Failure Mode Classification

Failure mode classification relies on labeled historical datasets where each fault event has been verified through physical inspection or laboratory analysis. Common algorithms include:

• Random Forest and XGBoost for tabular DGA data with engineered gas ratios
• Support Vector Machines (SVM) for high-dimensional feature spaces with clear margin separation
• Multi-layer Perceptrons for capturing non-linear relationships between gas concentrations and fault types

The quality of classification depends directly on the accuracy and completeness of ground-truth labels obtained during transformer teardowns or repairs.

The international standard IEC 60599 defines the canonical fault categories that classification models must predict:

• Partial Discharge (PD): Low-energy discharges in gas-filled voids, characterized by high hydrogen and methane
• Thermal Faults (<300°C, 300-700°C, >700°C): Overheating of cellulose or oil, indicated by ethylene and methane ratios
• Electrical Faults (D1, D2): Low-energy and high-energy arcing, marked by acetylene production

Models must distinguish between thermal faults in oil versus paper, as the latter indicates critical cellulose degradation.

Raw gas concentrations are rarely used directly. Instead, diagnostic ratios serve as the primary features:

• Duval Triangle ratios: %CH₄, %C₂H₄, %C₂H₂ for thermal/electrical discrimination
• Rogers ratios: C₂H₂/C₂H₄, CH₄/H₂, C₂H₄/C₂H₆ for fault severity staging
• Doernenburg ratios: Five-ratio method requiring minimum gas concentration thresholds

Lag features and rolling statistics from online DGA monitors add temporal context, enabling models to detect evolving faults before gas concentrations exceed alarm thresholds.

Real-world transformer fleets exhibit severe class imbalance—thermal faults dominate while dangerous arcing faults are rare. Mitigation strategies include:

• Synthetic Minority Oversampling (SMOTE) to generate synthetic arcing fault examples in feature space
• Cost-sensitive learning that penalizes misclassification of catastrophic faults more heavily than minor thermal anomalies
• Stratified k-fold cross-validation to ensure rare fault types appear in every training and validation fold

Without imbalance correction, models achieve high overall accuracy while missing the most critical failure modes.

Classification outputs must be interpretable to justify maintenance decisions. Techniques applied to transformer fault models include:

• SHAP (SHapley Additive exPlanations): Quantifies each gas's contribution to a specific fault prediction
• LIME (Local Interpretable Model-agnostic Explanations): Generates local surrogate models explaining individual predictions
• Decision tree visualization: For ensemble methods, extracting the most influential decision paths

Explainability bridges the gap between black-box neural networks and the engineering judgment required to authorize a $500,000 transformer outage.

Failure mode classification is not a standalone exercise—it feeds directly into Condition-Based Maintenance (CBM) workflows:

• Classification triggers specific inspection protocols (e.g., acetylene detection prompts immediate internal inspection)
• Fault type determines urgency scoring within Health Index calculations
• Outputs are mapped to IEC 61850 MMS logical nodes for SCADA alarming and automated load shedding

The classification model serves as the diagnostic brain within a broader predictive maintenance architecture that includes RUL estimation and Digital Twin synchronization.

A graphical diagnostic tool that plots relative proportions of %CH₄, %C₂H₄, and %C₂H₂ to classify faults into six zones:

• D1: Low energy discharge
• D2: High energy discharge (arcing)
• T1: Thermal fault < 300°C
• T2: Thermal fault 300-700°C
• T3: Thermal fault > 700°C
• PD: Partial discharge

Machine learning classifiers often use Duval Triangle zone assignments as categorical features or validate model predictions against this established diagnostic framework.

Modern failure mode classification leverages ensemble methods to improve accuracy over single-algorithm approaches:

• Random Forest: Handles non-linear gas ratio relationships and provides feature importance rankings
• XGBoost: Excels with imbalanced datasets where certain fault types are rare
• Stacking: Combines multiple base classifiers with a meta-learner for final prediction

These methods typically achieve >95% accuracy on labeled DGA datasets, significantly outperforming traditional ratio-based methods. Feature engineering includes gas ratios, rate-of-change calculations, and load-correlated gas levels.

Asset managers require interpretable justifications for automated fault classifications. XAI methods applied to transformer diagnostics include:

• SHAP (SHapley Additive exPlanations): Quantifies each gas's contribution to a specific fault classification
• LIME (Local Interpretable Model-agnostic Explanations): Generates local approximations to explain individual predictions
• Partial Dependence Plots: Visualize how acetylene levels influence arcing probability

These techniques bridge the gap between black-box neural networks and the IEC 60599 diagnostic logic that reliability engineers trust.

An emerging approach that embeds thermodynamic governing equations directly into the neural network loss function. For transformer fault classification, PINNs constrain predictions to physically plausible failure modes:

• Heat transfer equations prevent thermal fault predictions inconsistent with load and ambient temperature
• Gas generation rate laws ensure predicted gas trajectories follow Arrhenius reaction kinetics
• Dielectric breakdown physics constrain partial discharge classifications

This hybrid approach reduces false positives by 40-60% compared to purely data-driven models, particularly when training data is limited.

Unsupervised neural networks trained to reconstruct normal transformer operational data. Failure mode classification integrates autoencoders for:

• Novelty detection: Identifying fault signatures not present in training data
• Pre-classification filtering: Flagging anomalous conditions before supervised classification
• Degradation trajectory mapping: Tracking how reconstruction error increases as faults develop

When reconstruction error exceeds a dynamic threshold (typically 3σ above baseline), the sample is routed to the supervised classifier for specific fault type identification.