[All work](/work?type=work)

# Grid Diffusion AD

I studied whether diffusion models trained on normal magnetic-field signals could identify electrical faults, separating reconstruction quality from the detection behaviour needed in a monitoring system.

Academic research. Model implementation, experiments, evaluation and report.

![Original experiment plot comparing four channels of signal data with their reconstructed waveforms.](https://cdn.sanity.io/images/qnuj1c4o/production/ff23455da7c086b719431afaca10625eb391f5a9-1200x1200.png?w=960\&q=80\&auto=format)

The data was repetitive and multi-channel, and I worked on learning its normal structure with diffusion models before using reconstruction and denoising scores to look for fault-related departures, with each score needing to be tested against the normal variation that the model had already seen.

The pipeline included quantisation, categorical denoising and several scoring variants, with experiments comparing model settings and threshold calibration using normal data, as fault labels should not silently become part of the calibration process.

The report found that good reconstruction did not consistently produce useful detection, and the evaluated operating points still had high false-alarm rates, which made the deployment assumptions part of the result.

The catalogue date follows the first preserved commit on 12 November 2025.

## Outcome

A completed implementation and evaluation study with recorded experiments and a report. The findings limit the detector claims, particularly under normal-only threshold calibration.

[All work](/work?type=work)