熔池影像特徵與雷射畫線品質之預測研究

Abstract

本研究以雷射銲接加工頭結合六軸工業機械手臂為實驗平台，並搭配工業相機及定焦鏡頭，於銲接過程中即時擷取熔池影像。為系統性探討製程參數對熔池外形之影響，採用全因子實驗法規劃不同功率、雷射加工頭到工件距離及銲接速率之組合，並結合各條件下實際量測之銲道寬度與滲透深度，建立熔池特徵與銲道結果之關聯。 在影像處理流程方面，針對雷射銲接過程中對影像造成之雜訊干擾，運用工業相機搭配 Python 與 OpenCV 函式庫進行逐幀影像處理。經影像標定後，從原始 4096×3000 像素影像中截取包含熔池主體的 224×224 區域，經色彩空間轉換、分群演算法分割、形態學閉運算修補、濾波降噪與閾值二值化後，成功擷取熔池外形特徵並計算有效區域之像素數據，為後續分析奠定基礎。 實驗結果顯示，功率對熔池面積具有最顯著之正相關影響，為決定熔池尺寸的關鍵參數；雷射加工頭到工件距離的作用呈現非單調性，於距離 20 mm 時熔池面積達最大值；銲接速率則與熔池尺寸呈顯著負相關，速率愈快熔池面積愈小。 在銲道結果預測方面，本研究以熔池面積與長寬比為輸入特徵，利用二次多項式回歸分析建立銲道寬度與滲透深度之預測模型。銲道寬度模型具高解釋力（R²=0.871）、低誤差（MAE=0.185mm、MAPE=9.43%），預測表現優良；然而，滲透深度模型的解釋度較低（R²=0.347），即便新增長寬比後僅提升至 R²=0.3905，預測精度改善有限。誤差分析指出，當長寬比高於 2.1 時，模型低估極深熔深；而在中等熔深區間則易出現高估情形。 本研究建立了熔池影像處理流程及建立熔池特徵與銲道結果之關聯，初步驗證了熔池影像特徵於銲接品質預測上的可行性。所提出之方法可作為即時監控與製程參數調整的基礎，未來可透過擴充特徵變數、增加樣本數據及導入進階學習模型，以進一步提升滲透深度之預測能力，邁向智慧化銲接品質監控。

This study employs a laser welding head integrated with a six-axis industrial robot arm as the experimental platform, combined with an industrial camera and a fixed-focus lens to capture melt pool images in real time during welding. A full factorial experimental design was adopted to systematically investigate the effects of process parameters—including laser power, standoff distance, and welding speed—on melt pool geometry. The measured weld bead width and penetration depth under each condition were correlated with the extracted melt pool features. For image processing, an industrial camera with Python and the OpenCV library was utilized to address noise interference caused by intense reflections and molten metal fluctuations during laser welding. After image calibration, a 224×224-pixel region containing the melt pool was cropped from the original 4096×3000-pixel image. Subsequent steps, including color space conversion, clustering-based segmentation, morphological closing, filtering and denoising, and thresholding, enabled accurate extraction of melt pool geometric features and pixel-based measurements. Experimental results indicate that laser power exerts the most significant positive influence on melt pool area, while the effect of standoff distance is non-monotonic, with a maximum melt pool area observed at 20 mm. Welding speed shows a strong negative correlation with melt pool size. Furthermore, second-order polynomial regression models were developed to predict weld bead width and penetration depth based on melt pool area and aspect ratio. The bead width model demonstrated high explanatory power (R² = 0.871) with low prediction error (MAE = 0.185 mm, MAPE = 9.43%), whereas the penetration depth model exhibited limited accuracy (R² = 0.347), improving only slightly after incorporating aspect ratio (R² = 0.3905). Error analysis revealed underestimation in cases of extremely deep penetration (aspect ratio > 2.1) and overestimation in moderate penetration ranges. This research establishes a melt pool image processing workflow and quantifies the relationship between process parameters, melt pool features, and weld quality. The findings demonstrate the feasibility of using melt pool image features for weld quality prediction and provide a foundation for real-time monitoring and process parameter adjustment. Future work will focus on expanding feature sets, increasing data diversity, and applying advanced learning models to enhance penetration depth prediction and enable intelligent welding quality control.