A Variable Attention-Based Gated GRU Approach for Soft Sensor Models

Abstract

The actual industrial process is dynamic and nonlinear, so the collected data of the industrial process is a complex series of time series. Soft sensor modeling of nonlinear processes’ dynamic inputs and outputs is necessary for production processes. Gated neural networks with long-term memory can handle nonlinear time series data well and are suitable as a basic network model for soft sensor modeling. However, the simple Gated Recurrent Unit (GRU) does not consider the correlation between input and output variables nor the contribution of feature representation. A variable attention-based gated GRU (VAGGRU) network is proposed in this paper to solve the above problems. In this network, an attention mechanism is used to assign weights to different input variables according to the relevance of the inputs and outputs. Then, the contribution of the features proposed by each layer of the GRU is calculated by gated neurons to obtain further a feature representation that is highly relevant to the prediction target. Improving the model’s prediction accuracy is achieved by continuously extracting the feature representations related to the prediction target. Finally, the effectiveness of the proposed model was verified by applying it to the penicillin fermentation process and the actual industrial process through simulation experiments.