The Detection Of Outlier And Application With Neuron-Fuzzy Modeling

These years, neuro-fuzzy modeling has been applied to knowledge modeling which transformed knowledge in special fields into computer language. In all kinds of neuro-fuzzy modeling methods, the ANFIS(Adaptive-Network-based Fuzzy Inference Systems)method which was proposed by Jang in 1993 is the most prominent one.usually,the modeling data set is clean, we can acquire satisfied model mostly if it is clean,but if the data set is not clean,data with outliers,we will not acquire satisfied ANFIS model.In this paper,the main task is outliers detection before modeling use ANFIS on the data of estimation of SOC(state of charge) and the data of the remain fuel of plane fuel tank.Firstly,the paper proposed definition of outlier and make a summary of several usually outlier detection methods,indicate their strongpoint and deficiency,then give an example to explain the meaning of analyse outlier.Secondly,the paper show the disadvantage the outlier to ANFIS through a function example,such as the structure of model and RMSE of model.Thirdly,the paper proposed two advantage algorithms based on usuall detection methods,the PCA(Principle component analysis) and the PLS (Partial least square ), particular explained the theory of the PCA method and PLS method.Finally,the paper particular analyzed two modeling examples,proved that the errors and structure of two model with outliers or without outliers are difference,then contrast the performance of four outlier detection methods on these two examples.The result shows that the outlier detection method use PLS and PCA are reasonable and effective.The innovation in this paper is analyzing the influence of the outlier to ANFIS and effectively using the PLS and PCA in outlier detection.We can exactly find outliers use these two methods in data set that exist mutivariables and correclation.