Soft sensor for nonuniform sampling nonlinear dynamic process using irregular-time-interval latent probabilistic...

Area of Science:

• Chemical Engineering
• Data Science
• Machine Learning

Background:

• Dynamic latent variable (DLV) models are crucial for industrial soft sensing but struggle with nonlinear dynamics and nonuniform data.
• Conventional DLV models are limited to linear feature extraction and exhibit poor performance with irregularly sampled data.

Purpose of the Study:

• To propose an advanced soft sensor, the irregular-time-interval latent probabilistic predictability embedding supervised deep network (ILPPSDN), for nonlinear dynamic processes with nonuniform sampling.
• To enhance feature predictability and model latent temporal dependencies in soft sensing applications.

Main Methods:

• Developed an ILPPSDN incorporating a prediction regularization term in the autoencoder's decoding loss.
• Utilized an ordinary differential equation network to parameterize the internal state derivative within a variational recurrent neural network.
• Implemented unified training for all network components and employed pre-training and supervised fine-tuning for soft sensor development.

Main Results:

• The ILPPSDN achieved significant reductions in root mean square error (RMSE) across various uneven sampling ratios in debutanizer and sulfur recovery units.
• Demonstrated RMSE reductions of at least 21.1% to 26.1% in industrial case studies.
• Ablation studies confirmed the proposed method's effectiveness, reducing RMSE by at least 5% and 6% in respective industrial cases.

Conclusions:

• The proposed ILPPSDN offers a robust solution for soft sensing in nonlinear dynamic processes with nonuniformly sampled data.
• This deep learning approach enhances feature predictability and temporal dependency modeling, leading to superior performance over conventional methods.
• The ILPPSDN demonstrates practical applicability and significant improvements in industrial process monitoring and control.