A Novel Bidirectional Long Short-Term Memory Network With Weighted Attention Mechanism for Industrial Soft Sensor Development
Miao Zhang, Beike Xu, Jing Jie, Beiping Hou, Le Zhou
Abstract
Accurate measurement of key quality variables is of great significance for evaluating product quality and ensuring production safety. How to extract useful dynamic latent features from complex process data for effective quality prediction is still a difficult and hot topic in soft sensing modeling. The existing soft sensing models based on LSTM do not consider the correlation between hidden states of different time steps when extracting dynamic nonlinear features. To solve these problems, a novel bidirectional long short-term memory network with weighted attention mechanism (WA-BiLSTM) is proposed in this paper to improve the prediction performance for dynamic nonlinear industrial processes. In the WA-BiLSTM model, the bidirectional structure of BiLSTM is used to learn the dynamic information of the input sequence. A weighted attention mechanism layer is designed to dynamically extract the correlation information between the hidden states of different batches. In addition, the ratio between the historical hidden states and the current hidden state is adjusted reasonably by assigning adaptive parameters. Finally, the experimental results on a debutanizer column process and an IndPensim process demonstrate the effectiveness and superiority of the proposed method.