Work

Benchmarking the Robustness of Wearable HAR Pipelines Under Graded Sensor Noise and Missing-Data Corruption

Year: 2026
Type: conference-paper
Abstract: Sensor-based Human Activity Recognition (HAR) is now routinely reported above 95% accuracy on curated benchmarks, but the behaviour of the same pipelines under realistic data-quality degradation is much less well characterised. Previous studies typically address either sensor noise or missing values in isolation, which leaves the relative severity of the two corruption modes unquantified within a single experimental protocol. This paper presents a controlled, severity-graded robustness benchmark applied to the UCI-HAR dataset. A Random Forest classifier is trained on clean data and evaluated against test sets corrupted in two ways: additive Gaussian noise with standard deviation sigma in \{0.1, 0.3, 0.5, 1.0\} and random feature dropout with rates p in \{10%, 20%, 30%, 50%\}, the latter restored by mean imputation and, for comparison, by KNN imputation. Each condition is repeated for 30 random seeds, and pairwise differences are tested with the Wilcoxon signed-rank test under Bonferroni correction. Noise damages the classifier significantly more than missing data at every severity level: sigma = 1.0 reduces accuracy from 92.91% to 45.21%, whereas 30% missing values still allow 87.43% (mean imputation) or 92.33% (KNN). The same asymmetry is observed with a Multi-Layer Perceptron baseline, which indicates that it is not an artefact of the tree ensemble. To our knowledge, this is the first study to place sensor noise and missing-data corruption on a common, severity-graded footing within a single fixed HAR pipeline and to confirm the resulting asymmetry with paired statistical testing. The results provide quantitative guidance for allocating preprocessing effort in HAR systems where noise and intermittent dropout occur as distinct failure modes.
Language: English
Open Access status: hybrid
OpenAlex ID: W7204136385
FWCI:
Cites: 15
Cited by:
Related to: 10