Information Filtering Unit-Based Long Short-Term Memory Network for Industrial Soft Sensor Modeling

The use of soft sensors is an alternative solution to physical sensors for real-time acquisition of key quality variables that cannot be measured online in real time. Long short-term memory (LSTM)-based methods have been improving the state-of-the-art in soft sensors. In order to mine the correlation between historical information and current input and utilize the relatively important historical information, we propose an underlying recurrent neural network (RNN) unit called information filtering unit (IFU) and an IFU-based LSTM (IF-LSTM) soft sensor. In IFU, a feedforward neural network is first constructed to predict the hidden state of the previous time step based on the current input, and subsequently, a scaling factor is obtained; the hidden state vector is then scaled to highlight those important components that contain nonredundant information with the current input. Finally, an IF-LSTM model is constructed by combining IFU and LSTM units for predicting hard-to-obtain key quality variables from easily measured process variables, and its performance is validated on two benchmark industrial datasets (sulfur recovery unit (SRU) and debutanizer column (DC) datasets). Experimental results show that the performance of the IF-LSTM outperforms popular RNN variants, such as LSTM, GRU, Bi-LSTM, and Bi-GRU in all metrics [root-mean-square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and ${R}$ -squared ( ${R}^{{2}}$ )], and achieves ${R}^{{2}}$ of 0.9929 on the DC dataset, indicating that the predicted values are well-fit to the true values. Furthermore, comparative experiments with state-of-the-art methods in the literature show that the IF-LSTM model achieves the best RMSE of 0.0069 on the DC dataset and the second-best RMSE of 0.0124 on the SRU dataset.

I. Introduction

In modern industrial processes, real-time access to key quality variables is critical for improving the performance of industrial control and optimization algorithms, but some of these quality variables are not always readily available due to high costs, delays, or overhauls [1]. An alternative solution is that these variables that are not available in real time can be estimated from a number of relevant and easily measured process variables by using soft sensors [2]. As an effective complement to physical sensors, soft sensors are being used in modern industry to predict hard-to-obtain key quality variables from easily measured process variables [3]. A soft sensor modeling process can include the selection of relevant process variables, sampling of industrial data, preprocessing of data such as outliers and missing values, construction of the soft sensor models, and updating the models after online-use, of which this article focuses on the construction of soft sensor models.