支持向量機偵測風機葉片表層損傷之研究

Abstract

風力發電機價格成本極為昂貴，為降低葉片因環境影響造成故障之維修費用，故建立一套系統來判斷風機葉片表面是否為損壞，及早對葉片表層進行修補，避免因結構破壞而須更換新葉片，造成鉅額費用支出。本實驗收集陸地風機運轉所產生之噪音訊號，並將現場測得之不同風速範圍從4 m/s至10 m/s建立同一套判別標準，量測組數共817組風機運轉噪音，進行梅爾倒頻譜(Mel-Frequency Cepstral)訊號處理，擷取該訊號之梅爾倒頻譜係數(Mel-Frequency Cepstral Coefficients, MFCC)，以此係數及其微分等運算作為葉片之特徵訊號，每一組運轉噪音皆得156個特徵維度；接著，再透過支持向量機(Support Vector Machine, SVM)，建構偵測並診斷風機葉片表層損傷與否之分類訓練模組，並將所有MFCC特徵係數隨機抽取75%進行SVM模型訓練，剩餘25%樣本為測試資料，希望藉由此機器學習分析方式，提高判斷葉片為正常或損壞之準確率。 透過時頻圖觀察風機運轉之噪音，發現於頻率3100 Hz左右以下之能量相對較大，大多參雜發電機及風機運轉之低頻噪音，為避免該噪音對判斷結果造成影響，可利用高頻段(3100 Hz~12800 Hz)噪音訊號進行分析。經研究顯示，透過MFCC及SVM方法計算，得到葉片運轉之高頻訊號診斷結果準確率為92.3%；另外，也將風機運轉噪音以全頻段(~12800HZ)訊號進行偵測並診斷葉片損傷資訊，該判斷分類結果準確率為96.8%，較高頻訊號診斷結果準確。由此可知，判斷風機葉片表層損壞與否之方法，以全頻段訊號處理分析較符合實際應用，方能得到較高之診斷準確率。為減少運算量及維度資料儲存量，本研究針對全頻段訊號進行主成分分析法(Principal Components Analysis, PCA)計算，使原本156個維度係數刪去冗餘噪音資料，同時保留貢獻度較大之特徵作為主要訓練模型之數據，最終以12個維度係數進行SVM計算，得到診斷準確率與原本高維度(156維)相比，僅些微之差，可有效作為風機葉片表面損傷之診斷。

Due to the climate change in recent years, the development of renewable energy projects has widely been promoted worldwide to reduce greenhouse gas. Despite of the contaminated environment, people tended to build the variety of renewable power systems produced by wind, sunlight, rain and other approaches which could be easily obtained from nature. The wind turbine is one of the most popular power supplies while this generally causes defects on the surface of blades based on strong wind or harsh weather besides the ocean. Thus, in order to avoid the expensive maintenance fee and prolong wind turbine’s life, an aim of this research was to build a detected system to examine whether the condition on blade surface was damaged or not. By recording the sound of wind turbines on land is the main data sources since offshore wind turbine is difficult to obtain. There are 817 data including different wind speed from 4 m/s to 10 m/s. In this paper, we have proposed three approaches. First, using Mel-Frequency Cepstral Coefficient (MFCC) to extract feature from the signal noise of recorded sound of wind turbine, and every 817 data will obtain 156 dimensional coefficients. However, in order to reduce the calculation time and remove insignificant data, the second method is Principal Component Analysis (PCA), which will lower the dimensional coefficients. Finally, randomly choose 75% of the feature coefficients as the trained model data while the other 25% are the test data, and using a classification called Support Vector Machine (SVM) to build a model, which will show the accuracy of the classified result of normal or damaged rotor blades. Our research investigated both the whole frequency (~12800 Hz) and the high frequency (3100~12800 Hz) of the signal noise. After MFCC and SVM processing, the accuracy of the whole frequency signal result is 96.8%, which is better than the other one (92.3%). Thus, the whole frequency (~12800 Hz) signal noise could be more useful for the actual application. Except for MFCC and SVM approaches, we applied the PCA methods on the whole frequency (~12800 Hz) signal noise which was successfully lower the amount of 156 dimensional coefficients to only 12 dimensions. Also, the accuracy of that classified result had a slight difference between the original one. Hence, this recognizing system may be helpful to improve the lifetime of wind turbine as well as keeping the cost down from maintenance fee.