Gauss distribution of various intervals of a normal wind turbine blade were utilized to determine the optimal threshold for the anomaly of wind turbine blades. The two estimation models were applied to examine whether the wind turbine blades in operation in Changhua Coastal Industrial Park were damaged and the models were proven effective.

並列關鍵字

Short Time Fourier Transform ； Receiver Operating Characteristic Curve ； Optimal Threshold ； Fractal Dimension ； Gauss Distribution

參考文獻

[6]. C. C. Lee and B. W. Tsai, "The Relationship between the Spatial Pattern of Built-up Area and Landform-- An Application of Fractal Theory," Journal of Taiwan Geographic Information Science(3): pp.43-55, 2005.

[16]. 劉大均, "運用經驗模態分解法與包絡線分析於單/雙損傷軸承故障診斷," 國立中興大學機械工程學系所, 2015.

[32]. 秦正宇, "小波轉換於風力發電機葉片診斷之應用," 國立台灣大學工程科學及海洋工程學研究所, 2016.

[40]. 陳文哲, "以呼吸聲頻之數位訊號作哮喘病徵之辨識," 國立台灣大學工程科學及海洋工程學研究所, 2011.

[57]. C. N. Wang and Y. C. Tang, "Application of Morlet wavelet in Diagnosis of Noise Signal Characteristics of Wind Turbine Blade in Abnormal Operation", Journal of Marine Science and Technology, Vol. 25, No. 1, pp. 62-69 ,2017.