An interpretable fuzzy inference framework for transformer health assessment and remaining useful life estimation

Reliable transformer health assessment and remaining useful life (RUL) estimation are essential for maintaining the stability and efficiency of modern power systems. This study proposes an interpretable, standards-informed framework that integrates a fuzzy inference system with a nonlinear health-index-based RUL estimation mechanism for transformer condition and risk-state assessment under real-world operating conditions. Thermal and electrical measurements, including oil temperature, ambient temperature, oil level, and three-phase currents, were utilized together with derived features such as average load and current imbalance, and were processed by a fuzzy inference system, comprising an eighteen-rule, cascading-exclusion rule base grounded in relevant engineering standards where such guidance exists, to generate physically interpretable health scores categorized into five discrete health classes. Rather than relying on a simple linear interpolation between the health score and lifetime interval, the proposed framework maps the health scores to a continuous RUL estimate through a geometric baseline interpolation combined with an empirically calibrated thermal modulation term, referenced to the transformer’s own observed operating temperature, thereby introducing a nonlinear temperature-dependent adjustment relative to the transformer’s observed operating baseline. The resulting RUL estimates were further categorized into four operational risk levels: Critical, High Risk, Warning, and Normal. Because run-to-failure data and verified end-of-life records were not available for the studied transformer, these RUL values were treated as fuzzy- and literature-derived reference estimates rather than observed ground-truth values. Comparison with the available SCADA protection signals showed a statistically significant association between the fuzzy-derived health states and alarm occurrence ( 𝜒 2 = 451.47, 𝑝 < 0 . 0 0 1 , 𝑁 = 18,054), with an AUC-ROC of 0.77. However, because the thermal and oil-level alarm signals share underlying measurement channels with the fuzzy inference system, these associations were interpreted primarily as consistency evidence rather than independent validation. More importantly, 143 of the 181 observations classified as ‘Very Poor’ occurred without an active SCADA alarm and were instead characterized by elevated current imbalance, demonstrating the framework’s ability to identify electrically driven abnormal conditions not captured by the available hardware alarm channels. Parameter-level sensitivity analysis revealed a non-uniform and data-distribution-dependent response to membership-function and health-boundary perturbations, with particularly strong sensitivity at the Poor/Regular decision boundary. The proposed framework provides an interpretable, standards-informed basis for transformer condition monitoring and health-index-based risk assessment, whereas direct validation of the lifetime estimates against run-to-failure or verified end-of-life data remains an important direction for future work.