| 摘要: | 本論文針對塑膠齒輪進行瑕疵檢測，感測器包含將光纖光柵感測器(Fiber Bragg
Grating, FBG)、加速規與角度編碼器，其中光纖光柵感測器直接黏貼於塑膠齒輪齒根處，
加速規至於軸承座上，角度編碼器則設置於主動軸以及被動軸上。齒輪包含正常、磨損
與裂痕等三種狀態。
為有效獲取齒輪訊號中的特徵資訊，本論文採用兩種訊號處理技術進行前處理：一
為完全整體經驗模態分解（Complementary Ensemble Empirical Mode Decomposition,
CEEMD），將原始訊號分解為多個具有物理意義的本質模態函數（IMFs），藉此保留有
效成分；另一為同步平均法（Synchronous Averaging），透過多週期重疊平均以提升週期
性特徵訊號的訊雜比（signal-to-noise ratio）。之後根據處理後的訊號計算多項狀況指標
（Condition Indicators, CIs），如 RMS、峭度、偏度等，作為後續辨識模型之輸入特徵。
在辨識階段，分別採用支持向量機（Support Vector Machine, SVM）與隨機森林
（Random Forest）兩種模型進行分類訓練與測試，以建立齒輪健康狀況辨識模型，並比
較其準確率。為驗證 FBG 感測應用之可行性與效能，研究同時引入傳統的加速度感測
器與角度編碼器作為對照組，比較三種感測器之間的差異。實驗結果顯示，FBG在利用
CEEMD分解後以隨機森林模型辨識的準確率高達99%以上。FBG不僅能穩定擷取齒輪
磨損與裂痕訊號，且在多數分類場景中優於傳統感測器，證實其於健康監測領域的潛力。;This study focuses on defect detection of plastic gear by directly bonding a Fiber Bragg
Grating (FBG) sensor to the root of the gear teeth. The sensing system includes three types of
sensors: FBG sensor, accelerometer, and encoder. While the FBG sensor is directly attached to
the root of the plastic gear teeth, the accelerometer is mounted on the bearing housing, and the
encoders are installed on both the driving and driven shafts. Gear conditions under investigation
include three states: normal, wear, and crack.
To effectively extract characteristic information from the gear signals, two signal
processing techniques are used for preprocessing. The first is Complementary Ensemble
Empirical Mode Decomposition (CEEMD), which decomposes the original signal into several
Intrinsic Mode Functions (IMFs), thereby preserving essential signal components. The second
technique is Synchronous Averaging, which enhances the signal-to-noise ratio of periodic
features through multi-cycle overlapped averaging.
Subsequently, several condition indicators (CIs), such as Root Mean Square (RMS),
kurtosis, and skewness, are calculated based on the preprocessed signals and used as input
features for the classification models. In the classification stage, Support Vector Machine (SVM)
and Random Forest models are used to train and test classification performance, thereby
establishing models for gear health condition recognition and comparing their accuracy. To
validate the feasibility and effectiveness of FBG-based sensing, accelerometer and encoder are
also introduced as reference sensors. Experimental results show that FBG with CEEMD and
Random Forest, can achieve over 99% accuracy. the FBG sensor not only reliably captures gear
wear and crack signals but also performs better than traditional sensors in most condition,
confirming its potential for applications in condition monitoring. |
