Same Gas, Four Diagnoses

Power transformers are the most expensive and least replaceable components of the electrical grid, with lead times of 12–24 months for large units. Dissolved gas analysis (DGA) — measuring gases like hydrogen, methane, ethylene, and acetylene dissolved in transformer oil — is the primary diagnostic technique for detecting incipient faults before catastrophic failure. However, the major diagnostic interpretation methods (Duval Triangle, Rogers Ratios, IEC 60599, IEEE C57.104 key gas method) frequently give conflicting diagnoses for the same gas data, leaving utilities uncertain whether to de-energize an expensive asset or continue operating with monitoring.

The US power transformer fleet averages 40+ years old, and replacement of a large power transformer costs $3–10M. A single transformer failure can cause widespread outages, with repair costs exceeding $50M when fire, environmental cleanup, and lost revenue are included. The installed base of aging transformers is growing (new installations haven't kept pace with demand), making predictive maintenance increasingly critical. Yet when DGA results indicate possible faulting, utilities face an impossible dilemma: the cost of unnecessary de-energization (planned outage, replacement power procurement) is enormous, but so is the cost of a preventable failure.

Each diagnostic method maps gas ratios to fault types (partial discharge, low-energy sparking, high-energy arcing, thermal fault, cellulose degradation), but they use different gases, different ratio boundaries, and different classification logic. Duval Triangle uses three gases and graphical boundaries; Rogers Ratios use four gas ratios with numerical cutoffs; IEEE C57.104 uses absolute gas concentrations with population-based thresholds. Agreement between methods ranges from 50–75% — meaning a quarter to half of diagnoses are contradictory. Machine learning approaches trained on DGA databases show improved accuracy but require labeled training data (confirmed fault type after transformer inspection or teardown), which is scarce because utilities rarely open transformers to verify diagnoses. Online DGA monitors provide trend data but don't resolve the interpretation ambiguity — they just generate conflicting diagnoses more frequently.

A unified diagnostic framework that integrates multiple gas signatures, gas generation rates (trends), transformer load history, and design-specific factors (oil type, insulation class, cooling method) could resolve the contradictions between methods. Building a validated, open DGA-fault-type database — analogous to medical imaging databases for AI training — with confirmed fault types from transformer teardowns would provide the ground truth needed for supervised learning. Alternatively, complementary diagnostic techniques (frequency response analysis, partial discharge monitoring, infrared thermography) could be systematically combined with DGA to reduce diagnostic uncertainty.

A team could compile DGA data from publicly available utility reports and systematically compare diagnoses from all major methods, quantifying disagreement rates by fault type. Alternatively, a team could develop a Bayesian fusion framework that combines diagnoses from multiple methods (weighted by their historical accuracy for each fault type) and test whether the combined diagnosis outperforms any single method. Skills: power engineering, statistical analysis, Bayesian inference, data science.