Explainable AI (XAI)

A game-theoretic approach that assigns each input feature an importance value for a particular prediction. SHAP values quantify the marginal contribution of each dissolved gas—such as acetylene (C₂H₂) or hydrogen (H₂)—to a fault classification output.

• Based on Shapley values from cooperative game theory
• Provides both global (model-level) and local (prediction-level) interpretability
• Example: A SHAP waterfall plot reveals that elevated ethylene (C₂H₄) contributed +0.42 to a thermal fault prediction while methane (CH₄) contributed -0.15
• Guarantees consistency: if a feature's contribution increases, its SHAP value never decreases
• Computationally intensive for high-dimensional data; TreeSHAP variant optimizes for ensemble models like XGBoost

A perturbation-based method that explains individual predictions by approximating the complex model locally with an interpretable surrogate model. LIME generates synthetic samples around a specific transformer fault instance and fits a simple linear model to reveal which gas ratios drove the classification.

• Model-agnostic: works with any black-box classifier including neural networks and gradient-boosted trees
• Creates local fidelity by weighting perturbed samples by proximity to the original instance
• Example: For a partial discharge alert, LIME identifies that the hydrogen-to-methane ratio was the dominant factor
• Limitations include instability—repeated runs on the same instance may yield different explanations
• Best used for ad-hoc spot checks rather than systematic feature attribution across a fleet

A global visualization technique that shows the marginal effect of one or two features on the predicted outcome, averaged over the distribution of all other features. PDPs reveal how fault probability changes as acetylene concentration increases from 0 to 50 ppm.

• Displays monotonic, non-linear, or threshold relationships between gas levels and failure modes
• Example: A PDP for dissolved hydrogen shows a sharp increase in arcing fault probability above 100 ppm
• Assumption of independence: PDPs can be misleading when features are strongly correlated (e.g., ethane and ethylene in thermal faults)
• Accumulated Local Effects (ALE) plots offer a correlation-robust alternative
• Essential for sanity-checking that learned relationships align with domain knowledge from IEC 60599 and Duval Triangle heuristics

A gradient-based attribution method designed for deep learning models that satisfies the completeness axiom—attributions sum to the difference between the prediction and a baseline. It computes the path integral of gradients as inputs transition from a neutral baseline to the actual gas concentration values.

• Particularly suited for LSTM and Temporal Fusion Transformer architectures used in time-series DGA forecasting
• Requires careful selection of a baseline (e.g., zero gas concentration or historical mean)
• Example: For a transformer hot-spot temperature prediction, Integrated Gradients attributes 62% of the increase to load current and 28% to ambient temperature
• Satisfies sensitivity: if a feature differs from baseline and changes the output, it receives non-zero attribution
• More computationally efficient than SHAP for deep networks but limited to differentiable models

A method that generates minimal changes to input features that would flip a model's prediction to a desired alternative outcome. For transformer diagnostics, counterfactuals answer: "What dissolved gas levels would have resulted in a normal classification instead of a thermal fault alarm?"

• Produces actionable insights: "Reduce acetylene from 15 ppm to below 5 ppm to avoid arcing classification"
• Uses optimization-based approaches like Wachter et al. or genetic algorithms to find minimal perturbations
• Example: A counterfactual reveals that lowering ethylene by 8 ppm while holding other gases constant would reclassify a T3 fault as normal operation
• Must respect feasibility constraints—some gas combinations are physically impossible in transformer oil
• Complements SHAP by showing what to change rather than just what contributed

A rule-based explanation method that identifies high-precision if-then rules (anchors) that sufficiently guarantee a prediction regardless of other feature values. Unlike LIME's local approximations, anchors provide coverage statistics showing how broadly each rule applies.

• Generates explanations like: "If acetylene > 10 ppm AND ethylene/ethane ratio > 3, then arcing fault is predicted with 97% precision"
• Precision measures how often the rule holds; coverage measures how many instances satisfy the condition
• Example: An anchor for cellulose degradation identifies that carbon monoxide > 500 ppm AND furan content > 2 ppm guarantees the classification for 89% of cases
• Particularly valuable for audit trails in regulated utility environments
• Builds on the anchoring concept from Ribeiro et al. (2018) as a successor to LIME

A game-theoretic approach to model interpretability that assigns each input feature an importance value for a particular prediction. SHAP values are computed by averaging the marginal contribution of a feature across all possible feature subsets, ensuring local accuracy and consistency.

• Based on Shapley values from cooperative game theory
• Provides both global feature importance and local per-prediction explanations
• Commonly visualized via waterfall plots and summary bar charts
• Example: A SHAP analysis of a DGA fault model may reveal that acetylene concentration contributed 62% to an arcing classification while hydrogen contributed only 8%

A perturbation-based technique that explains individual predictions by approximating the complex model locally with an interpretable surrogate model, typically a sparse linear model or decision tree. LIME generates synthetic samples around the instance of interest and weights them by proximity.

• Model-agnostic: Works with any black-box classifier or regressor
• Produces explanations showing which features pushed the prediction toward or away from a specific class
• Limitations include instability across repeated runs and sensitivity to the neighborhood definition
• Example: For a transformer thermal fault alert, LIME might show that top-oil temperature rise rate and load current were the dominant local drivers

A global interpretability method that visualizes the marginal effect of one or two features on the predicted outcome of a machine learning model, averaged over the distribution of all other features. PDPs reveal whether the relationship is linear, monotonic, or non-linear.

• Computed by forcing the feature of interest to take a fixed value across all instances and averaging predictions
• Assumption: The feature of interest is independent of other features (violated in correlated sensor data)
• Often paired with Individual Conditional Expectation (ICE) plots to show heterogeneity
• Example: A PDP for methane-to-hydrogen ratio may show a sharp increase in predicted fault probability above a threshold of 0.8

A form of explanation that answers the question: 'What would need to change in the input for the model to produce a different output?' Counterfactuals identify the minimal set of feature perturbations required to flip a prediction from one class to another.

• Generates actionable insights: 'If acetylene were below 5 ppm, the model would not have flagged an arcing fault'
• Computed via optimization methods like Wachter's algorithm or genetic algorithms
• Must satisfy constraints: sparsity (few changes), feasibility (realistic values), and diversity (multiple paths)
• Critical for maintenance planning: tells engineers exactly which sensor readings to monitor for reversal

An attribution method designed specifically for deep neural networks that satisfies the sensitivity and implementation invariance axioms. It computes feature importance by accumulating gradients along a straight-line path from a baseline input (e.g., zero or historical average) to the actual input.

• Overcomes the saturation problem where gradients of saturated neurons approach zero
• Requires a meaningful baseline; for transformer DGA, the baseline might be normal operating gas levels
• Produces pixel-level or feature-level attribution maps
• Example: In a 1D-CNN trained on time-series DGA data, Integrated Gradients can highlight the exact time window where gas ratios began deviating from normal

A high-precision rule-based explanation method that identifies sufficient conditions for a prediction. An anchor is a set of feature predicates that 'anchors' the prediction locally—meaning changes to other features do not alter the outcome as long as the anchor conditions hold.

• Extends LIME by providing if-then rules with formal coverage and precision guarantees
• Example anchor: 'If ethylene exceeds 50 ppm AND the ethylene-to-acetylene ratio exceeds 3, then the model predicts a thermal fault with 98% precision'
• Rules are intuitive for domain experts who already use IEC 60599 gas ratios for manual diagnosis
• Computationally expensive to construct due to the need for perturbation-based statistical validation