An adaptive multi-fuzzy logic model for diagnosing transformer faults using dynamic weight optimization

Traditional Dissolved Gas Analysis (DGA) methods—such as the Duval Triangle or IEC ratio methods—often struggle with inconsistent results, particularly when dealing with complex, overlapping, or multiple fault conditions. This paper addresses the need for a more robust, flexible diagnostic tool that can integrate multiple DGA techniques while overcoming the limitations of static, rule-based interpretation.

The authors propose an Adaptive Multi-Fuzzy Logic (AMFL) model. Unlike traditional ensemble systems that assign fixed weights to different diagnostic methods, the AMFL model employs a dynamic weight adjustment mechanism. This system iteratively evaluates the diagnostic performance of five classical DGA methods (IEC Ratio, Roger’s Ratio, Doernenburg’s Ratio, Duval’s Triangle, and Key Gas) and recalibrates their influence based on historical fault prediction accuracy. The researchers also refined fuzzy membership functions by introducing intentional overlaps and statistical adjustments to better handle boundary cases where gas concentrations do not clearly fall into a single fault category.

The AMFL model demonstrates significant improvements in diagnostic accuracy, consistency, and reliability compared to traditional fixed-weight multi-fuzzy systems. By utilizing a feedback-based optimization loop, the model effectively manages "Undefined Diagnosis" outcomes and complex, mixed-fault scenarios (such as simultaneous thermal and electrical faults). The study reports that the model maintains high diagnostic accuracy (exceeding 99%) across diverse datasets without requiring massive or perfectly balanced training sets, making it a highly adaptable solution for real-world transformer condition monitoring.

Power transformers are critical infrastructure assets; unexpected failures lead to significant economic losses and operational hazards. By providing a more nuanced, adaptive, and interpretable diagnostic framework, this research supports better asset management decisions. The ability of the AMFL model to handle uncertainty and overlapping fault signatures makes it a practical, industry-ready tool for utility operators seeking to transition from reactive maintenance to proactive, condition-based monitoring.