Dynamic Soft Sensor Modeling and Its Application Using TCN-CBAM-LSTM Method

Data-driven soft sensor modeling has gained significant traction in modern chemical process monitoring and quality prediction. However, persistent challenges remain in accurately characterizing the complex dynamics inherent in chemical production systems, which typically exhibit significant time delays, strong nonlinearity, and time-varying characteristics. To address these critical challenges, a dynamic soft sensor modeling method based on temporal convolutional network (TCN) combined with channel spatiotemporal attention module and long short-term memory network (TCN-CBAM-LSTM) is proposed. Firstly, TCN is employed to extract deep nonlinear dynamic dependencies from process variables through its dilated causal convolution architecture, secondly, a convolutional block attention module (CBAM) is incorporated to enhance feature representation by adaptively focusing on critical spatiotemporal information across different sensor channels, finally, a long shortterm memory network (LSTM) is integrated to model intricate temporal patterns and long-range dependencies between process variables and quality indicators. This multi-stage architecture enables comprehensive learning of both local temporal features and global dynamic relationships within complex chemical processes. To verify the effectiveness of the proposed method, TCN-CBAM-LSTM was applied to a soft sensor modeling example for calculating the exhaust gas composition in a sulfur recovery unit (SRU). Under the same experimental conditions, it was also compared with convolutional neural network (CNN), variable weighted stacked autoencoder (VWSAE), spatiotemporal attention LSTM (STA-LSTM), CNN-LSTM, TCN, and TCN-LSTM. The results show that the TCN-CBAM-LSTM method has better performance and modeling accuracy, and its performance meets the needs of practical engineering applications.