System Modeling And Data Feature Construction Based On Granular Computing

With the advent of the information age,the advancement of science and technology,the data generated by daily life and industrial environments continues to grow,such as network monitoring,vehicle service systems,etc.How to analyze and use these data,mine its internal meaning,and make it better serve human beings,has become a challenging problem.Faced with these challenges,people have developed and designed various intelligent systems for different fields with corresponding success.The purpose of this dissertation is to establish a general granular model/linguistic model through the research on the construction method of information granules.In this dissertation,the concepts and ideas of granular computing are applied to the fields of time series classification and reinforcement learning control,so as to expand the application scope of granular computing and new directions worthy of exploration.Based on the existing reasonable granularity criterion,this dissertation proposes an augmented principle of justifiable granularity.The principle of justifiable granularity does not fully consider the influence of output spatial information on the distribution of input spatial data,so the obtained information granules may not be optimal.How to make full use of the information contained in the output space and construct the optimal information granules in the input space is a problem worthy of further exploration.The gain reasonable granularity criterion proposed in this dissertation uses the information implicit in the output space to quantify the performance of the information granularity,using three criteria of coverage,specificity and output space distribution to express its optimality.The flexibility of this method stems from the introduction of an adaptive weighting scheme for output spatial information,which constructs optimal information granules by selecting weights of weighting factors.When designing the hidden layer of the granular model,based on the connection between the information granules in the input and output spaces,an efficient reference scheme is constructed for the calculation between the input and output information granules.The information granules constructed by the method in this dissertation can be directly used for the prediction model of spatially distributed data,and the output results of the model are expressed in the form of information granules.The position and size of the output information granules are used to predict the results and reflect the accuracy of the prediction results.Aiming at the problems of poor interpretability and large output error in traditional linguistic models,this dissertation proposes a linguistic model based on conditional fuzzy clustering optimization.First,by using semantics to divide the output space,the division of the output space fully combines the data distribution characteristics of the input space,so that the division of the output space has the characteristics of interpretability.Secondly,according to the information of output space division,the input space is divided,and the method of conditional fuzzy clustering is used in the divided input subspace to generate information particles.This segmentation process makes full use of the experimental data of the input and output space,and realizes the segmentation of the output and input space through the corresponding semantics.When new data enters the linguistic model for inference,in the corresponding input subspace,the model calculates the relationship and activity between the new data and the subspace information grains,aggregates and outputs their approximate output grains.The method proposed in this dissertation effectively improves the interpretability and accuracy of the linguistic model by using semantics and combining the data distribution characteristics of the input space to segment the output and input spaces,and perform conditional clustering in the input subspace.In the face of the problem of time series classification,how to better mine and present the features of time series and use them to improve the classification efficiency is a key issue.In this dissertation,a convolutional neural network classification method based on granular computing is proposed.First,by using the fuzzy mean method,a set of concept centers is formed,and a two-dimensional grayscale image is constructed through the membership matrix of the sequence to be classified and the concept center,which is used as the input of the convolutional neural network classifier.Secondly,optimize the convolutional neural network classifier,and replace the traditional convolution kernel with a wavelet kernel function for the grayscale image input converted in this dissertation to improve the classification efficiency.Combining the idea of granular computing,this dissertation proposes an image transformation classification method for time series,which converts one-dimensional time series into two-dimensional grayscale images,and redesigns the kernel function of convolutional neural network,which effectively improves the classification efficiency.In the traditional machine learning and industrial control fields,the division of state space and state estimation have always been issues worth exploring.This dissertation proposes an application method of reinforcement learning based on granular computing in the control field.This method uses reasonable granularity criteria to divide the state space granularly to form a set of information granules.At the same time,a reasonable inference system is designed to iteratively train the granules and test the final model.This method,using the principle of justifiable granularity,can perform state space segmentation for different systems or subjects flexibly.Based on granular computing,this dissertation aims at the construction method of information granules and the application of information granules in system modeling,and combines the principle of justifiable granularity with the fields of time classification and reinforcement learning,which provides valuable research directions and fields of application for granular computing.(1)An augmented principle of justifiable granularity is proposed,which makes full use of the information in the output space and optimizes the position and radius of the information grains in the input space.(2)For the granular model,a reasonable inference scheme is proposed,which flexibly calculates the granular output of new data and has good accuracy.(3)Combined with semantics,the output space is divided,and the input subspace information granules are formed according to the conditional fuzzy clustering,and the formed information granules have stronger interpretability.(4)For different fields such as time series classification and reinforcement learning,this dissertation flexibly applies the idea of granular computing to provide a new research idea for data feature detection,reconstruction,state space division,and prediction.In the follow-up research,the application of different types of information granules in system modeling would be discussed in detail,and the application of granular computing in different fields could be optimized.