Novel virtual sample generation using Gibbs Sampling integrated with GRNN for handling small data in soft sensing

In order to optimize complex industrial processes, an accurate model is essential. The mainstream approach for complex industrial modeling is data-driven soft sensors. However, the accuracy of the established models is often low due to an insufficient amount of effective data, so the method of generating virtual samples has been proposed to achieve data augmentation, but the previous virtual sample generation methods have ignored the correlation between samples. To solve this problem, an effective virtual sample generation method based on Gibbs Sampling algorithm (GS-VSG) is proposed in this paper. In the proposed method, virtual input samples are first generated using the prior knowledge of the original data through the Gibbs Sampling method. Next, a generalized regression neural network (GRNN) model is constructed from the raw data, which is used to predict the output values of the virtual samples. Finally, the input and output parts of the virtual samples are combined to create a virtual sample set, which completes the extension of the original data set. To demonstrate the feasibility of the proposed GS-VSG method, numerical example and real industrial process dataset are used for simulation experiments. The results show that GS-VSG generated samples can improve the model accuracy and is a good technique for virtual sample generation.