.png){kind=icon}
.png){kind=icon}
Return
A Pre-impact Fall Algorithm Based on a Lightweight Re-Parameters-Parallel Convolutional-TCN
.png){kind=icon}
PRE
AI
DOI：10.13164/re.2026.0084.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the aging society intensifying, the problem of elderly falls has become a key issue of social concern. Research on fall prediction based on Internet of Things (IoT) technology has received widespread attention. To effectively predict fall events, a lightweight IoT-based fall prediction model called lwRPPC-TCN (lightweight Re-Parameters-Parallel-Convolutional Temporal Convolutional Network) is proposed. The model utilizes the temporal data collected by IoT sensors in the input stage and achieves efficient decoupled extraction of temporal and spatial features through lwRPPC blocks. The subsequent Temporal Convolutional Networks (TCNs) further strengthens the ability of modeling the global temporal dependency, thus optimizing the processing capability of sensor time-series data. To validate the generalization ability of the model and mitigate fall data scarcity, two public datasets, SisFall and KFall, are fused, and the performance of the model is evaluated by five-fold cross-validation. In addition, a homogeneous (models belong to the same model family) knowledge distillation technique is introduced to improve the performance of the model. Experimental results demonstrate that the proposed lwRPPC-TCN achieves an accuracy of 98.88% on the fused dataset, outperforming existing fall prediction models, with a fall prediction lead time (interval between the fall prediction time and the collision time) of 250 ms, and a compact model size of 60 KB, which makes it suitable and possible to deploy in a resource-constrained wearable device.
Keywords:
Deep learning
fall prediction
knowledge distillation
Re-Parameters-Parallel-Convolutional-Temporal Con-volutional Network
inertial sensor
spatio-temporal feature decoupling
Journal
R
IF:
0.7
Papers:
37
Citations:
994