Abstract
盤尼西林系列的抗生素是目前使用最廣泛的抗生素之一﹐為了控制饋 料批式生產盤尼西林G之生化程序﹐使得其在相同的時間下﹐且加入的總 糖量為一定時﹐能有最大的盤尼西林G產量﹐我們將應用實數編碼遺傳演 算法(real coding genetic algorithm)找最適的糖進料策略。實數編碼 遺傳演算法有高精確度和高效率的優點，並能確定找到的結果是整體最適 值，是一種很有效率的方法。 解最適化問題時﹐通常需要有一個數學 模式﹐我們將採用Nicolai等在1991年所提出的數學模式來模擬饋料批式 生產盤尼西林G系統。 針對生化程序模式之建立是十分困難的﹐ 為了能快速且準確地建立一未知的生化程序模式﹐可應用類神經網路( artificial neural networks)以做模擬的工具。在只有少數實驗數據和 一些可得的資訊下﹐類神經網路可經由學習﹐進而模擬真實的生化程序。 類神經網路中又以混成(hybrid)類神經網路易於分析和應用﹐混成類神經 網路可以清楚地將反應模式分成基本系統方程式和反應程序參數估算器( 類神經網路)。混成類神經網路模式可視為含有反應程序參數的基本系統 方程式﹐其中反應程序參數是與狀態變數有關﹐並且由類神經網路來模擬 。 應用混成類神經網路﹐我們可建立一個模擬饋料批式生產盤尼西林 G系統之模式。經由試驗﹐此混成類神經網路模式可以成功地描述饋料批 式生產盤尼西林G系統。故若有生化程序的實驗數據﹐即使沒有其數學模 式﹐我們也能應用混成類神經網路﹐找到一個適當的模式以模擬生化程序 。 在饋料批式生產盤尼西林G系統的模擬分析中﹐經過120小時的醱酵 後﹐兩種模式(Nicolai數學模式和混成類神經網路模式)產生差不多的最 大盤尼西林G產量(分別為76.10695g和77.62g)﹐但其糖進料策略和醱酵過 程卻不相同。這可能是因為以混成類神經網路模擬饋料批式生產盤尼西林 G系統時﹐會與Nicolai數學模式有些許的差別﹐而最適化控制對模式的敏 感度相當高。所以﹐不同的模式所找的最適糖進料策略也會不同。 未 來在有饋料批式生產盤尼西林G系統的實驗數據下﹐可以應用混成類神經 網路來模擬﹐並以實數編碼遺傳演算法來找最適糖進料策略﹐再將此最適 糖進料策略應用於饋料批式生產盤尼西林G系統上﹐看看是否能提高盤尼 西林G產量。若可提高產量﹐則可證明此法是一個模擬與最適化的好方法 ，並可應用於任何生化程序上。 Penicillin is one of the most popular antibiotics produced from fermentation. In order to control the fermentation process for penicillin G production, we applied a real coding genetic algorithm to search the optimal feeding strategy. The real coding genetic algorithm has the advantages of high accuracy and efficiency and is easy to find the global optimum. A mathematical model is usually required to determine the optimal condition. The mathematical model developed by Nicolai in 1991 is applied to the fed-batch process for penicillin G production. In order to develop a model of an unknown process, neural networks can be employed. The model based on neural networks requires the training data from the system. In the present study, a model of fed-batch process for penicillin G production is developed by using hybrid neural networks. The hybrid neural networks model can be divided into two parts﹕one is the fundamental system equations ﹔the other is estimation of the reaction process parameters using the neural networks. The simulation results showed that hybrid neural networks model could successfully describe the fed-batch process for penicillin G production. Under modeling and analysis of the system of penicillin G production, both Nicoali and hybrid neural networks models have similar results in the maximum amount of penicillin G production. However, the feeding strategies and time courses of the fermentation were different. The season may be that optimal control is highly sensitive to system models. Applying the proposed method to fermentation system is straight forward. With the experimental data from the fed-batch operation for penicillin G production, hybrid neural networks are employed to setup the model. A real coding genetic algorithm is then carried out to determine the optimal feeding strategy. If this optimizing method can be employed to fermentation for penicillin production system in fed-batch culture, extension of this approach to other fed-batch fermentation systems should be applied.