Cross-Sensor Diffusion for Wearable Sensor Reconstruction

The problem

Human activity recognition over body-worn sensors assumes the sensors are all reporting. In practice they fail, fall off, or run out of battery — and a recognition pipeline trained on complete data degrades badly when they do. The question is whether the missing signal can be reconstructed well enough to keep downstream recognition reliable.

Approach

I architected a cross-sensor diffusion framework that reconstructs missing IMU signals from the sensors that are still reporting, and built a complete training and evaluation pipeline over 14 body-worn IMU sensors from the OPPORTUNITY dataset, using both locomotion and high-level activity labels.

Modeling realistic failure

Rather than dropping sensors at random, I designed three training regimes that model how sensing actually fails in the field:

• 35% whole-body-location dropout — an entire body location goes dark
• 40% single-sensor dropout — one sensor fails
• 25% random 2–3 sensor dropout — scattered simultaneous failures

Results

A DDPM-based reconstruction model feeding a downstream C-LSTM-A activity classifier significantly outperforms mean-fill imputation when the missing sensors are strongly correlated with the observed ones. Just as usefully, the experiments identify the regimes where that correlation is weak and simple statistical imputation remains competitive — diffusion is not uniformly the right answer, and knowing when it is not is part of the contribution.

In progress

Extending the evaluation to the CogAge and PAMAP2 datasets in collaboration with a visiting researcher.