Multiscale Convolutional Stochastic Configuration Network Soft Sensor Modeling Method.

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Industrial process modeling faces challenges due to multiscale spatiotemporal coupling.
• Existing soft sensor methods struggle with complex, interconnected process dynamics.

Purpose of the Study:

• To propose a novel soft sensor method to address multiscale spatiotemporal coupling in industrial processes.
• To enhance the accuracy and adaptability of soft sensors in complex industrial environments.

Main Methods:

• Developed a multiscale convolutional stochastic configuration network (MSC-SCN).
• Incorporated parallel multiscale feature extractors with incremental learning.
• Utilized cross-scale feature fusion for integrating diverse feature maps.
• Optimized output weights using low-rank matrix approximation and regularization.

Main Results:

• The MSC-SCN method demonstrated superior performance compared to state-of-the-art techniques.
• Achieved high adaptability in handling multiscale spatiotemporal coupling.
• Experimental validation on three industrial soft sensor tasks confirmed effectiveness.

Conclusions:

• The proposed MSC-SCN offers a robust solution for industrial process modeling with multiscale spatiotemporal coupling.
• This approach enhances soft sensor capabilities for complex industrial applications.
• The method shows significant potential for improving process monitoring and control.