軸承為機械工業中不可或缺的零部件，其剩餘使用壽命(Remaining Useful Life， RUL)預測是受到重點關注的研究議題。本研究旨在模擬解決實務上軸承數據有限，且訓練軸承與待預測軸承故障特徵不同之困境，並將提出的預測方法以NASA IMS與IEEE PHM 2012公開軸承數據集進行驗證與評估。本論文分為四部分–(1)實驗數據與訊號分析：說明兩組數據集之實驗平台與資料擷取方式，並針對NASA IMS軸承數據進行詳細的故障特徵分析，藉由倒頻譜預白化(Cepstrum Pre-Whitening， CPW)與平方包絡譜(Squared Envelope Spectrum， SES)評估各軸承之故障狀態，以作為軸承壽命重新評估之依據；(2)分段線性健康指標(Health Indicator， HI)建立：以三倍標準差定律設定第一預測點(First Predicting Time， FPT)，將個別軸承之HI定義於[1， 0]範圍內，使不同故障狀態軸承統一壽命終點(EoL)之模型訓練標籤，健康狀態標籤定義為 RUL=1，壽命終點則定義為 RUL=0，以此建立各軸承之分段線性HI；(3)特徵提取與篩選：每顆軸承提取19項幅域特徵，以相關系數與方差擴張因子(Variance Inflation Factor， VIF)進行特徵篩選，剔除相關係數小於0.4、大於0.99且VIF大於10之特徵，簡化模型複雜度並提高準確率，並保留全部軸承篩選特徵之聯集；(4)極限梯度提升(eXtreme Gradient Boosting， XGBoost)模型建立與訓練：以留一驗證(Leave-One-Out Cross Validation， LOOCV)與網格搜索的方式進行訓練，確保模型在避免數據洩漏前提下找到最佳超參數。本研究也計算兩組數據於不同失效閾值(Failure Threshold， FT)之相對誤差百分比，以評估預測方法於實際軸承維護之可行性；NASA IMS四顆軸承於失效閾值設為0.2、0.1與0.05時，其相對誤差百分比分別落於11.85%至-5.59%、3.75%至-3.63%，以及4.49%至-1.25%，而IEEE PHM 2012七顆軸承之相對誤差百分比則分別落於19.84%至-5.40%、18.64%至-3.99%，以及19.19%至1.11%；從預測結果可得知，本研究提出之預測方法具備軸承健康狀態之追蹤能力，且兩組數據皆存在失效閾值(FT)越低，則相對誤差百分比越小之趨勢，顯示模型於失效期間具備更高的準確率，可在軸承失效前準確提出警報，提前進行維護作業。;Rolling bearings are indispensable components in mechanical industries， and accurate Remaining Useful Life (RUL) prediction has become a critical research topic. This study aims to address practical challenges where bearing data are limited and the fault characteristics between training bearings and target bearings are inconsistent. The proposed prediction framework is validated and evaluated using the NASA IMS and IEEE PHM 2012 public bearing datasets. This thesis is divided into four parts–(1) Experimental data and signal analysis: The experimental platforms and data acquisition procedures of both datasets are introduced. Detailed fault signature analysis is conducted on the NASA IMS dataset， where Cepstrum Pre-Whitening (CPW) and Squared Envelope Spectrum (SES) are employed to evaluate bearing fault conditions and reassess bearing lifetime; (2) Piecewise linear Health Indicator (HI) construction: The First Predicting Time (FPT) is determined using the three-sigma rule. The HI of each bearing is normalized within the range of [1， 0]， enabling unified End-of-Life (EoL) labeling across bearings with different fault modes. The healthy state is defined as RUL = 1， while the EoL is defined as RUL = 0， thereby constructing a piecewise linear degradation trajectory for each bearing; (3) Feature extraction and selection: Nineteen amplitude-domain features are extracted from each bearing. Feature selection is performed using correlation analysis and the Variance Inflation Factor (VIF)， where features with correlation coefficients below 0.4， above 0.99， and VIF greater than 10 are removed. This process reduces model complexity and improves prediction accuracy， while the union of selected features across all bearings is retained; (4) eXtreme Gradient Boosting (XGBoost) model development and training: Leave-One-Out Cross Validation (LOOCV) combined with grid search is employed to identify optimal hyperparameters while preventing data leakage. To evaluate practical maintenance applicability， relative percentage errors under different Failure Threshold (FT) are calculated. For the NASA IMS dataset， the relative errors of four bearings fall within 11.85% to −5.59%， 3.75% to −3.63%， and 4.49% to −1.25% under FT values of 0.2， 0.1， and 0.05， respectively. For the IEEE PHM 2012 dataset， the relative errors of seven bearings range from 19.84% to −5.40%， 18.64% to −3.99%， and 19.19% to 1.11% under the same FT settings. The results demonstrate that the proposed method effectively tracks bearing degradation. Furthermore， both datasets exhibit a consistent trend in which lower FT levels lead to smaller relative errors， indicating improved prediction accuracy during the failure progression stage. The proposed approach can therefore provide reliable early warnings prior to bearing failure， supporting proactive maintenance decision-making.
