AI-Powered FRA Diagnostics: Using Machine Learning to Classify Winding Faults

Traditional Transformer Frequency Response Analyzer interpretation relies on statistical indices (CC, SDR, ASLE) and engineer experience. While effective, these methods can miss subtle patterns or generate false positives. Machine learning—particularly convolutional neural networks (CNNs) and support vector machines (SVMs)—offers automated pattern recognition that classifies fault types and severity with accuracy approaching expert human interpretation. This article presents the state of AI for FRA diagnostics.

Why Machine Learning for FRA?

FRA signatures are high-dimensional data (500–5000 frequency points per measurement). ML excels at:

• Detecting non-linear patterns that statistical indices miss (e.g., specific notch shapes characteristic of spacer displacement vs. radial buckling).
• Classifying fault type from a single measurement without baseline (using features extracted from the signature shape).
• Fusing FRA with other diagnostic data (DGA, TTR) for multi-modal classification.
• Reducing false positives by learning instrument-specific or site-specific noise patterns.

Feature Engineering for FRA Signatures

Before applying ML, extract relevant features:

• Spectral features: Position, amplitude, and width of the first 10 resonant peaks.
• Shape features: Slope between peaks, roll-off rate above 1 MHz, low-frequency plateau level.
• Statistical features: CC, SDR, ASLE computed against a reference (if available).
• Wavelet features: Decompose signature into time-frequency representations using discrete wavelet transform (DWT).

For deep learning, raw FRA amplitude and phase vectors can be fed directly into a 1D CNN.

Supervised Learning for Fault Classification

Train classifiers on labeled FRA datasets:

• Normal vs. fault (binary classification): SVM with RBF kernel achieves >95% accuracy on balanced datasets.
• Fault type (multi-class): Distinguish between axial displacement, radial buckling, shorted turns, core movement, and lead damage. CNN architectures (ResNet-18 adapted for 1D signals) achieve 85–90% accuracy.
• Severity regression: Predict the degree of displacement (e.g., 5 mm vs. 15 mm) using random forest regression.

Case Example: CNN-Based Fault Classification on Field Data

A research team trained a 1D CNN on 2,000 FRA measurements from laboratory-induced faults (axial displacement, radial buckling, shorted turns) and 2,000 normal signatures. The model achieved:

• Normal vs. any fault: 98% accuracy
• Fault type classification: axial displacement 92%, radial buckling 88%, shorted turns 96%, core movement 85%
• Severity estimation: ±2 mm for axial displacement (R² = 0.91)

When tested on 50 field transformers with known fault histories (validated by internal inspection), the CNN correctly classified 47 of 50 (94% accuracy). In two misclassifications, the FRA deviation was due to temperature variation (not damage)—the model lacked temperature metadata. Adding temperature as an input feature improved accuracy to 96%.

Unsupervised Learning for Anomaly Detection

When labeled fault data is scarce, use unsupervised methods:

• Autoencoders: Train on normal FRA signatures; reconstruction error > threshold indicates anomaly.
• One-class SVM: Learn the boundary of normal signatures; points outside are anomalous.
• Isolation Forest: Isolates anomalies by random partitioning; effective for high-dimensional FRA data.

Unsupervised methods detect novel fault types not seen in training, making them suitable for fleet-wide screening.

Practical Implementation for Utilities

Deploy ML-based FRA analysis in stages:

1. Curate a labeled dataset from your fleet (minimum 200 normal, 50 fault signatures).
2. Train a binary classifier on your instrument's specific frequency range and resolution.
3. Integrate the model into the FRA software or cloud platform.
4. Use model output as a second opinion alongside traditional statistical indices.
5. Periodically retrain with new field data to adapt to instrument drift or new fault types.

Limitations and Challenges

AI for FRA faces several hurdles:

• Data scarcity: Few publicly available FRA fault datasets; utilities must generate their own.
• Instrument variability: Models trained on one FRA instrument may not generalize to another. Use domain adaptation techniques or train on multiple instruments.
• Explainability: Neural networks are black boxes; techniques like SHAP or LIME can highlight which frequency bands drove the classification.

Artificial intelligence does not replace human engineering judgment but augments it. A Transformer Frequency Response Analyzer coupled with machine learning can screen hundreds of transformers, flagging only the most suspicious for expert review—enabling predictive maintenance at fleet scale.