Signal-Quality-Aware Multimodal Fusion for Reliable Context Recognition at the Edge: Code and Experimental Results

Curated code and aggregated experimental results for a study on signal-quality-aware multimodal fusion for human activity recognition (HAR) at the edge. The system estimates per-modality signal quality/reliability from sensor data that is present but degraded (noise, bias, mis-orientation, clipping, sample drop, cross-modal conflict) and evaluates whether quality-gated fusion improves robustness, calibration, and selective-prediction/abstention compared to fixed and attention-based fusion baselines. The package includes: the core fusion model implementations and metrics (macro-F1, ECE, NLL, Brier, quality/error AUROC, risk-coverage), the experiment-driver scripts used to run the degradation benchmark and multi-seed evaluations across six public HAR benchmarks (HHAR, WISDM, MotionSense, MHEALTH, PAMAP2, UCI HAR, OPPORTUNITY), and the aggregated, paper-ready result tables (CSV/JSON/Markdown/SVG) that the reported metrics are computed from. Raw third-party sensor datasets and per-seed model checkpoints are not redistributed; see README.md for pointers to the original data sources and instructions to reproduce a run from the included configs. See METHODOLOGY_AND_RESULTS.md for the full protocol, baseline ladder, and headline results tables. Submitted anonymously for review. Provided for reference/review; redistribution of modified versions and commercial use are not permitted (see license).