為驗證所提系統之效益，本研究設計兩項實驗。第一項實驗比較本系統與人工串接流程在新樣態瑕疵適應流程上的效率差異。結果顯示，在相同輸入資料、任務流程與驗收條件下，本系統可將人工介入次數降低約 85%，人工操作時間降低約 99%，且整體流程總耗時降低約 40%。第二項實驗驗證生成經驗記憶機制對生成流程效率與穩定性之影響。結果顯示，在相同生成模型、輸入條件與迭代上限下，未啟用記憶機制之流程未能於迭代上限內通過品質驗證；相較之下，具備記憶機制之流程於三次重複實驗中皆能於迭代上限內達標，且平均約 3.3 輪迭代即可通過品質驗證。綜合實驗結果可知，本研究所提出之系統能降低 PCB 新樣態瑕疵適應流程中的人工操作負擔與重複調整成本，並提升流程自動化程度與生成迭代穩定性。;In PCB manufacturing lines， when new defect patterns emerge， only a small number of real samples can usually be obtained. Nevertheless， the adaptation process still requires multiple stages， including defect preprocessing， data augmentation， generation quality verification， and classification model updating. Since these stages are often distributed across different tools and independent procedures， users must repeatedly connect data across stages， convert formats， configure parameters， and inspect results. This not only increases the manual burden required to adapt to new defect patterns， but also makes the overall workflow difficult to execute in a stable and reproducible manner. In addition， defect image generation under limited-sample conditions usually requires multiple rounds of adjustment and verification. Without preserving previous generation results and refinement experience， the system may repeatedly attempt generation conditions with limited effectiveness， thereby increasing the trial-and-error cost of the generation process.
This study proposes a closed-loop adaptation system for new PCB defect patterns built around a multi-agent system， integrating defect preprocessing， defect generation and verification， and classification model training into an end-to-end workflow. Through role specialization and task scheduling among different agents， the system automatically performs most workflow coordination and processing tasks， including defect region extraction， defect description generation， synthetic image generation， generation quality verification， and classification model updating. Meanwhile， user confirmation and correction mechanisms are retained at important stages to reduce the risk of erroneous results being propagated to subsequent processes. To address the problem that generative models require repeated adjustment under limited-sample conditions， this study further introduces a dual generation quality verification mechanism and a generation experience memory mechanism. These mechanisms allow the system to preserve parameter settings， verification results， and refinement experience from previous generation and verification processes， and to use them as references for subsequent defect generation tasks.
To evaluate the effectiveness of the proposed system， two experiments were conducted. The first experiment compared the efficiency of the proposed system with that of a manually connected workflow in the new defect adaptation process. The results show that， under the same input data， task workflow， and acceptance criteria， the proposed system reduced the number of manual interventions by approximately 85%， reduced manual operation time by approximately 99%， and reduced the total workflow time by approximately 40%. The second experiment evaluated the effect of the generation experience memory mechanism on the efficiency and stability of the generation process. The results show that， under the same generative model， input conditions， and iteration limit， the workflow without the memory mechanism failed to pass quality verification within the iteration limit. In contrast， the workflow with the memory mechanism successfully satisfied the quality criteria within the iteration limit in all three repeated experiments， requiring an average of approximately 3.3 iterations to pass quality verification. Overall， the experimental results indicate that the proposed system can reduce the manual operation burden and repeated adjustment cost in the new PCB defect adaptation process， while improving workflow automation and the stability of the iterative generation process.
