Data Augmentation Method For Industrial Soft Sensor Modeling

Soft sensor technology is widely used for industrial manufacturing and provides a new way of thinking,bringing new challenges.The modern industrial process tends to be complex and large-scale,resulting in many key quality variables in the industrial production process that are difficult to be directly measured or cannot be measured.As a result,the collected data presents the characteristics of a small total sample,strongly nonlinear,and strong correlation.The quality of data samples will seriously affect the predictive performance of data-driven soft sensor modeling methods.This work focuses on using a small number of complete samples and robust nonlinear and strong correlation data for soft sensor modeling.In this work,the data augmentation method and deep learning method are used for soft sensor modeling.The generative adversarial networks,convolutional autoencoder,and self-attention mechanism are generated to supplement and filter the data to ensure that the soft sensor model can conduct a reliable real-time online monitor,which is of great significance for intelligent industry.The main research content of this paper is summarized as follows:(1)Due to the problem of insufficient data samples in industrial data,a generative adversarial model based on stack variational autoencoder is proposed to construct new and high-value data samples.In this method,stacked variational autoencoders are used as generators of Wasserstein’s generated adversarial network for depth feature extraction.Moreover,adversarial training is used to train and optimize generators and discriminators to generate new samples similar to the original data.The results show that the data generated by the SVAE-WGAN network have better data distribution and matric value than those of VAE,GAN,and WGAN.(2)The soft sensor model based on a stacked convolutional autoencoder network is proposed to solve the problem of high dimensional and strong nonlinear characteristics of industrial data.In this method,the stacked convolutional autoencoder is used to extract the local features of the data at a deep level.The prior information of the original data is obtained from the input data using the encoder and decoder structure of the network model to improve the nonlinear representation ability of the network,and a soft sensor model is constructed.The SVAE-WGAN network was used to generate supplementary data samples for regression prediction.The results show that the proposed SCAE soft sensor model has good prediction performance.(3)The soft sensor model based on a self-attentional variational autoencoder is proposed for the strong correlation between industrial process data.The self-attention mechanism is used to capture the correlation information between the data and construct the strong correlation weight matrix between the data.The self-attention mechanism is embedded in the variational autoencoder network,and the fully connected neural network is used to build the soft sensor model.The SVAE-WGAN network is used to generate supplementary data samples for experiments.The results show that the prediction effect of the proposed SA-VAE model is significantly better than that of other contrast models.