A new diagnostic tool catches hidden clock drift and sensor desync in time-series AI data - without touching the labels or the model.
Researchers built a classifier- and label-free diagnostic based on minimum description length (MDL) to detect when multichannel sensor streams used for time-series classification have drifted out of sync. The method shifts groups of sensor channels against each other and measures how compactly one group can be encoded from the rest; a shift that destroys shared structure makes the encoding longer, while the shift with the shortest code length marks the best alignment. Because it needs neither retraining nor a trusted reference signal, it can check both training data and live deployment data. Tested on two synthetic tasks and nine real-world datasets, it recovered accuracy that had been lost to deliberately introduced drift.
More telling: auditing four widely used benchmarks - FordChallenge, Opportunity, PAMAP2, and UCIActivity - turned up stable, nonzero misalignment that standard accuracy metrics never flagged. That means some published time-series classification results may already be quietly skewed by sync errors nobody checked for. Any wearable, industrial sensor rig, or multi-device setup where clocks can drift has the same blind spot.
Benchmarks are supposed to be the control group. If their timestamps are already off, the field has been grading itself on a shifted curve.