Grid Diffusion AD
Diffusion models trained on normal magnetic-field signals, scored on whether they notice a fault they've never seen.
Academic research. Model implementation, experiments, evaluation and report.
The data is repetitive and multi-channel, which is what makes normal-only training plausible at all. Reconstruction and denoising scores both had to be tested against the variation the model had already seen, and the thresholds were calibrated on normal data so fault labels never crept into training.
The report keeps the false alarm side visible. A detector that flags everything isn't a detector, and that's the failure mode this kind of work falls into.
State
An implementation, recorded experiments and a report. The claims stay narrow under normal-only training.