Project background
Distinguishing fault types — phase-to-ground, phase-to-phase, transient vs. sustained — quickly after an event is critical for dispatch decisions. The client wanted automated classification rather than relying on engineer interpretation.
Challenge
Training on an imbalanced set of real fault waveforms, avoiding overfitting to site-specific noise, and producing classifications fast enough to feed dispatch workflows.
Approach & solution
We combined DSP feature extraction with ML classification, trained against a curated waveform dataset. Models were calibrated per substation where needed, with a fallback to generalized models for new sites. Every classification ships with its supporting features for engineer review.
Results & benefits
Classification accuracy now exceeds manual triage speed with comparable correctness on common fault types. Dispatchers use the output as a starting point rather than generating it themselves, freeing engineering attention for harder cases.
