A multidimensional reliability-enhanced belief rule base model for fault diagnosis

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Fault diagnosis is critical for industrial equipment safety and maintenance efficiency.
• Belief Rule Base (BRB) excels in knowledge representation and uncertainty reasoning for complex fault diagnosis.
• Traditional BRBs face limitations in multi-class scenarios, including rule-scale expansion and inadequate reliability utilization.

Purpose of the Study:

• To propose a novel reliability-enhanced one-versus-rest BRB (RE-OvR-BRB) method for multi-class fault diagnosis.
• To address limitations of traditional BRBs in handling rule complexity and incorporating reliability information.
• To improve the accuracy, confidence, and robustness of fault diagnosis in industrial settings.

Main Methods:

• Implemented a one-versus-rest (OvR) strategy to decompose multi-class problems into binary subtasks using parallel BRB submodels.
• Developed a multidimensional reliability modeling scheme incorporating feature-level and submodel output-level reliability into inference and decision fusion.
• Introduced a decision fusion mechanism that leverages submodel output reliability to enhance multi-class decision robustness.

Main Results:

• The RE-OvR-BRB method demonstrated superior performance on the CWRU bearing and UConn gear datasets.
• Achieved higher classification accuracy compared to existing methods.
• Showcased improved decision confidence and enhanced noise robustness in fault diagnosis tasks.

Conclusions:

• The proposed RE-OvR-BRB method effectively overcomes the limitations of traditional BRBs for multi-class fault diagnosis.
• The integration of multidimensional reliability significantly enhances diagnostic performance.
• The method proves effective for complex industrial equipment fault diagnosis, improving safety and maintenance efficiency.