基於生成式 AI 的瑕疵資料生成與銲錫瑕疵檢測方法之研究;Leveraging Generative AI for Defect Data Generation and Solder Defect Detection in PCB Manufacturing

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98628

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98628

Title:	基於生成式 AI 的瑕疵資料生成與銲錫瑕疵檢測方法之研究;Leveraging Generative AI for Defect Data Generation and Solder Defect Detection in PCB Manufacturing
Authors:	賴民淵;Lai, Min-Yuan
Contributors:	資訊工程學系
Keywords:	生成式模型;瑕疵資料生成;擴散模型;焊點瑕疵檢測;印刷電路板;Generative Models;Defect Data Generation;Diffusion Model;Solder Joint Defect Detection;Printed Circuit Board
Date:	2025-08-28
Issue Date:	2025-10-17 13:01:37 (UTC+8)
Publisher:	國立中央大學
Abstract:	自動化瑕疵檢測一直是PCB製造流程中的一項關鍵技術挑戰。相較於傳統的人工作業，自動化檢測能大幅降低人力資源的消耗。然而，在實務應用中訓練瑕疵檢測用的深度學習模型時，常會面臨瑕疵資料量不足的問題。隨著人工智慧技術的發展，許多生成式模型被提出，例如 Stable Diffusion 等模型。然而經實際測試後發現，Stable Diffusion 所生成的資料與真實樣本差異甚大，且難以生成包含瑕疵的樣本。此外，在本研究使用的資料集中，部分不同種類的瑕疵在外觀上極為相似，例如少錫（solder_less）與多錫（solder_more）瑕疵，兩者外觀幾乎無異，僅在錫量上有所差異。對於這兩個瑕疵，較有效的檢測方式是透過 3D-AOI 設備計算錫量體積，然而目前工廠配備的 AOI 設備還是以 2D 為主，若更換成 3D 硬體設備是一筆不小的開銷。而若在2D條件限制下，兩瑕疵差異細微，因此無論是在生成瑕疵資料，或進行瑕疵檢測時，皆是一個困難的挑戰。因此，本研究提出一套針對瑕疵資料不足問題的解決方案，能在資料稀缺的情況下，快速生成具真實感且與實際樣本高度相似的瑕疵影像。將此方案生成的合成資料應用於YOLOv11模型微調(fine-tuning)後，整體性能顯著提升，平均 Precision 提升 0.014（1.5%）、平均 Recall 提升 0.036（4.2%）、mAP@50 提升 0.02（2.1\%）。其中在多錫（solder_more）瑕疵類別上，Precision 提升 0.009（0.9%）、Recall 提升 0.127（15.6\%），而在少錫（solder_less）瑕疵類別上，Precision 雖下降 0.059（6.6%），但 recall 提升 0.033（3.9%）。;Automated defect detection has long been a critical technical challenge in the PCB manufacturing process. Compared to traditional manual inspection, automated methods significantly reduce labor costs. However, in real-world applications, training deep learning models for defect detection often suffers from a lack of sufficient annotated defect data. With the advancement of artificial intelligence, many generative models—such as Stable Diffusion—have been proposed. Nonetheless, empirical results show that images generated by Stable Diffusion differ significantly from real samples and fail to accurately reproduce defect patterns. Moreover, in the dataset used in this study, certain types of defects exhibit highly similar visual characteristics. For example, solder_less and solder_more defects are nearly indistinguishable in appearance, differing only in solder volume. Although 3D-AOI systems can estimate solder volume for more accurate classification, most manufacturing facilities are still equipped with 2D-AOI systems. Replacing these with 3D hardware incurs substantial cost. Under the constraint of 2D inspection, such subtle differences between solder defects present considerable challenges for both data generation and defect detection. Therefore, this study proposes a solution to the problem of insufficient defect data, enabling the rapid generation of realistic defect images that closely resemble actual samples even under data-scarce conditions. When the synthetic data produced by this method is applied to fine-tune the YOLOv11 model, the overall performance improves significantly, with the average precision increasing by 0.014 (1.5%), the average recall increasing by 0.036 (4.2%), and mAP@50 improving by 0.02 (2.1%). In particular, for the solder_more defect category, precision increases by 0.009 (0.9%) and recall improves by 0.127 (15.6%). For the solder_less defect category, although precision decreases by 0.059 (6.6%), recall improves by 0.033 (3.9%).
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

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

社群 sharing

Loading...