Gaussian mixture TimeVAE for industrial soft sensing with deep time series decomposition and generation

Abstract

Most industrial process data is time series data and contains multi-mode characteristics, which poses difficulties and challenges in the establishment of soft sensing models. To address these issues, this paper proposes a Gaussian mixture based time series decomposition model. This model innovatively introduces Gaussian mixture distributions into the latent space and utilizes a time series decomposition module in the decoder to decompose complex distributions. On one hand, the latent variables of the Gaussian mixture distribution can better extract complex features from time series inputs. On the other hand, the time series decomposition module can break down and extract disentangled features from the time series perspective. Furthermore, to tackle the problem of poor fitting in peak or extreme data due to information imbalance, it generates virtual time series data. The generated virtual time series can complement the information of poorly fitted data, supplementing the original data, and contribute to a better soft sensing model. Finally, to validate the effectiveness of the proposed methods, the soft sensors based on the proposed model are applied to two real industrial cases. The experimental results show that the proposed models have superior predictive performance compared to other state-of-the-art methods.