Research On Penicillin Potency Prediction Algorithm Based On Machine Learning

Penicillin,as a widely used antibiotic in clinical practice,has a great demand both domestically and internationally.The most important fermentation stage in penicillin production is usually time-consuming and costly.In order to improve the yield and quality of penicillin during the fermentation process and reduce production waste,it is necessary to optimize and control its production process.Real time monitoring of penicillin efficacy is crucial in optimizing the fermentation process.At present,manual sampling and testing are used in industrial production to obtain the efficacy value of penicillin.Usually,the testing time interval is long and has a significant time lag,making it difficult to provide accurate and high-quality reference for the production process.The fermentation process is an extremely complex biochemical reaction process with strong nonlinearity and time-varying characteristics.In addition,the complex industrial production environment makes it difficult to construct a very accurate mechanism model.The development of machine learning provides new ideas for modeling fermentation processes.This article conducts in-depth research on machine learning based penicillin potency prediction algorithms.At present,most data-driven model studies use laboratory scale penicillin fermentation data.This article uses data generated by industrial scale batch fermentation with feed.Firstly,the Pearson correlation coefficient is used to calculate the characteristic variable with the greatest correlation with the efficacy value as the input of the network.Finally,the fermentation time Time,oxygen absorption rate OUR,carbon dioxide concentration in the exhaust gas,container weight Wt,aeration rate Fg,volume V,oil flow Foil,and bottom flow rate Fb are determined as the input variables,and the penicillin efficacy value is determined as the output variable.The data is normalized.This article mainly uses recurrent neural networks and Bayesian neural networks to predict the potency of penicillin.Firstly,an attempt was made to predict the potency of penicillin from a time series perspective,using RNN and LSTM networks for modeling.In order to improve the efficiency and performance of searching for the optimal model,the PSO-LSTM model is constructed,and the PSO algorithm is used to find the optimal hyperparameter for LSTM network.In order to solve the problem of poor generalization ability and easy overfitting of general neural networks,Bayesian neural network modeling is proposed.In order to verify the performance of this model,BP neural networks with the same structure are used to compare the prediction results.The final experiment proved that the LSTM network has better predictive performance than the RNN network,and the PSO-LSTM algorithm improves efficiency while ensuring model performance.Under the same structure,Bayesian neural networks have higher accuracy than BP neural networks.