| 摘要: | 隨著電子業對產能的需求提升，為降低生產設備維護成本及提升生產效率，焊接良率高低成為電子製造業生產力的重要指標，製程監測與管控問題在生產製程變得極為重要；常見的應對方法是透過自動化光學檢測(Automated optical inspection, AOI)的方式取得焊點影像，基於規則或機器學習方法對焊點進行瑕疵檢測，藉此管控產品良率。然而AOI無得知製程狀態的訊息，對於製程狀態異常的應對仍極度仰賴製程工程師經驗進行停機、調整製程參數，當異常發生時若未及時得知及判別故障類別(如溫度異常、轉臂轉速異常等)，將導致產線後端組裝產出大量不良品，且耗費大量的生產時間。
基於上述，本研究以廠商現有RJ45連結器之高頻被動元件自動化產線製程為例，探討智能性不良率評估與異常診斷方法。首先於產線評估階段，利用部分因子實驗設計以少量的實驗組數求得關鍵參數(如焊接溫度、浸焊轉速及浸焊停留時間)，以利於架設適當感測器(如熱電偶及慣性感測器)；接著透過實驗設計模擬產線容易出現的異常狀態，利用自組織映射(Self-organizing feature maps, SOM)將數據映射於二維拓樸結構，觀察群集的分布以篩選有效特徵。爾後以多層感知機(Multilayer perceptron, MLP)建構不良率評估模型，提供不良率估計，以提早得知製程不良，減少後端失效的料件成本。最終在30組測試數據上得到1.1%的平均絕對誤差；另以SOM的輸出作為K-means++算法的訓練數據來建構異常狀態模型，在測試數據集上得到100%的F1度量(F1-measure)。本研究透過不良率評估與異常診斷模型的相互輔助，建構不良率評估及異常診斷系統，當異常發生時，產線人員可即時判斷異常製程狀態，減少調整製程所費時間成本，藉此提升產線稼動率。
;With the increasing demand for production capacity in the electronics industry, in order to reduce the maintenance cost of production equipment and improve production efficiency, the soldering yield rate has become an important indicator of the productivity of the electronics manufacturing industry. Process monitoring and control issues have become extremely important in the production process; the common way to solute the problem is to capture the solder joint image by means of automated optical inspection (AOI), and perform defect detection on the solder joint based on the rule or machine learning method, thereby controlling the product yield. However, the AOI has no information on the process status. The process status abnormality still relies on the process engineer experience to stop and adjust the process parameters. If the abnormality occurs and fault category (such as temperature abnormality and arm revolution speed) is unknown on time, that will result in a large number of defective products at the back end of the production line, and it takes a lot of production time.
Based on the above, this study takes the RJ45 connector of high-frequency passive component of automatic production line as an example to discuss the intelligent defect rate evaluation and abnormal diagnosis methods. First, in the production line evaluation stage, the partial factor experiment design is used to obtain key parameters (such as soldering temperature, dip soldering speed and dip soldering dwell time) with a small number of experimental groups to facilitate the installation of appropriate sensors (such as thermocouples and habits). Then, through experiments, the abnormal state that is occur frequently in the production line is simulated. The self-organizing feature maps (SOM) mapping the data to the two-dimensional topological structure, and the distribution of the cluster is observed to select effective features. Subsequently, the multi-layer perceptron (MLP) used to construct a defect rate evaluate model to provide an estimate of the defect rate, so as know the poor process and reduce the cost of the back-end failure. Finally, the defect rate evaluate model obtain the score of average absolute error 1.1% on 30 sets of test data. The output of SOM was used as the training data of K-means++ algorithm to construct the abnormal state model, and 100% F1 measure (F1-measure) was obtained on the same test data set as above. In this study, through the mutual assistance of the defect rate evaluate and the abnormal diagnosis model, the defect rate evaluation and abnormality diagnosis system are constructed. When an abnormality occurs, the production line personnel can immediately judge the abnormal process state and reduce the time cost of the adjustment process, thereby improving the production line rate. |
