Long-short Memory with Feature Representation by Latent-variable Model for Industrial Soft Sensor

Abstract

In modern industry, soft sensor of quality variables is essential for the safety, stability, and optimization of the processes. However, it is very hard to construct accurate soft sensor models in the horrible industrial environment. Firstly, the complex nonlinearity and variability are widespread in industrial process data. In order to extract useful information efficiently, feature representation is needed. As a probabilistic latent-variable model, variational auto-encoder (VAE) can map the observed data to the latent variable space and represent features in data robustly. Hence, the VAE is used as a feature extractor in this article. In order to solve KL vanishing in the standard VAE, a weighted VAE is developed to control the process of training. Besides, dynamic feature is also an important characteristic that should be considered in the process industry. To capture the temporal dependence in the time series, a long-short memory is utilized in this article, which can deal with the data in sequence and is used as a regressor. Finally, a long-short memory with feature representation by latent-variable model is proposed, which takes into account the nonlinearity, dynamics, and variability simultaneously. In order to prove the feasibility of the designed method, experiments were carried out on the real blast furnace ironmaking process. Through comparison with other methods, it is proved that this model can improve the prediction accuracy.