基于注意力模型的多传感器人类活动识别

深度循环神经网络适用于处理时间序列数据, 然而循环神经网络特征提取能力差, 时间依赖关系挖掘不足。针对此问题, 提出了3种注意力机制和长短时记忆(LSTM)神经网络结合的模型用于人类活动识别问题, 并研究了3种注意力机制在不同数据集上单独及配合使用时对模型精度的影响。对于UCI_HAR数据集, 3种注意力LSTM模型准确率分别为94.13%、95.15%和94.81%,高于LSTM模型识别准确率93.2%。此外, 针对人类活动识别的传感器时间序列数据的标签特点, 提出将时间段分类任务转化为分割任务, 设计了2个基于分割任务的注意力门控循环单元(GRU)神经网络模型, Bahdanau注意力GRU模型在Skoda数据集和机会(Oppor)数据集准确率为84.61%和89.54%, 高于基准UNet模型的70.40%和88.51%。

Multi-sensor Human Activity Recognition Based on Attention Model

The deep recurrent neural network was suitable for processing time series data. However, the feature extraction ability of traditional recurrent neural network was poor, and the time dependency mining was insufficient. In response to the mentioned problem, three models of attention mechanism and long-term short-term memory (LSTM) neural network were proposed for human activity recognition application problems. The effects of these three mechanisms on the accuracy of the models were studied separately and in combination with different datasets. In the UCI_HAR data set, the accuracy rates of the three attention LSTM models were 94.13%,95.15% and 94.81%, respectively, which were higher than the identification accuracy rate of LSTM model (93.2%). In addition, for the label characteristics of sensor time series data for human activity recognition, it was proposed to convert the time segment classification task into a segmentation task. Therefore, two attention-based gate recurrent unit(GRU) models based on segmentation tasks were designed. The accuracy of Bahdanau attention GRU model were 84.61% and 89.54% in the Skoda data set and the opportunity data set which were higher than the benchmark UNet model’s 70.40% and 88.51%.