隨著電子產業對生產效率需求提升，焊接良率成為連接器製造過程中最重要指標。焊接過程產生瑕疵焊點時，會導致產品功能直接失效，因此發展預測性維護是極為重要。本研究以電子零組件拉焊製程工站為智慧化目標，以RJ45 電子零組件為拉焊對象，利用製程之參數建構深度學習模型及長短期記憶模型預測設備元件之剩餘使用壽命，並以感光耦合元件擷取其影像進行分析，求出製程參數優化組及烙鐵頭剩餘使用壽命預測。 本研究以拉焊製程之參數建構深度神經網路(Deep Neural Network, DNN)，從製程參數中選出重要幾項參數因子為以下3 項：烙鐵頭溫度、拉焊速度及錫絲出錫速度。透過類神經網路訓練後得到一組製程參數優化，烙鐵頭溫度：363 C、拉焊速度：7 mm/s、錫絲出錫速度：15 mm/s。此製程參數優化組預測焊點良率為99.998%且模型平均誤差為0.0179。得知製程參數優化組後進行實驗驗證，並利用2 台照相機分別將烙鐵頭及電路板影像記錄下來，隨後經影像處理後得知烙鐵頭上所沾錫區域值、沾錫區域重心位置及烙鐵頭磨耗長度。將此數據資料利用長短期記憶模型建立預測烙鐵頭剩餘使用壽(Remaining Useful Life, RUL)之模型，以在發生瑕疵焊點導致電路板失效前發出適當預警，及時更換烙鐵頭。 本研究目標將達成拉焊製程參數優化以及烙鐵頭剩餘使用壽命評估，達到智慧製造的目的。準確更換烙鐵頭時機並提升產品良率，以及降低物料庫存之成本。;With the increasing demand for production efficiency in the electronics industry.Electronic soldering is the most important process in the connector manufacturing process. When this process produces defective solder joints, it will lead to the direct failure of product functions. Therefore the development of predictive maintenance is extremely important. This study applies machine learning on the drag soldering process for printed circuit board (PCB) of RJ45 electronic connector. Using the process parameters to construct the deep neural networks(DNN) model and the long short-term memory (LSTM) model enables us to predict the remaining useful life of the soldering tip based on the captured images of the soldering tip. In this study, a DNN model was built based on the parameters of the drag soldering process. After capturing the soldering images by cameras, DNN was trained to achieve the optimum process parameters, including soldering temperature of 363℃, soldering speed of 7mm/s, and soldering tin feeding rate of 15mm/s. A high PCB yield of 99.998% was predicted by using the attained optimum drag-soldering process parameters. Then, the average error of this model is 0.0179. Experimental verification was performed after learning the process parameter optimization group. Two cameras were used to record the soldering tip and PCB images respectively. After image processing, the soldering area on the soldering tip, the position of the center of gravity of the soldering area and the abrasion length of the soldering tip were obtained. The images of soldering iron tip and soldering joints were then used by LSTM to predict the RUL of the soldering iron tip. According to the predicted RUL, the soldering iron can be replaced in advance to prevent defective soldering solder joints on PCB. The goal of this research is to achieve optimum drag soldering parameters and to evaluate the remaining useful life of the soldering tip. Timely replace ment of the soldering tip improves product yield, as well as reduces the cost of material inventory.
