传感技术学报2026，Vol.39Issue(1)：58-65,8.DOI:10.3969/j.issn.1004-1699.2026.01.008
一种基于Attention-TCN的跌倒预测算法
A Fall Prediction Algorithm Based on Attention-TCN
摘要
Abstract
With the intensifying global aging population,fall events among the elderly are happening increasingly,posing a serious threat to their physical and mental health.Although many studies have been devoted to fall detection and prediction,most of them primarily analyze the temporal features of the time-series data acquired by inertial sensors,while the study of spatial features is less.In order to predict fall events among the elderly more effectively and activate protective devices such as airbags promptly,a new fall prediction algo-rithm based on Attention-TCN is proposed,which combines the attention mechanism and temporal convolutional networks(TCN),capa-ble of extracting global spatial-temporal features from fall time-series data and autonomously fusing the features through the adaptive fea-ture fusion(AFF)method to provide accurate fall classification.Meanwhile,a downsampling technique is used to reduce the model size and inference time while improving the model prediction performance.In offline experiments conducted on the SisFall public fall dataset on PC,the method achieves an accuracy of 98.67%,a sensitivity of 98.89%,and a specificity of 98.52%,with an average fall prediction leading time of 221.16 ms,a model inference time of 0.19±0.05 ms,and a model size of 673 KB,which confirms its high efficiency and practicality in predicting falls among the elderly.关键词
深度学习/跌倒预测/时序卷积网络/注意力机制/惯性传感器Key words
deep learning/fall prediction/temporal convolutional networks/attention mechanisms/inertial sensors分类
信息技术与安全科学引用本文复制引用
王宏宇,潘巨龙,周辰,宋炜..一种基于Attention-TCN的跌倒预测算法[J].传感技术学报,2026,39(1):58-65,8.基金项目
浙江省教育厅科研项目(专业学位研究生专项)(Y202456356) （专业学位研究生专项）
