應用聲音訊號與深度學習進行纜索張力估算與腐蝕辨識

Abstract

本研究探討以麥克風量測振動纜索之聲壓訊號，應用於纜索張力估算與腐蝕狀態辨識之可行性。在張力估算方面，使用弦振動理論與雙振頻法進行反算，實驗結果顯示於各種結構條件下，其張力反算誤差普遍低於 3 %，驗證本方法具備良好的準確性與穩定性。後續進行整體與局部腐蝕模擬試驗，探討不同腐蝕條件下頻譜特徵之變化，並以多層感知器模型輸入頻譜特徵進行訓練與預測，驗證集與測試集的預測結果分別達到 R^2 為 0.9577 與 0.9387，顯示本方法在纜索腐蝕辨識上的發展潛力。

This thesis explores the feasibility of using a microphone to measure sound pressure signals from vibrating cables for cable tension estimation and corrosion identification. For tension estimation, string vibration theory and the dual-frequency method were employed, resulting in estimation errors generally below 3% under various structural conditions, verifying the accuracy and stability of the proposed method. Corrosion tests were conducted to examine variations in vibration spectral features under different corrosion conditions. The extracted features were used to train a multilayer perceptron model, which achieved R^2 values of 0.9577 and 0.9387 on the validation and test sets, demonstrating the potential of this method for cable corrosion identification.