Improved Random Forest Algorithm Based on Decision Paths for Fault Diagnosis of Chemical Process with Incomplete Data

Area of Science:

• Chemical Engineering
• Data Science
• Industrial Process Control

Background:

• Data-driven fault detection and diagnosis (FDD) is crucial for industrial processes.
• Existing FDD methods struggle with accuracy when data is missing.
• Big data analytics has increased the need for robust FDD techniques.

Purpose of the Study:

• To propose an improved random forest (RF) based FDD method, DPRF, that effectively compensates for incomplete data.
• To enhance the reliability of FDD in the presence of missing sensor readings.
• To demonstrate the superiority of DPRF over existing methods in handling data anomalies.

Main Methods:

• Developed a Decision Path Random Forest (DPRF) model incorporating correction coefficients.
• Utilized intact training samples to build decision trees within the RF.
• Inferred sample reliability scores from decision paths and node importance for each tree.
• Implemented a majority voting system combining predictions and reliability scores.

Main Results:

• The DPRF model demonstrated superior performance in fault detection and diagnosis with incomplete data.
• Tested on the Tennessee Eastman (TE) process, DPRF showed enhanced accuracy compared to other FDD methods.
• Correction coefficients effectively compensated for the influence of missing data points.

Conclusions:

• The proposed DPRF method offers a robust and accurate solution for FDD with missing data.
• DPRF significantly improves the reliability of fault diagnosis in industrial big data scenarios.
• This approach provides a valuable tool for maintaining operational integrity in complex industrial systems.