Research Of Virtual Sample Generation Technology Based On T-SNE And Modeling Application

In the process industry,it is necessary to establish a data-driven soft sensor model to obtain some key variables that are difficult to measure directly.The performance of data-driven soft sensor model largely depends on the quality of data.Unfortunately,in some cases,it is difficult to obtain sufficient data and data fluctuation is small.It is difficult to obtain sufficient information data in limited samples,which will lead to the problem of small samples.In this paper,a novel virtual sample generation method based on tdistributed stochastic neighbor embedding(t-SNE)is proposed.Firstly,tdistributed stochastic neighbor embedding(t-SNE)is used to extract the features of input data.The interpolation algorithm is used to obtain the virtual t-SNE input features,and then the random forest algorithm is used to construct the relationship between the original distribution and t-SNE features and the relationship between input and output to obtain the input and output value of virtual samples.In this paper,a novel compensatory function linked neural network(FLNN)model is proposed to improve the accuracy of the FLNN.Based on the traditional FLNN model,the additional model between the input value and the error is constructed,and the result is added to the first FLNN model to obtain a more accurate result.In this paper,the effectiveness of the t-SNE-VSG is verified by using standard data set and small data set of industrial process of purified terephthalic acid(PTA)production process.The simulation results show that the virtual sample generation method based on t-SNE proposed in this paper effectively solves the problem of small samples.The PTA data set after data expansion and ethylene industry dataset are used to verify the effectiveness of the proposed compensatory FLNN.The simulation results show that the proposed compensatory FLNN can further improve the accuracy of the soft sensor on the base of solving the small sample problem.