小波分析與卷積神經網路於風機齒輪箱振動故障診斷之應用

Vibration-based Fault Diagnosis of Wind Turbine Gearbox Using Wavelet Analysis and Convolution Neural Network

黃柏諭(Po-Yu Huang)

指導教授 ： 宋家驥

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102551

摘要

離岸風力發電近年來逐漸發展為綠能產業中重要的一環，其中變速齒輪箱在整個風力發電機系統中是為最關鍵的組件之一。因風機在長時間運轉下維修不易，變速齒輪箱中的零件一旦損壞，除了維修成本巨大，嚴重甚至可能造成整台風機的崩潰。為了降低維護成本，狀態監測(condition monitoring)與故障診斷的技術也成為當今重要課題。
　本文主要目的在於研究風力發電機的故障分類及診斷，並提出下列方法：(a)小波分析(wavelet analysis)，一種可以得知訊號在不同時間段頻率資訊的時頻分析，以及(b)卷積神經網路(convolution neural network, CNN)，一種常應用於影像分類的神經網路(neural network)，並應用上述兩種理論作為故障診斷的方法，利用振動訊號，針對模擬風機傳動系統的振動測試平台進行實驗。為模擬風機變轉速的情形，實驗僅訓練少數特定轉速，以其他中間轉速進行測試，並分別模擬齒輪箱在健康、齒輪故障、軸承故障和複合故障四種不同模式下的狀態。結果表明，本實驗在振動測試平台5-15 RPM的控制轉速下，能成功判斷出變轉速中四種不同的模式，準確率最高可達到97.91%，可靠度則為97.78%。

關鍵字

風力發電機 ； 變速齒輪箱 ； 小波分析 ； 神經網路 ； 故障分類及診斷

並列摘要

Offshore wind power has gradually developed into an important part of the renewable energy industry in recent years. The gearbox is one of the key components in the wind turbine system. Gearbox failure can cause a wind turbine breakdown, expensive both in subsequent repair and lost output. In order to reduce maintenance costs, condition monitoring and fault diagnosis have been a hot topic nowadays.
　　The aim of this thesis is to solve fault diagnosis problems for wind turbine with variable speed. The proposed methods have been based on: (a) wavelet analysis, a time-frequency analysis which can reveal a certain frequency exists in the temporal domain and (b) convolution neural network, a class of artificial that can be used for image classification. The two methods are tested on test rig to simulate wind turbine drivetrain with vibration signal. In order to simulate the variable speed wind turbine, the experiment only trains a few specific speeds, tests with other intermediate speeds, and considers the state of the gearbox in four different modes: health, gear fault, bearing fault and compound fault. The results show that the proposed methods are able to classify four types of faults with gearbox under the control speed 5-15 RPM, the accuracy rate can reach 97.91%, and the reliability is 97.78%.

並列關鍵字

wind turbine ； gearbox ； wavelet analysis ； neural network ； fault classification and diagnosis

參考文獻

[1] REN21, “Renewable Energy Policy Network for the 21st Century.” Available: https://www.ren21.net/wp-content/uploads/2019/05/gsr_2020_full_report_en.pdf.

[2] 經濟部能源局，〈能源統計資料查詢系統：109年發電結構〉。檢自：www.moeaboe.gov.tw/wesnq/Views/B01/wFrmB0102.aspx?fbclid=IwAR20GtPnHYFIOjFBfy4mzma5_mmthMXNsi70bggn5wqIB5Pm1Dtjm_8LNMg。

[3] 4C Offshore, “Global Offshore Wind Speeds Rankings.” Available: https://www.4coffshore.com/windfarms/windspeeds.aspx.

[4] GWEC, “Global Wind Power Growth Must Triple over Next Decade to Achieve Net Zero.” Available: https://gwec.net/global-wind-power-growth-must-triple-over-next-decade-to-achieve-net-zero.

[5] The Welding Institute, “How Long do Wind Turbines Last?” Available: https://www.twi-global.com/technical-knowledge/faqs/how-long-do-wind-turbines-last.