Nonlinear Dynamic Soft Sensor Development with a Supervised Hybrid CNN-LSTM Network for Industrial Processes

Adaptive Temporal-Spatial Pyramid Variational Autoencoder Model for Multirate Dynamic Chemical Process Soft Sensing Application

Data-driven soft sensors play an important role in practical processes and have been widely applied. They provide real-time prediction of quality variables and then guide production and improve product quality. In practical chemical production processes, nonlinear dynamic multirate data is widespread and challenging to model. This paper innovatively proposes a temporal-spatial pyramid variational autoencoder (TS-PVAE) model for the nonlinear temporal-spatial feature pyramid extraction from multirate data. This structure not only selectively utilizes multirate data but also handles complex nonlinear time-series data. Based on this, integrated with just-in-time (JIT) learning, an adaptive TS-PVAE (ATS-PVAE) model is developed. In this model, historical data are used for real-time fine-tuning of the model, leading to the development of an adaptive model. Finally, the proposed models are validated by an industrial case of a methanation furnace, demonstrating a superior estimation performance.

Multivariable System Prediction Based on TCN-LSTM Networks with Self-Attention Mechanism and LASSO Variable Selection

Intelligent prediction of key output variables that are difficult to measure online in complex systems has important research significance. In this paper, by using the least absolute shrinkage and selection operator (LASSO) algorithm to analyze the principal elements of input variables, a temporal convolutional network fused with long short-term memory (TCN-LSTM) network and self-attention mechanism (SAM) is designed to realize dynamic modeling of multivariate feature sequences. For complex processes with multiple input variables, each variable has different effects on the output, so it is necessary to use the LASSO algorithm to perform regression analysis on the input and output data for selecting the principal component variables and reducing the redundancy and computation burden of the network. The TCN network is used to extract the features of the input variables efficiently. The long-term memory performance of time series is enhanced by applying an LSTM network. The multihead SAM is used to optimize the network, and the role of key features is enhanced by assigning weights with probability to further improve the accuracy of sequence prediction. Finally, by comparison with the existing network model, the offline data generated by the high and low converters in the synthetic ammonia industry is used to predict the CO content so as to verify the superiority and applicability of the proposed network model.

A New Regularized Spatiotemporal Attention-Based LSTM with Application to Nitrogen Oxides Emission Prediction

The data collected from complex process industries are usually time series with considerable nonlinearities and dynamics, as well as excessive redundancy. Moreover, there are temporal and spatial correlations between input variables and key performance variables. These characteristics bring great difficulties to data-driven modeling of the key performance variables. To overcome the problems, a new regularized spatiotemporal attention (STA)-based long short-term memory (LSTM) was developed. First, a standard LSTM network with an STA module was trained to capture the dynamic relationship between input and target variables. Second, the least absolute shrinkage and selection operator was introduced to optimize the STA module. Third, the hyperparameter representing the regularization strength of the algorithm was determined using a moving window cross-validation strategy. Finally, the proposed algorithm was compared to other state-of-the-art algorithms using artificial data, and then it was used to predict the nitrogen oxide emissions of a selective catalytic reduction denitration system. Simulation results showed that the proposed algorithm achieved more accurate predictions than the other algorithms. Furthermore, the statistics and analysis of the importance of the variables are consistent with known chemical-reaction mechanisms and observations of field experts. Thus, the proposed method can provide technical support for the predictive control and optimization of such systems.

Yield and Properties Prediction Based on the Multicondition LSTM Model for the Solvent Deasphalting Process

Solvent deasphalting (SDA) is a complex multiscale continuous process. The operation mode of the SDA process is not considered in the related data-driven model. Therefore, this paper proposes a time lag process prediction model with multiple operation modes to solve the above problem. First, based on random forests, the relative importance of initial input variables in the SDA process on DAO yield and Conradson carbon residual are studied and features are selected according to the results. Then, the stack denoising autoencoder (SDAE) is used to reconstruct the data and obtain the nonlinear mapping information of hidden layers of SDAE and achieve feature dimension reduction. SDAE can improve clustering accuracy of fuzzy c-means, and the operation mode of SDA process is accurately divided. Long short-term memory (LSTM) is used to establish a multicondition LSTM model. Compared with the traditional LSTM model, the multicondition LSTM model has a higher prediction accuracy with R ² > 0.95. The sensitivity analyses of the properties of feed and operating conditions on DAO yield are consistent with the principle of two-phase countercurrent extraction in the SDA process. In addition, the benchmark test of the Tennessee Eastman process shows that the proposed method is also effective in the fault detection of other processes. Because the multicondition LSTM can predict the future process measurement data according to operating mode, it can better avoid the false alarm problem and predict the fault earlier.

Structural Analysis and Classification of Low-Molecular-Weight Hyaluronic Acid by Near-Infrared Spectroscopy: A Comparison between Traditional Machine Learning and Deep Learning

Confusing low-molecular-weight hyaluronic acid (LMWHA) from acid degradation and enzymatic hydrolysis (named LMWHA-A and LMWHA-E, respectively) will lead to health hazards and commercial risks. The purpose of this work is to analyze the structural differences between LMWHA-A and LMWHA-E, and then achieve a fast and accurate classification based on near-infrared (NIR) spectroscopy and machine learning. First, we combined nuclear magnetic resonance (NMR), Fourier transform infrared (FTIR) spectroscopy, two-dimensional correlated NIR spectroscopy (2DCOS), and aquaphotomics to analyze the structural differences between LMWHA-A and LMWHA-E. Second, we compared the dimensionality reduction methods including principal component analysis (PCA), kernel PCA (KPCA), and t-distributed stochastic neighbor embedding (t-SNE). Finally, the differences in classification effect of traditional machine learning methods including partial least squares-discriminant analysis (PLS-DA), support vector classification (SVC), and random forest (RF) as well as deep learning methods including one-dimensional convolutional neural network (1D-CNN) and long short-term memory (LSTM) were compared. The results showed that genetic algorithm (GA)-SVC and RF were the best performers in traditional machine learning, but their highest accuracy in the test dataset was 90%, while the accuracy of 1D-CNN and LSTM models in the training dataset and test dataset classification was 100%. The results of this study show that compared with traditional machine learning, the deep learning models were better for the classification of LMWHA-A and LMWHA-E. Our research provides a new methodological reference for the rapid and accurate classification of biological macromolecules.

Feature Selection Assists BLSTM for the Ultrasensitive Detection of Bioflavonoids in Different Biological Matrices Based on the 3D Fluorescence Spectra of Gold Nanoclusters

Rapid and on-site qualitative and quantitative analysis of small molecules (including bioflavonoids) in biofluids are of great importance in biomedical applications. Herein, we have developed two deep learning models based on the 3D fluorescence spectra of gold nanoclusters as a single probe for rapid qualitative and quantitative analysis of eight bioflavonoids in serum. The results proved the efficiency and stability of the random forest-bidirectional long short-term memory (RF-BLSTM) model, which was used only with the most important features after deleting the unimportant features that might hinder the performance of the model in identifying the selected bioflavonoids in serum at very low concentrations. The optimized model achieves excellent overall accuracy (98-100%) in the qualitative analysis of the selected bioflavonoids. Next, the optimized model was transferred to quantify the selected bioflavonoids in serum at nanoscale concentrations. The transferred model achieved excellent accuracy, and the overall determination coefficient (R²) value range was 99-100%. Furthermore, the optimized model achieved excellent accuracies in other applications, including multiplex detection in serum and model applicability in urine. Also, LOD in serum at nanoscale concentration was considered. Therefore, this approach opens the window for qualitative and quantitative analysis of small molecules in biofluids at nanoscale concentrations, which may help in the rapid inclusion of sensor arrays in biomedical and other applications.