High-Dimensional Data Global Sensitivity Analysis Based on Deep Soft Sensor Model.

Area of Science:

• Industrial process modeling
• Data science
• Machine learning

Background:

• High-dimensional data presents computational challenges for sensitivity analysis (SA).
• Existing SA models often lack generalization capacity for complex industrial processes.
• Identifying key process variables influencing outputs is crucial for industrial soft sensor modeling.

Purpose of the Study:

• To develop a novel global sensitivity analysis (GSA) approach for high-dimensional data in industrial soft sensor modeling.
• To address the high computational cost and limited generalization of existing SA techniques.
• To accurately identify the effects of process variables on the quantity of interest (QoI).

Main Methods:

• Proposed a high-dimensional data GSA approach using a deep soft sensor model.
• Developed approximately incremental grouping (AIG) and region-based cooperative co-evolution (RBCC) algorithms for data decomposition.
• Designed a multihead deep soft sensor model with region of interest (RoI) alignment for feature extraction.
• Utilized Monte Carlo dropout (MC-dropout) and a joint loss function for uncertainty analysis and GSA index measurement.

Main Results:

• The proposed approach effectively decomposes high-dimensional data into manageable regions.
• The multihead deep soft sensor model accurately determines GSA indices for each region.
• Experimental results on benchmark and real-world datasets validate the method's effectiveness.
• The approach successfully addresses GSA challenges in high-dimensional industrial processes.

Conclusions:

• The novel GSA approach significantly improves the analysis of high-dimensional industrial data.
• Deep soft sensor models offer a powerful solution for computationally intensive SA tasks.
• The method provides accurate insights into process variable impacts on QoIs, enhancing industrial modeling.