Domain adaptation · UCI gas sensor array

Gas classification that survives sensor drift.

Sensor arrays drift for months. CORAL realigns the distribution and residualization strips dose bias, holding accuracy where baselines collapse.

sensor manifold · 16 channels

Two shifts. Two fixes.

Temporal drift and concentration shift both wreck a sensor classifier, but for opposite reasons. Align the wrong statistic and accuracy gets worse, not better.

Temporal drift, aligned with CORAL
Source · batches 1-5 Target · batch 10
CORAL alignment
50.7%

target-domain accuracy

No alignmentCovariances matched

Drifted. The target domain's second-order statistics no longer match the source, collapsing the classifier.

Train batches 1-5, test batch 10. Sensors age over months.

CORAL aligns the covariance across time, so the decision boundary transfers. Accuracy climbs from 47% to 79.9%.

Baseline MLP47.0%
CORAL + Robust MLP79.9%

Results that hold under drift.

Every number below comes from the held-out evaluation: train on early batches, test on the drifted batch the model never saw.

0.0%
Severe-drift accuracy
CORAL + Robust MLP
+0 pts
Gain over baseline
47% → 79.9%
0.0%
Concentration accuracy
Residualization
0.0%
MLP over Transformer
vs 64.3% on tabular

Temporal drift across severity

CORAL + Robust MLP dominates under severe aging. Stacking edges ahead only when drift is mild, an honest trade-off the system exposes rather than hides.

Architecture: MLP vs Transformer

Improvement trajectory

01

Different shifts, different fixes

Temporal drift needs covariance alignment (CORAL). Concentration shift needs dose-bias removal (residualization). One method cannot do both.

02

Alignment beats ensembles under severe drift

CORAL + Robust MLP reaches 79.9% where a Random-Forest/SVM stack manages 65.4%, a 14.5-point gap.

03

MLP inductive bias wins on tabular data

At ~13k samples the Robust MLP scores 78.3% against a sensor-tokenized Transformer's 64.3%.

Probe the model live.

Pick a gas, set the dose, then turn up sensor drift and watch the 16-channel fingerprint distort until the classifier breaks.

Target gas

Concentration

Dose level1.0 ppm
0.15.0

Sensor drift

Aging0 months
FreshSevere
LiveStep 0Re-samples 2.5s

Sensor array fingerprint

16-channel response signature. Each gas paints a distinct shape.

Response magnitude

How a reading survives the pipeline.

From raw conductance to a drift-proof verdict, each stage targets one failure mode the previous stage cannot fix.

Built on the UCI Gas Sensor Array Drift dataset: thousands of recordings captured as the hardware physically aged, the canonical benchmark for distribution shift in chemical sensing.

~13k
recordings
16
sensors
6
gases
36mo
of aging
  1. 01

    Capture

    Sixteen metal-oxide sensors log conductance across 10 batches spanning 36 months of continuous operation.

  2. 02

    Detect the drift

    A t-SNE projection shows early and late batches pulling apart while the six gas classes stay internally coherent.

  3. 03

    Align with CORAL

    Recolor the target domain's covariance to match the source. No target labels required.

  4. 04

    Residualize the dose

    Regress out concentration so the model reads which gas is present, not how much of it there is.

  5. 05

    Classify

    A Robust MLP with batch-norm and dropout, tuned by Optuna, reads the realigned signature.