Development of adaptive moving window LSTM‐EKF state and parameter estimator through maximum likelihood framework

For efficient control, monitoring, and online optimization of any processes, accurate information of the key process variables is critical. Data driven soft‐sensors, which can predict the key quality variables using historical data, have gained significant attention in the process industry in the recent years. These soft‐sensors are generally developed under normal operating conditions. In reality, chemical processes are subjected to abnormalities such as process parameter drift, sensor bias, or unmeasured disturbances. Under these abnormal conditions, the performance of data‐driven soft sensors developed under normal data conditions will deteriorate over a period of time. To account for these data abnormality issues, soft‐sensor models need to be adapted and robustified. In this work, a dynamic recurrent neural network‐based state and parameter estimator is developed through moving window maximum likelihood framework. Initially, a machine learning state estimator is developed by integrating long short‐term memory (LSTM) networks with the extended Kalman filter (EKF), which can estimate the critical quality variable using real‐time process measurements. To address the data quality issues such as process drift and sensor bias, in the presence of measurement delays, an innovation error‐based moving window maximum likelihood approach state and parameter estimator is proposed, where the LSTM integrated EKF is used as the state estimator. The performance of the proposed state and parameter estimator is evaluated on benchmark systems. From the analysis of results, it is observed that the proposed machine learning‐based state and parameter estimator accurately estimates the states in presence of data abnormalities.