Prediction Of Sensory Indicators For Cigarette Based On Priori Knowledge

With the improvement of national living standard in our country,the market demand for tobacco products is also increasing instantly.Facing this trend,tobacco companies need to improve their core technologies,promote the product research level and enhance the market competitiveness of products.Product sensory evaluation is the foundation of cigarette formula design,which is the key problem of cigarette product research.The quality of sensory evaluation directly affects the quality of cigarette products.At present,the sensory evaluation of tobacco companies mainly relies on the evaluation of tobacco experts.Experts need to catch various characteristics of flue gas in a short period of time and get evaluation results.However,in practice,making objective assessments of tobacco products quality is also complex and difficult.At the same time,the description of sensory characteristics is not detailed enough and lack of stability,so it is difficult to ensure the evaluation quality.In the past years,tobacco companies have accumulated some historical data in the process of cigarette production and technology test,which provides an important prerequisite for the research and optimization design of cigarette products.In order to establish intelligent sensory smoking cigarettes prediction model,some data mining methods are used to find the relationship between cigarette sensory index and chemical composition based on existing historical data.However,existing cigarette sensory prediction models only consider the feature of historical data,and ignore the role of prior knowledge.When there is complex data distribution or there are some strong noise data,the prediction performance of the model will be greatly affected.The cigarette data has the characteristics of small samples,high dimensionality and strong noise.It is difficult to achieve satisfactory prediction results using a single data mining algorithms.In this thesis,it is considered that sensory evaluation experts have accumulated a lot of prior knowledge in the process of many years' sensory evaluation which can be applied to the data mining algorithm and used to build a sensory prediction model based on prior knowledge.The model using prior knowledge can help itself to learn the inherent characteristics of data and improve its forecast performance.The main research contents in this article are as follows:(1)The data of cigarette sensory evaluation are preprocessed,and the rules of data can be obtained by ID3 decision tree.Then the rules are reduced by removing the extra rules and excess attributes.Finally,the prior knowledge of correlation between chemical components of tobacco is obtained and it is applied in the sensory prediction models which are based on domain knowledge.(2)The prediction model of cigarette sensory evaluation indicators based on BP neural network(BPNN)is established.The optimal hidden layer node numbers of BP Neural Network is found and the best model in prediction performance is chosen.By combining the prior knowledge that ID3 provides,the Knowledge-Based Artificial Neural Network(KBANN)and the Revising Approximate Probabilistic Theories Using Repositories of Example(RAPTURE)are established.According to the experiment with these models,it can be found that the prediction accuracy and the stability of KBANN and RAPTURE are better than those of BPNN.(3)According to the characteristics of cigarette sensory evaluation indicators,the prediction model of cigarette sensory evaluation indicators based on Bayesian Networks is established.The performance of the prediction model is analyzed.By using the correlation among the chemical components of tobacco,a prediction model based on prior knowledge of Bayesian Networks is established.Through the experiment with these models,it can be found that the prediction model based on prior knowledge of Bayesian Networks is better than the model based on Bayesian Networks.In addition,because the prediction model based on prior knowledge of Bayesian Networks does not have network structure learning process,it has shorter learning time than the prediction model based on Bayesian Networks.