Description
The progressive degradation of a sensing model's accuracy over time after deployment, even in the same physical environment, due to slow accumulation of micro-changes (furniture nudged, antennas reoriented, AP reseated, humidity cycle, equipment aging). Calibration drift is the temporal cousin of environment-dependence: not "I trained in room A and deploy in room B" but "I trained in room A on Monday and the same room is subtly different on Friday". This is the central operational pain point that the thesis's periodic BLE calibration campaigns are designed to detect and correct in a closed loop.
Why it's hard
- Drift is incremental and silent — there is no error signal during routine inference.
- Detecting drift typically requires re-acquisition of ground truth, which is the expensive step we want to avoid.
- Distinguishing genuine drift from a transient anomaly is statistically delicate.
- The cadence of recalibration is application-specific and rarely studied empirically.
- Drift compounds across all the failure modes of environment-dependence.
Common approaches
- Periodic calibration campaigns with a complementary sensor (BLE in the thesis approach) producing fresh ground truth.
- Online learning / continual learning to adapt incrementally.
- Drift-detection statistics on intermediate representations (mean / variance shift, KS tests).
- Conformal prediction / uncertainty estimates that grow as drift accumulates.
- Self-supervised pretexts that double as drift indicators.