Research Of Rule Extraction From Neural Network And Its Application To Strip Hot-dip Galvanizing Quality Modeling

Although ANN (Artificial Neural Network) has been widely used in pattern recognition, control and decision-making, system modeling, the inherent"black box"characteristic of ANN has greatly limited their further application. This work studies the rule extraction method of ANN, make clear the relationship between inputs and outputs of ANN and make the model easy to understand. What is more, we apply the"knowledge"extracted from ANN to production quality modeling in order to automatic set the quality parameters combining with quality prediction model, which avoid the artifical seting by people's experience. It has the important significance for deeply understanding the production law, improving the production technology and the product quality.The main content of this work is as follows:(1) The method of the rule extraction of ANN based on optimized activation functions is put forward, which is applied to data mining and knowledge discovery in the production of hot-dip galvanizing and overcomes the defect of"poor explanation"of the traditional ANN. The penalty term of the exponential variable is applied to make the values of the activation function have a better approximation to binary values: 0 or 1,which is helpful for rule extraction. The corresponding relationships among the raw materials parameters, control parameters and the product quality are extracted in the form of production rules, which provides effective means of analysis and control in production process monitoring and quality management. The results of model verification using actual product data of zinc coat weight showed that the coverage rate of the knowledge rules extracted from our method has reached 94.7%.(2) The prediction-control method of process control parameters in the multivariable production quality model is proposed. The knowledge rules in existing production data are extracted to be constraint conditions of parameters setting problem. Simultaneity, the quality prediction modeling was built based on process control parameters, which was used to find the exactly control parameters in the range of the rules. The proposed method has the advantages of small search space of problem domain and fast convergence during optimization, and has been successfully applied to off-line navigation system for zinc weight control in hot-dip galvanizing strip.