Soft Sensor Modeling Based on Gated Stacking of Extended Long Short-Term Memory Networks

In many industrial processes, traditional hard sensors are unable to provide accurate data in complex environments due to technical limitations or high costs. Soft sensing technology can predict key process parameters that are difficult to directly measure or are expensive to measure based on easy-to-measure process variables. This can significantly improve the monitoring capabilities and control accuracy of the production process, reduce reliance on expensive detection equipment, and reduce production costs. Therefore, data-driven soft sensing techniques have been widely used to estimate hard-to-measure quality variables in industrial processes. In the debutanizer industrial process, soft measurement technology is used to predict butane content. However, existing models have limited memory capacity when processing long sequences of debutanizer process data and cannot take into account the longer historical time range. Extract effective feature representation within. To address these issues, based on Extended Long Short-Term Memory Networks (xLSTM), this paper proposes a Gated Stacking Soft Measurement Model called G-xLSTM. This model incorporates an exponential gating mechanism and a matrix memory structure to enhance the information filtering process, and capture useful historical information over a longer time frame. In addition, the model improves prediction performance by adding gated neurons to fully utilize the information flow represented by the features of each hidden layer, and weightedly integrates the prediction results of different levels to output the final value. Finally, the effectiveness and superiority of the proposed method were verified through an industrial case of debutanizer.