← All work

Sense

Independent research on keeping electronic noses accurate as their sensors drift.

Role
Independent researcher
Timeline
Jul 2026 – present
Stack
Python, scikit-learn, Continual learning

The problem

Electronic noses could catch gas leaks, spoiled food or disease markers, but only if they stay accurate over time. Their gas sensors drift: the same gas reads differently months later.

On the public UCI Gas Sensor Array Drift dataset, 36 months of readings from 16 sensors in ten time-ordered batches, six gases are about 99% separable within a single batch. A model trained on the first batch and left alone falls to about 47% on the batches that follow.

What I've built so far

A clean, reproducible baseline on that public dataset, so the work needs no special hardware to start. It measures how badly drift hurts a fixed model, and reproduces two standard drift-correction methods, CORAL and DRCA, as honest comparison points before trying anything new.

Key decisions

Keep time in order. The batches stay in sequence, so drift is measured the way a deployed sensor would meet it, as one task after another, never as a shuffled split.

Tune only on what you'd really have. The comparison methods' settings are chosen using the first batch's labels alone, never the later batches'.

What's next

Two directions. Continual learning, so the model keeps up as new batches arrive. And open-set recognition, so it can tell when it smells a gas it has never seen.