A Study Of Processing Method Of Small Sample Size Data Based On Conditional Generative Adversarial Networks And Extreme Learning Machine

With the continuous advancement of science and technology,the importance of data has become increasingly prominent.The knowledge gained from a large amount of data is the goal of data processing.However,many fields such as medical,defense industry,and aerospace cannot obtain enough data due to various factors such as environment,time,and cost.Due to the small number of samples,existing data-driven machine learning methods do not perform well on it.How to effectively process small-sample data poses new challenges for existing machine learning methods.Based on the Conditional Generative Adversarial Networks to expand the small sample data,this thesis uses the particle swarm optimization-based extreme learning machine to realize the feature extraction and classification of small sample data.The main work of this thesis is as follows:(1)The existing conditions generate the samples generated by the Conditional Generative Adversarial Networks,and there are defects in which the degree of difference between classes is small and the generated samples of different categories are difficult to distinguish.In view of this,this thesis proposes a method for Conditional Generative Adversarial Networks based on the distance between classes.First,the maximum expected algorithm is used to fill the missing values in the original sample,and then the inter-class discrimination information is added to the generation network of the Conditional Generative Adversarial Networks model.In the case of the category information tag data,different class samples having discrimination are efficiently generated.The proposed method can positively distinguish each category of samples while expanding the number of samples,and helps the machine learning algorithm to further analyze the expanded data set containing the generated samples.The experimental results show that on multiple small sample data sets and MNIST,the improved Conditional Generative Adversarial Networks generated samples have larger inter-class distance and more obvious inter-class discrimination than traditional methods.(2)Aiming at the robustness of the traditional extreme learning machine method due to the influence of random initialization weights on its generalization performance,this paper proposes a class of double hidden layer over-limit learning machine based on particle swarm optimization algorithm,aiming at improving small sample data.Feature extraction and feature classification performance.The first layer of the model adopts a variable length overrun learning machine based on the automatic encoder structure to extract effective data features.The second layer of the model adopts a variable length overrun learning machine based on particle swarm optimization algorithm,that is,the use of particle swarm The algorithm optimizes the weights in the double hidden layer network by fitting error and classification cross entropy as the fitness function,so as to classify the small sample data more accurately.In addition,this method effectively limits the excessive growth of hidden layer nodes and avoids the over-fitting problem of extreme learning machines.The experimental results show that on the six benchmark classification datasets,the improved algorithm can obtain the effective feature representation on the small sample dataset,and achieve better classification accuracy than the traditional algorithm.