Hierarchical Knowledge Distillation for Human Activity Recognition
Di Hu, Xingchen Wu, Dezhu Xiong · Advances in transdisciplinary engineering · 2025
The widespread adoption of IoT and mobile devices with a variety of sensors has made it possible to develop applications that combine time-series data from multiple sensor inputs. However, current deep neural network architectures for multimodal fusion struggle with consecutive missing data and noise, common in real-world scenarios. We introduce a Hierarchical Knowledge Distillation Framework (HKDF) designed to address noise and data missingness challenges in Human Activity Recognition (HAR). Specifically, we introduce a cross-sample correlation distillation mechanism, leveraging contrastive learning to capture and transfer holistic semantics across samples, thereby enabling effective feature reconstruction. Additionally, a cross-category variation distillation strategy is proposed to learn inter- and intra- category dynamics based on category prototypes to produce robust joint multimodal representations. The superior performance achieved on both datasets demonstrates the remarkable advantages of HKDF.