基于LSTM模型的人体情景多标签识别研究

Multi-Label Classification for Human Context Recognition with LSTM Model

针对人体情景识别问题,本文提出了一种以多模态传感器数据为特征的多标签分类问题解决方案。以z-score实现传感器数据的标准化,根据时序相邻原则及多元线性回归,实现对缺失数据的填充,利用情景标签间的互斥和依赖关系过滤无效标签,通过LSTM模型实现从传感器数据到标签的非线性映射,同时在损失函数中引入代价矩阵,解决标签数据不平衡问题。与之前建立的模型相比,LSTM模型具有长期记忆的特性,并在原数据集上分别考察了模型的精度、召回率、特异度及均衡精度,各项识别指标较前人模型有明显提升。研究结果表明,LSTM模型利用记录间的时序信息提高了识别性能,具有广泛适用性。该研究揭示了时序关系在人体情景识别中的重要性,为进一步提升识别性能指明了方向。

Aiming at human context recognition, this paper proposes a multi-label classification solution with the feature of multi-modal sensor data. We use z-score to accomplish the standardization of sensor data, and fill up the missing data by the principle of neighboring relationship or multivariable linear regression, filter out invalid labels by mutual exclusion and dependency between labels, and actualize the nonlinear mapping from sensor data to labels through LSTM model. We also introduce a cost matrix into the loss function, which solves the problem of label data imbalance. Compared with the former model, the long-term memory is a characteristic feature of LSTM model. The accuracy, sensitivity, specificity and balance accuracy of the model are investigated on the original dataset. These indicators are significantly improved compared with the former models. The results show that the LSTM model improves the recognition performance by using the sequence information between records. The timing relationship is ubiquitous in sensor data, and the research findings can be widely adopted. The study also reveals the importance of sequence relationships in human context recognition, and indicates the direction for improving recognition performance.