Triple Attention-based deep convolutional recurrent network for soft sensors

Abstract

Soft sensors have been widely applied in industrial processes, where exist nonstationary conditions, spatial correlations, and temporal dependencies. Without considering these characteristics, many previous methods failed to adaptively extract the characteristics of process data. To handle the abovementioned issues in soft sensor modeling, a triple attention-based deep convolutional recurrent network (TADCRN) is proposed in this paper. Firstly, multiscale 1d-CNN with scale-wise attention is designed to learn critical features from multiple receptive fields to eliminate the adverse influence of nonstationary characteristics. Secondly, space-wise attention is utilized to extract spatial correlations among multiple sensors. Thirdly, BiGRU with time-wise attention is designed to capture temporal dependencies among the consecutively collected data samples. To verify the effectiveness and efficiency of the proposed method, the experiments were conducted on the debutanizer column. The experimental results show that the proposed method outperforms both the conventional machine learning and deep learning methods.