Predictive Modeling With Multiresolution Pyramid VAE and Industrial Soft Sensor Applications.

Area of Science:

• Industrial Process Control
• Data-Driven Modeling
• Machine Learning

Background:

• Industrial processes often have discrepant sampling rates due to instrument limitations and varying measurement demands.
• This discrepancy poses a significant challenge for developing accurate data-driven soft sensors.

Purpose of the Study:

• To propose a novel multiresolution pyramid variational autoencoder (MR-PVAE) predictive model.
• To effectively address the multirate modeling issue in data-driven soft sensor development.

Main Methods:

• A multirate data filter was designed using a resolution searching strategy to create a multiresolution dataset.
• A pyramid variational autoencoder (PVAE) was developed for deep nonlinear feature extraction across different resolutions.
• An augmented feature pyramid was constructed layer-by-layer to fuse features from low to high resolutions.

Main Results:

• The MR-PVAE model effectively extracts and fuses features from data with discrepant sampling rates.
• The layer-by-layer feature enhancement led to gradual improvements in soft sensing model prediction accuracy.
• An optimized training strategy was established to select the best feature pyramid for prediction.

Conclusions:

• The proposed MR-PVAE model demonstrates effectiveness and superiority in handling multirate data for soft sensor development.
• The model fully utilizes process information from data with discrepant sampling rates.
• Validation through numerical experiments and an industrial case confirms the model's practical applicability.