Due to the dynamics of industrial production, it is a great challenge to learn robust feature representations from industrial process data for building an accurate soft sensor model. The traditional variational autoencoder (VAE) can learn robust features that can adapt to the dynamic process better, but cannot be directly applied in soft sensor modeling. Therefore, a novel regression-based dynamic VAE (REGDVAE) is introduced in this paper. Firstly, the encoder of the DVAE, which is constructed by the graph attention and the convolutional neural networks, is utilized to obtain the robust spatio-temporal features. The decoder of the DVAE is responsible for reconstructing the input data via transposed convolution. Secondly, the Transformer is employed to capture the dynamic associations between the robust spatio-temporal features and corresponding outputs. Moreover, with the purpose of guaranteeing the confidence of the proposed method, the Gaussian loss function is used as the optimization target of the REG-DVAE to enhance the confidence of each predicted value. The proposed REG-DVAE is implemented in a real-world melt index modeling of the polypropylene production process. The results show that compared with other baseline methods, the REG-DVAE not only can achieve the best performance, but also can provide a confidence interval, which greatly enhances the credibility of the prediction results.
原文链接
分享
引用
