Ensemble Learning Modeling Method Based On Feature Weighting For Strip Steel Quality In Continuous Annealing

Hardness is an important indicator of steel strip quality in steel production process.The traditional off-line methods have obvious hysteresis while measuring the hardness of steel,and often lead to fluctuations of product quality,which brings considerable economic loss.With the development of data acquisition technology,the data-analytics modeling methods can be applied to the online prediction of strip hardness.As a data-analytics method,Least Squares Support Vector Machine(LSSVM)has high prediction accuracy and solving speed,however,its generalization ability still needs to be improved.The primary reason is that traditional LSSVM assumes the attributes of samples have the equal importance to the output and ignores the impact of different attributes,which may influence the predicted results.For the online prediction of steel strip hardness in continuous annealing unit,an ensemble learning modeling method based on feature weighted LSSVM is therefore proposed in this article.The feature weights and learning machine parameters are optimized collaboratively,and ensemble learning method is utilized to fuse different local learning machines,in this way,the prediction accuracy and generalization ability of the model is improved,the accuracy and stability of product prediction is ensured.The research contents include as following:(1)The initialization of feature weights:firstly,the production process data are preprocessed and abnormal samples are removed.Then analysis of the sample data is carried out to determine the initial weight value of each feature.(2)Feature weighted LSSVM modeling method:Based on the idea of feature weighting according to the different effects of features,the hardness prediction model based on feature weighted LSSVM model is constructed.The effectiveness of the feature weighted method is verified by the comparison results of two modeling methods,i.e.,feature weighted LSSVM and traditional LSSVM.(3)The optimization of model parameters:Two methods are designed to solve the parameter optimization problem of the feature weighted LSSVM.One method is to optimize feature weights and learning machine parameters at the same time.And another method is to optimize the parameters in a decomposition-coordination way.The latter method utilizes differential evolution algorithm based on the decomposition-coordination strategy to optimize feature weights and learning machine parameters concurrently,and then combine them to obtain the global optimal parameters.The results of experiment indicate the effectiveness of the decomposition-coordination strategy.(4)Ensemble learning modeling method:Single learning machine cannot guarantee a good generalization ability,and thus an ensemble learning modeling method based on feature weighted LSSVM is proposed.A comparative experiment of ensemble learning with different number of learning machines is designed to analyze the effect of parameter change.