基於數值分析與深度學習之感測器佈設最佳化與結構健康監測方法建立

Optimized Sensor Placement and Structural Health Monitoring Method base on Numerical Analysis and Deep Learning

陳楚云(Chu-Yun Chen)

指導教授 ： 黃心豪

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202203189

摘要

近年來的船舶海損事故，一艘滿載2,200萬加崙燃料油的油輪，行駛於西班牙外海時不幸發生船難，船體結構呈現舯垂現象，最終裂成兩半後沉沒，船上的燃油全部外漏，造成極為嚴重的環境汙染。一艘貨櫃輪航行至距葉門外海200英里處時，結構從船舯處突然折斷成兩截，船齡僅僅只有五年，損失超過2.5億美元，綜合以上兩個海損事故案例可知，結構健康監測在船舶產業領域上扮演了很重要的角色。因此，本文針對一艘大型貨櫃船感測器佈設最佳化及結構健康監測方法之建立進行研究。利用法國驗船協會之水動力分析工具，計算不同船速、航向角及角頻率下船體殼元素的應力反應振幅運算子（Stress Response Amplitude Operator, Stress RAO），經由26個北大西洋海況的短期響應分析（Short-term Analysis）及5個等效規則波設計（Equivalent Design Wave, EDW）之計算結果統計出大型貨櫃船感測器佈設最佳化的位置，將感測器佈設最佳化位置上之殼元素取出，針對不同的波高進行短期響應分析，本文建立6種不同波高大小的資料集，根據應力訊號之特性設計出不同的深度學習模型架構，本研究共提出五種不同的深度學習模型架構：RNN、LSTM、GRU、CNN結合LSTM及CNN結合GRU，觀察不同深度學習模型架構對多維度時間預測之影響，在北大西洋海況條件中表現最佳的模型為LSTM模型，資料集平均均方根誤差為0.008，並利用LSTM模型架構建立出一套針對不同海況條件的結構健康監測方法。

關鍵字

短期響應分析 ； 等效規則波設計分析 ； 多維度時間序列預測 ； 感測器佈設最佳化 ； 結構健康監測

並列摘要

In recent years of ship accidents, an oil tanker loaded with 22 million gallons of fuel oil, unfortunately suffered a shipwreck. Finally, it split in half and sank in the sea. All the fuel oil are leaked out, causing extremely serious environment problem. The container ship was cracked at the midship in harsh marine environment and broken into two pieces. The ship was only five years old, and lost more than 250 million US dollars. Above two accidents show that structural health monitoring play an important role in shipbuilding industry. Therefore, this paper is study optimized sensor placement and establishment of structural health monitoring method for large container ship. The stress response amplitude operator of shell elements under different ship speeds, heading angles and frequencies are calculated by using hydrodynamic analysis tool. Through results of short-term response analysis with 26 sea conditions and 5 equivalent regular wave design, the optimalized sensor position of large container ship is calculated, and shell elements at the optimalized sensor position are taken out. For short-term response analysis with different wave heights, this paper is established 6 data sets with different wave heights, and design different deep learning architectures according to the characteristics of stress signals. This paper is proposed five different deep learning architectures: RNN, LSTM, GRU, CNN combined with LSTM and CNN combined with GRU. Observe the influence of different deep learning architectures on multi-dimensional time series forecasting. The model with best performance in the North Atlantic sea conditions is LSTM model, the average root mean square error in each dataset is 0.008, and LSTM model architecture is used to establish structural health monitoring for different sea conditions.

並列關鍵字

Short-term Response Analysis ； Equivalent Regular Wave Design ； Multi-dimensional Time Series Forecasting ； Optimized Sensor Placement ； Structural Health Monitoring

參考文獻

[1] 社團法人台灣環境資訊協會. "油輪失事漏油."

[2] "郁志榮：日本工匠精神失傳 建造貨船誰人敢用."

[3] H. S. Chan, "A three-dimensional technique for predicting first-and second-order hydrodynamic forces on a marine vehicle advancing in waves," University of Glasgow, 1990.

[4] S. Malenica, F. o.-X. Sireta, F. Bigot, C. Wang, and X.-B. Chen, "Some aspects of hydro-structure coupling for combined action of seakeeping and sloshing," in International Conference on Offshore Mechanics and Arctic Engineering, 2009, vol. 43468, pp. 461-467.

[5] S. Malenica, E. Stumpf, F. o.-X. Sireta, and X.-B. Chen, "Consistent hydro-structure interface for evaluation of global structural responses in linear seakeeping," in International Conference on Offshore Mechanics and Arctic Engineering, 2008, vol. 48234, pp. 61-68.