Graceful Degradation by Design for Sensor-Aware Inference in Resilient AIoT Human Activity Recognition

Artificial Intelligence of Things (AIoT) systems often assume that every sensing channel remains available, although deployed devices can lose signals through faults, power management, poor contact, or device removal. This study compares five inference strategies under changing sensor availability. The UCI Human Activity Recognition Using Smartphones (HAR) dataset provides all 64 availability states of six inertial channels. When one channel remains, fixed complete-input models average about 23% accuracy, compared with 59.3% for mask-aware inference and 67.6% for specialized fallback. At a 50% independent channel-failure probability, sensor-aware strategies retain 75.8–78.9% expected accuracy, compared with 42.5–45.7% for fixed models. An ablation separates the effects of replacement values and availability indicators. A five-channel configuration selected using training subjects reaches 86.2% accuracy versus 85.8% for all six on the official test partition also used for exhaustive sensor-state evaluation. In MHEALTH leave-one-subject-out sensitivity analysis, complete three-location sensing averages 86.2%, compared with 77.5% for chest-only sensing. The findings demonstrate improved robustness under missing sensing and substantial variation in the usefulness of individual sensors.