Study On Machine Learning Based Product Design And Its Application On Lubricant Formula

In today's world,the research and application of artificial intelligence(AI)are in full swing.The core algorithm of artificial intelligence is relatively mature and difficult to change,but the application of AI technology is just ascending.As one of the key parts in artificial intelligence,the feature of machine learning technology is that it can efficiently and accurately summarize and learn regression,clustering,classification models about a complex problem from a large amount of accumulated data and improve the performance of the system itself.Using machine learning for product design is also an important direction for AI applications.Despite the explosive development of lubricant formulations in the past,the blending of formulations involves complex physical and chemical processes,and is still in a"semi-experience" design mode up-to-date.Few people explore the use of machine learning methods to understand and optimize the blending process in order to achieve "rational design".This thesis takes lubricant formulation as a probe case to explore the application of machine learning models in product design.Based on the collected actual lubricating oil formulations and their corresponding physicochemical properties,combined with the data obtained from the empirical calculation methods of existing industrial applications,a classification model for oil component packages and certain physical and chemical performance indicators and decision factors were established.Model of the mapping relationship between them were built,and the generalization ability of the model was evaluated.The research contributions are as follows:Firstly,the oil blending mechanism and preprocess the recipe data were investigated so that it can be used as an "input" for subsequent machine learning.That is to do feature extraction,missing patch and abnormal values normalized,data noise filtering and other processing on the data of the lubricant formula,to create a "clean" oil component data set,and then to standardize the data to adapt it to the relevant machine algorithm.Secondly,according to the data characteristics,the core algorithm of the machine learning method suitable for the data was selected,the core algorithm is used to construct the overall algorithm program framework,and the formula was classified and mapped.In the classification of oil component data sets,four core algorithm models namely decision tree,support vector machine,Bagging parallel tree,and RUSboost random under sampling tree were used.Each model used a 50%cross-validation method to avoid over-simulation.At the same time,the confusion matrix,the susceptibility curve(ROC)curve,and the multi-class evaluation matrix were used to judge and evaluate the model,and then the classificationaccuracy of the oil component data set of each model was compared.The final accuracy of the RUSboost tree is 96.8%as the best.In terms of index mapping regression prediction,the three models of regression tree,support vector machine and gradient lifting tree were also used to verify the five-fold cross-validation method.Response graph,prediction graph and response graph,and residual graph were used to evaluate the model performance,and the root mean square error(RMSE)of the six indicators under the three models was statistically compared and showed that the gradient boosting tree(GBDT)was the best among the methods used here in terms of predicting the indicator value.Finally,a sequential vector representation method is proposed for the oil component data set.Using two popular model neural networks(DNN)and long and short-term memory(LSTM)networks as core algorithms,the model algorithm architecture programming to classify indicator mapping was established.Combining traditional bagged vectors to train and verify the DNN,the classification of oil components is 91.38%and the regression evaluation index RMSE is 0.0993.Using the sequence vector of the simulated recipe process to input into the LSTM,the classification accuracy rate is 97.3%,and the regression evaluation index RMSE is 0.0964.