中文摘要：

研究了输入是可穿戴传感器获得的多通道时间序列信号，输出是预定义的活动的活动识别模型，指出活动中的有效特征的提取目前多依赖于手工和浅层特征学习结构，不仅复杂而且会导致识别准确率下降；基于深度学习的卷积神经网络（CNN）不是对时间序列信号进行手工特征提取，而是自动学习最优特征；目前使用卷积神经网络处理有限标签数据仍存在过拟合问题。因此提出了一种基于融合特征的系统性的特征学习方法用于活动识别，用ImageNetl6对原始数据集进行预训练，将得到的数据与原始数据进行融合，并将融合数据和对应的标签送入有监督的深度卷积神经网络（DCNN）中，训练新的系统。在该系统中，特征学习和分类是相互加强的，它不仅能处理端到端的有限数据问题，也能使学习到的特征有更强的辨别力。与其他方法相比，该方法整体精度从87．0％提高到87．4％。

英文摘要：

An activity recognition model, with its input being the multi-channel time series signals obtained by wearable sensors and output being a predefined activity, was studied. It was pointed that extracting effective features from ac- tivity is a key in activity recognition. Most of the existing work relies on manual extraction of the features and the shallow learning structure, which makes the work complex and the recognition unaccurate. However, the convolu- tional neural network （CNN） based on deep learning does not manually extract the time series signals, but auto- matically learns the best feature. At present, using convolutional neural network to process limited labeled data still has the overfitting problem. Therefore, a systematic feature learning method based on fusion characteristics was presented for activity recognition. The method uses the ImageNetl6 to pre-train the original data set to fuse the ob- tained data with the original data, and puts the fused data and the corresponding tag into a supervised depth convo- lutional neural network （DCNN） to train the new system. In this system, the characteristics of learning and classi- fication are mutually reinforcing, which can not only deal with the problem of limited data from end to end, but also make the learning more discriminative. Compared with other methods, the overall accuracy of the proposed method is increased from 87% to 87.4%.