Self-Attention Soft Sensor Model Based on Long-Short Time Span *

Quality monitoring of industrial process exhibits a real-time characteristic, which necessitates the use of process variables to predict values of quality variables. This approach is commonly known as soft sensor modeling. Many previous data-driven methods have shown limitations in handling nonlinearity and temporal relationships. Utilizing deep learning to extract features reflecting the chronological relationship is an ideal approach, yet it requires concerted efforts to fully exploit the time dependence of data. Therefore, a self-attention soft sensor model based on long-short time span (LSTS-SA) is proposed. It is composed of an autoencoder that utilizes long short-term memory neural networks to extract shallow long-term features, and a self-attention mechanism to focus on shallow features from the time dimension to obtain deep features. In addition, it utilizes short-term features extracted by a time window and maps the fused features into the prediction space. The superiority of proposed method in dynamic soft sensor modeling is demonstrated in experiments on simulated dataset, the Tennessee-Eastman Process (TEP) dataset.