Supervised Multi-Layer Conditional Variational Auto-Encoder for Process Modeling and Soft Sensor.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Process Engineering

Background:

• Variational Auto-Encoders (VAEs) are valuable for process modeling, offering deep feature extraction and noise robustness.
• Supervised VAEs face challenges in stable and controllable data generation, hindering prediction performance.
• Existing methods suffer from unstable outputs due to random latent subspace resampling.

Purpose of the Study:

• To develop a novel multi-layer conditional variational auto-encoder (M-CVAE) for enhanced supervised process modeling.
• To improve the stability and controllability of generated data in VAEs by injecting label information.
• To enable accurate online quality prediction in industrial processes.

Main Methods:

• Constructed a multi-layer conditional variational auto-encoder (M-CVAE) by incorporating label information into the latent subspace.
• Utilized label information as input alongside process variables to strengthen input-output correlations.
• Embedded a neural network layer within the encoder for online quality prediction.

Main Results:

• The M-CVAE demonstrates controlled output generation towards actual values, enhancing prediction accuracy.
• Strengthened correlation between input and output variables through integrated label information.
• Successfully validated the method's superiority and effectiveness on two real industrial process cases.

Conclusions:

• The proposed M-CVAE effectively addresses limitations of existing supervised VAEs in process modeling.
• The method provides stable, controllable data generation and accurate online quality prediction.
• M-CVAE shows significant advantages over other methods in industrial process applications.