Dynamic soft sensor modelling based on data imputation and spatiotemporal attention

Soft sensor technology is essential for achieving precise control and improving product quality in industrial processes, with broad application potential in chemical engineering as well. In industrial soft sensor modelling, while most models can capture the nonlinear and dynamic characteristics of time series, they often neglect the potential influence of spatial features. Additionally, due to factors such as signal instability, equipment failure, and sensor data packet loss, missing values are common in industrial data, which can compromise model accuracy. To address these issues, this paper proposes a soft sensor modelling framework based on a spatiotemporal attention network for quality prediction with missing data. The method first utilizes a generative adversarial imputation network (GAIN) to impute in the missing data. Then, a bidirectional long short-term memory (BiLSTM) encoder integrated with a spatial attention module is employed to more precisely capture spatial correlations among variables in industrial processes, enhancing the capacity of the model to handle complex spatial dependencies. Furthermore, a temporal attention mechanism is incorporated to strengthen the extraction of dynamic dependencies across different time steps, further improving the ability of the model to capture nonlinear and dynamic features in industrial processes. Extensive experiments on debutanizer and steam flow processes validate the superior performance of the proposed method, laying a foundation for its application in chemical engineering and other complex industrial processes.