| 摘要: | 機器良率一般是從機器輸入的產品數量中輸出的良品比例而得，但若是遇到因生產成本問題，無法在每部機器結束後設置產品檢驗站，而無法得到每部機器輸入的產品數量與輸出的良品數量，則無法得到每部機器的良率。陶瓷基板其材質較容易破裂，但是檢驗站卻不容易觀測到機器所造成的細微裂痕，產生了不易歸咎機器之責任的暗裂問題。破裂的瑕疵成為陶瓷基板生產線上高成本耗費的來源，而錯綜複雜的生產線也使得機器良率的計算困難。綜合上述問題，若是得知生產線中每部機器的機器良率，則可找出造成產品破裂的元兇。在先前的研究中，已有可以克服上面所描述的問題，從已完成產品之生產資料中，以EM演算法估計過去的機器良率之方法。然而，機器會因使用而造成零件損壞或異常，機器的異常會造成更多的有破裂瑕疵的產品。若是只知道過去的機器良率，只能從機器其歷史機器良率中得知是否曾經異常，無法得知哪部機器將要異常。但若是能知道未來的機器良率，則可知道有哪些機器將會因為異常而產生大量的破裂瑕疵產品，進而提前阻止成本的耗費。本研究針對上述問題，提出一個基於時間序列預測的機器良率預測方法，使用過去的機器良率來預測未來的機器良率。並以Online Learning的模式，隨著時間推進及過去機器良率的更新，預測出各個時間點的未來機器良率。再將預測結果與實際的生產資料相比較，其中以週為單位的預測模型之平均誤差為2.85%，而以天為單位的預測模型之平均誤差為2.76%。此外，本研究之未來機器良率預測方法可應用於機器維修預警上。其中以週為單位的預測模型估計平均每週可以挽回13%的破片不良品，以天為單位的預測模型估計平均每天可以挽回17%的破片不良品。減少破片不良品的產生可降低生產成本，進而減少備品數，加速生產。;The machine yield rate is generally derived from the number of good products output from the machine divided by the number of products input to the machine. But if it is due to the production costs, the inspection station cannot be set after every machine in the production line. Because we cannot get the number of products input to the machine and the number of good products output from the machine, we cannot know the yield rate of each machine. The ceramic substrates are easy to crack, but the inspection stations usually ignore the products with micro-crack, so we cannot know which machine makes the product crack. Therefore, the cracked defects cost much on the ceramic substrate production line, but the complicated production line also makes the calculation of the machine yield rate difficult. Based on the above problems, if we know the machine yield rate of each machine in the production line, we can know which machines make the product crack.
In previous research, there has been an approach that can overcome the above problems. The approach uses the EM algorithm to estimate the past machine yield rate from the finished production data. However, machines may be abnormal after we start to use them, and the abnormal machines will produce more defective products. If we only know the past machine yield rate, we can only know the machines have been abnormal or not from their historical machine yield rate, and we cannot know the machine will be abnormal or not in the future. But if we know the future machine yield rate, we can know which machines will be abnormal and produce a lot of defective products, and then we can save the costs in advance.
Aiming at the above issues, this research proposes a future machine yield rate forecasting approach based on the time series forecasting, using the past machine yield rate to forecast the future machine yield rate. And we forecast the future machine yield rate for each period based on Online Learning. We compare the forecasting results with the real production data. The average error of the weekly forecasting model is 2.85%, and the average error of the daily forecasting model is 2.76%. Besides, the future machine yield rate forecasting approach can be applied to the machine maintenance early warning. The weekly forecasting model is estimated to save 13% of micro-crack defects per week on average. The daily forecasting model is estimated to save 17% of micro-crack defects per day on average. Reducing the micro-crack defects can save production costs. |
