開發新型領域適應方法於深度學習與混合資料庫應用於複合式損傷結構健康監測系統

Abstract

為了降低例如離岸風電等在高度複雜操作環境下的的運營和維護成本並提高在能源穩定性，本研究提出了一種利用深度學習和大數據的力量來自動化和簡化監測過程的方法。本研究嘗試將傳統的結構健康監測技術、信號處理知識和深度學習技術相結合以克服這一挑戰。本研究試圖使用由真實世界數據和電腦模擬數據組成的混合資料庫訓練深度神經網絡（DNN），以解決在取得有限的真實數據下從而導致數據匹配度問題，並且取得良好的監測效果。本研究開發了一種基於卷積的創新技術以實現領域適應，增加模擬數據和真實世界數據之間的相似性，使DNN在有限數據領域中也能有良好的表現。卷積領域展開（CDE）的方法將不同來源的數據映射到新的領域供DNN分類，然後將分類後的數據重新映射回其原始的健康/損壞狀態。本研究應用了分散式的系統設計和多種深度學習技術來識別在無先驗知識下的未知損壞組合，並且與我們的CDE技術結合後有很好的領域適應效果。一個實驗室尺度的離岸風電塔架與水下機基礎被使用於採收數據且分析，最後應用開發的算法進行監測任務。本論文希望對大數據數據的土木結構監測研究提供貢獻，並且期待於未來可以用用於於風力發電機、電塔和電網等，提高能源的安全性和穩定性。

In order to lower the Operation and Maintenance cost and increase energy stability on structures in complex conditions such as offshore wind turbines, this research presents a method which leverages the power of deep learning and big data to automize and simplify the process of monitoring. In this research, an attempt to merge traditional Structural Health Monitoring techniques, Signal Processing knowledge, and Deep Learning technology is used to overcome this challenge. This research attempts to train a DNN with the use of a hybrid database comprised of real-life and computer simulated data to achieve good monitoring results with limited usage of real-life data, which would lead us to a data mismatch problem. A novel convolution-based technique is developed to increase the similarity between the simulated data and real-life results to achieve domain adaption in order for the DNN to perform well on a limited data domain. The Convolution Domain Expansion (CDE) method maps different sources of data into new domains for the DNN to classify, and then re-maps the data back to its original healthy /damaged state. A decentralized design and multiple deep learning techniques are applied to identify unknown damage combinations without prior knowledge and is proven to have great results combined with our CDE technique. A lab scale structure of a foundation of an Offshore Wind Turbines is used for modelling and a monitoring task is performed with the developed algorithm. This paper hopes to contribute to data-centric civil structure monitoring research that could potentially be used on OWTs, electric towers and grids, and many more and help increasing the safety and stability of energy.