Research On Virtual Sample Generation Technology Based On Information Diffusion

The advent of big data promotes the germination and maturity of data-driven modeling,and countless data bursts and accumulates in different fields.However,due to the low probability of sample occurrence and high cost of acquisition,the number of useful samples for modeling research is limited.The key to the performance of data-driven model lies in the number of training samples,uneven or unbalanced distribution of samples which results in poor learning ability,poor generalization ability and low accuracy of the model on the small sample set.It is very important and urgent to solve the problem of how to use small samples to build a robust and accurate model.At present,there are two ways to deal with the problem of small samples:one is to directly use relevant theories and knowledge,such as grey theory and machine learning;the other is to expand the original small sample set indirectly,such as adding virtual samples to it.In order to generate reasonable virtual samples,this paper proposes a virtual sample generation technology based on the mega-trend-diffusion (MTD) technology and Monte Carlo to improve the performance of the model on a small sample set.The proposed method uses the MTD to calculate the acceptable range of samples,at the same time utilizes the triangular membership function to estimate the distribution trend of samples and create the triangle probability distribution model,and then applies the Monte Carlo method to extract virtual samples to effectively fill the information gap of samples,which is used to improve the performance of the extreme learning machine (ELM).The method is further validated by two industrial data sets:multilayer ceramic capacitors (MLCC)and purified terephthalic acid (PTA).The experimental results show that the proposed method is an effective,reliable and advanced virtual sample generation method.