應用深度學習於微型軸承量測主軸之振動值預測與異常分類研究

NCUIR > Engineering College > Executive Master of Mechanical Engineering > Electronic Thesis & Dissertation > Item 987654321/98624

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98624

Title:	應用深度學習於微型軸承量測主軸之振動值預測與異常分類研究
Authors:	廖國宏;Liao, Kuo-Hung
Contributors:	機械工程學系在職專班
Keywords:	軸承;軸頸軸承;振動檢測;深度學習;機器學習;Bearing;Journal Bearing;Vibration Detection;Deep Learning;Machine Learning
Date:	2025-07-28
Issue Date:	2025-10-17 13:01:18 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究所設計之主軸專為微型軸承與小徑軸承量測應用，透過振動感知器監控軸承運行情況。測試過程中將待測軸承安裝於主軸軸心並施加適當推力載荷，以獲取軸承運轉時之振動數據。該主軸結構類似軸頸軸承，並搭配Turbine Oil No.56 作為潤滑介質。隨著運行時間增加，軸心、銅襯間之磨耗將導致振動值上升，進而影響量測準確性。本研究提出三種建模策略以預測主軸振動值並進行良品與不良品判別。方法一以振動值反向預測配合間隙，因特徵過於單一，模型表現不佳，測試 R2 多為負值。方法二改以幾何尺寸與表面粗糙度等數值特徵預測振動值，表現顯著提升，MLP 模型測試 R2 可達 0.493。方法三進一步結合 ResNet101 擷取之圖像特徵與數值資料，建構多模態模型，整體表現穩定，但高維特徵導致部分模型產生過擬合，第一層預測趨於一致，限制堆疊效果。進一步將圖像特徵降維後，雖抑制過擬合，但預測能力同步下降。分類方面，模型多能穩定辨識良品，但不良品 Recall 偏低，受限於樣本不平衡與邊界樣本比例過高。整體而言，融合圖像與數值特徵有助提升振動預測能力，未來可從特徵選擇、資料擴增與不平衡學習進行優化，以強化缺陷辨識與實務應用效能。;The spindle developed in this study is designed for measuring miniature and smalldiameter bearings. A vibration sensor monitors the bearing’s condition during testing, where the bearing is mounted on the spindle shaft under thrust load to collect vibration data. The spindle resembles a journal bearing and uses Turbine Oil No.56 for lubrication. Over time, wear between the shaft and copper bushing increases vibration, affecting measurement accuracy. Three modeling strategies are proposed to predict spindle vibration and classify product quality. Method 1 uses vibration to inversely predict clearance but performs poorly due to limited input features, yielding mostly negative test R2 values. Method 2 predicts vibration using numerical features like geometry and surface roughness, showing notable improvement with MLP reaching a test R2 of 0.493. Method 3 combines ResNet101-extracted image features with numerical data to build a multimodal model. Performance remains stable, though high-dimensional features lead to overfitting and similar outputs from first-layer models, limiting stacking effectiveness. Dimensionality reduction mitigates overfitting but reduces predictive power. For classification, models reliably identify good samples, but defective recall remains low due to data imbalance and borderline cases. Overall, combining image and numeric features improves vibration prediction. Future work may enhance defect detection through feature selection, data augmentation, and imbalanced learning.
Appears in Collections:	[Executive Master of Mechanical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	6	View/Open

社群 sharing

Loading...