Penicillin End Point Moisture Prediction Algorithm Based On Machine Learning And Its Implementation

Into the 21st century,with the continuous improvement of technology,machine learning has already gained further growth,and been applied broadly in each field.However,as the largest producer of penicillin,the inspection of the yield and quality of penicillin products in China is still in the stage of manual testing.To evaluate the quality of penicillin production,one of the most important standards is the terminal moisture value of penicillin after crystallization.When the end-point moisture of penicillin crystallization is generally controlled between 0.6% and 1.2%,the yield and quality of penicillin will be at a relatively higher level.But in the real production,the end-point moisture value is simply estimated by manual sampling test method for numerical determination.Meanwhile,the crystallization process needs to be in a vacuum environment.All of the above illustrates the necessity of establishing a prediction model for penicillin end-point moisture,by which the yield and quality of penicillin could be predicted more precisely.Applying the theory of machine learning algorithm which includes ANN algorithm and supporting vector machine algorithm,this thesis concentrates on the study of penicillin end-point moisture predict algorithm.According to the actual production data of penicillin,the factors affecting the end-point water value of penicillin are vacuum degree,initial moisture,initial potency,etc.The prediction model sets the influencing factor as the input variable,and it sets the end-point moisture value as the output variable.In this paper,three kinds of prediction models of penicillin end-point moisture based on machine learning are established,including BP neural network prediction model,support vector regression model and radial basis(RBF)neural network model.These models respectively predict penicillin end-point moisture.First of the estimating process,the actual production data is normalized preprocessed,and then the three models are trained through the processed data.In each model,different methods are used to compare,that is how the best model is selected.Finally,the best model is applied to actual production,and the prediction results show that the method proposed in this paper has certain engineering application.The result of BP neural network prediction model based on LM algorithm training is better than that of BP neural network built by other algorithms.At the same time,the prediction accuracy of support vector regression model based on radial basis kernel function is improved by about 9% compared to BP neural networks.The prediction accuracy of the generalized radial basis function neural network model based on the orthogonal least-square method is about 15% better than the prediction accuracy of the BP neural network.