Toward Lightweight Dynamic Convolutional Neural Network Modeling for Soft Sensors.

Area of Science:

• Chemical Engineering
• Artificial Intelligence
• Process Control

Background:

• Soft sensors are crucial for industrial monitoring and quality control.
• Industrial data often exhibits complex nonlinear and correlated behaviors.
• Existing deep learning models like RNNs and LSTMs can be overly complex for limited data scenarios.

Purpose of the Study:

• To propose a novel lightweight dynamic CNN (LDCNN) for soft sensor applications.
• To address the challenge of modeling complex dynamics and nonlinearities with limited training data.
• To develop a more efficient and generalizable soft sensor model.

Main Methods:

• Utilizing 1-D CNNs for time-series data modeling.
• Integrating positional embedding (PE) and simplified temporal attention for enhanced dynamic modeling.
• Employing dilated convolutions and layer normalization (LN) to reduce network complexity and prevent over-parametrization.

Main Results:

• The proposed LDCNN demonstrated superior performance compared to traditional methods on a real industrial case.
• The lightweight network achieved strong generalization capabilities despite limited training samples.
• The LDCNN effectively captured the nonlinear and dynamic behaviors of the industrial process.

Conclusions:

• The LDCNN presents a viable and effective solution for soft sensor development in data-scarce industrial environments.
• Lightweight deep learning models can outperform complex traditional methods for soft sensing.
• The proposed architecture offers a balance between performance and model complexity.