| 摘要: | 電子焊接是連接器製造過程中最重要的一環，因為電子焊接會產生瑕疵焊點，直接導致產品功能失效，因此發展預測性維護是重要的。本研究以拉焊(Drag Soldering)工站為智慧化目標，以RJ45連接器為焊接對象，利用生產之參數建構機器學習模型，建立感測系統即時輔助產線人員判斷產品是否通過。
研究分為三個部分，分別是機台狀態評估與失效文獻回顧、設計實做軟硬體整合、實驗結果分析討論。首先對於拉焊製程進行評估，整理焊接瑕疵種類並評估感測器佈置之優先順序，決定以烙鐵失效為優先監測。再來以佇列式訊息處理器(Queued Message Handler)做為軟體系統架構，實做軟硬體模組並整合在主人機介面，佈署與烙鐵相關的影像與溫度訊號擷取、訊號前處理、訊號特徵擷取，佈署與焊接平台相關之步進馬達控制、焊點影像擷取、影像分析，佈署與外部系統或功能串連如MySQL資料庫模組，預測生產狀態之機器學習模型。
最後規劃焊接實驗，以多個分類算法作為生產預測狀態，以回歸算法作為因果關係確認。以生產參數作為輸入，焊接通過與否與不良率分別作為模型分類標記，其中最佳的分類模型是支持向量機，實驗預測準確度為88.88%，AUC值達0.97。回歸模型使用之參數與良率t檢定有顯著，證明結果非隨機。
;Electronic soldering is the most important part of the connector manufacturing process. Because electronic soldering can produce solder joints, which directly leads to product failure, it is important to develop predictive maintenance. In this study, the Drag Soldering station was used as the intelligent target, and the RJ45 connector was used as the soldering object. The machine learning model was constructed by using the parameters of the production, and the sensing system was established to assist the production line personnel to judge whether the product passed.
The research is divided into three parts, namely, machine state evaluation and failure literature review, design implementation of software and hardware integration, and analysis of experimental results. First, the soldering process was evaluated, the types of soldering defect were sorted and the priority of the sensor arrangement was evaluated, and the failure of the soldering iron was prioritized. Then use the Queued Message Handler as the software system architecture, and implement the software and hardware modules and integrate them into the host computer interface. The image and temperature signals related to the soldering iron are extracted and pre-processed. Signal feature acquisition, stepper motor control related to soldering platform, solder joint image capture, image analysis, subordinate and external system or function serial connection such as MySQL database module, machine learning model for predicting production status.
Finally, the soldering experiment is planned, and multiple classification algorithms are used as the production prediction state, and the regression algorithm is used as the causal relationship confirmation. Taking the production parameters as input, the soldering pass or fail and the defect rate are respectively used as model classification labels. The best classification model is support vector machine, the experimental prediction accuracy is 88.88%, the AUC value is 0.97. The parameters used in the regression model and the yield rate were significant, indicating that the results were not random. |
