Research On Classifier Based On Relative Density Noise Filter And Its Application In Aluminum Electrolysis

The generalization ability of the classifier is affected by the classification noise data on different degrees.The classification noise filtering method can significantly improve the generalizability of the classifier in the presence of class noise.The noise filtering method based on relative density classification is effective and widely applicable in dealing with symmetric class noise.However,its performance would be deteriorated by the non-uniform distribution and big volume of data.Based on the relative density noise filtering method,the thesis studied related research and applied it to the aluminum electrolysis industry data.The main research results are as follows.At present,the relative density method is mainly aimed at uniformly distributed data sets,and the performance on non-uniformly distributed data is poor.To solve this dilemma,this thesis proposes two methods:the noise filtering method based on threshold classifier and the noise filtering method based on sample classifier respectively,which make the methods robust to detect class noise in complex data environment.The proposed methods based on relative density experimentally show better generalizability than the original relative density model in the public data set.The time complexity of the original relative density model is O(N~2).Inspired by the relative density,a granular ball computing framework is proposed to improve the efficiency of the class noise filtering classifier,which is scalable,robust and efficient.In addition,the basic model of the relative density granular ball k-nearest neighbor classifier is given.The experimental results on each data set exhibit an efficient advantage on large data sets and better generalizability on some data sets.To further verify the effectiveness of the proposed method,the thesis will apply the proposed method to industrial aluminum electrolysis data.Due to the number of features of aluminum electrolysis data is small and missing values are large,a series of data preprocessing procedures were carried out.The missing values are deleted and populated,and redundant features are removed.A series of new features were constructed using the feature engineering method for the original data based on the data analysis.Moreover,standardization,feature transformation and feature selection were carried out for aluminum electrolysis data.In addition,the actual production environment of industrial aluminum electrolysis is complicated.For instance,the high-frequency fluctuation of the tank voltage caused by the needle vibration,the sudden increase of the temperature caused by the anode effect,and the fault caused by the external environment and the detection equipment,etc.,produce different degrees of classification noise.In addition,the superheat degree will significantly affect production efficiency of aluminum electrolysis.Therefore,this thesis employs the proposed methods to predict the superheat of aluminum electrolysis.The experimental results demonstrate that the proposed methods obtain better predictive performance.Finally,based on the optimal superheat prediction model,an aluminum electrolysis superheat prediction system was designed and completed.