An Improved Normalized Mutual Information Variable Selection Algorithm for Neural Network-Based Soft Sensors.

Area of Science:

• Data Science
• Machine Learning
• Process Engineering

Background:

• Soft sensors are crucial for estimating process variables that are difficult to measure directly.
• Effective variable selection is essential for building accurate and efficient soft sensor models.
• Existing mutual information-based methods can be prone to local optima and redundancy.

Purpose of the Study:

• To develop an advanced variable selection algorithm for soft sensors.
• To improve the accuracy and reduce the complexity of soft sensor models.
• To overcome limitations of existing mutual information feature selection techniques.

Main Methods:

• Integration of Normalized Mutual Information Feature Selection (NMIFS) for identifying influential and non-redundant variables.
• Application of Tabu Search (TS) to prevent the NMIFS algorithm from converging to local optimal solutions.
• Combination of entropy information, mutual information, and Artificial Neural Network (ANN) error validation for robust variable selection.

Main Results:

• The proposed algorithm demonstrated superior performance in variable selection compared to state-of-the-art methods.
• Achieved higher model accuracy with a significantly reduced number of selected variables.
• Validated effectiveness on diverse simulation datasets and real-world power plant production data.

Conclusions:

• The developed NMIFS-TS algorithm offers an effective approach for variable selection in soft sensor applications.
• This method leads to more accurate and parsimonious soft sensor models.
• The algorithm provides reliable and improved results for practical industrial implementation.