Miscellaneous mini-wearable devices (e.g. wristbands, smartwatches, armbands) have emerged in our life, capable of recognizing activities of daily living, monitoring health information, so on. Conventional activity recognition (AR) models deployed inside these devices are generic classifiers learned offline from abundant data. Transferring generic model to user-oriented model requires time-consuming human effort for annotations. To solve this problem, we propose SS-ARTMAP-AR, a self-supervised incremental learning AR model updated from surrounding information such as Bluetooth, Wi-Fi, GPS, GSM data without user's annotation effort. Experimental results show that SS-ARTMAP-AR can gradually adapt individual users, become more incremental intelligence.

Deep learning-based methods demonstrate promising results in continuous non-invasive blood pressure measurement, whereas those models trained on large public datasets suffer from severe performance degradation in predicting from real-world user data collected in home settings. Transfer learning has been recently introduced to personalize the pre-trained model with unseen users' data to solve the problem. However, the existing methods based on network fine-tuning for model personalization require a large amount of labeled data, lacking practicality due to labeling using a cuff-based blood pressure monitor is extremely tedious and laborious for home users. In this paper, we propose a novel few-shot transfer learning approach named FewShotBP, which addresses the above-mentioned challenges by introducing a personalization adapter at the personalization stage (i.e., the transfer learning stage), and a multi-modal spectro-temporal neural network at the pre-train stage, to bridge the gap between data-hungry models and limited labeled data in realistic scenarios. To evaluate the approach's significance, we conducted experiments using both a publicly available dataset and a real-world user experiment. The results demonstrated that the proposed approach achieves similar accuracy of blood pressure prediction with 10× less data for personalization compared with the state-of-the-art method in the public dataset and achieves a mean absolute error of 6.68 mmHg (systolic blood pressure) and 3.91 mmHg (diastolic blood pressure) with only 10 personal data samples in the real-world user experiment.

For low-semantic sensor signals from human activity recognition (HAR), contrastive learning (CL) is essential to implement novel applications or generic models without manual annotation, which is a high-performance self-supervised learning (SSL) method. However, CL relies heavily on data augmentation for pairwise comparisons. Especially for low semantic data in the HAR area, conducting good performance augmentation strategies in pretext tasks still rely on manual attempts lacking generalizability and flexibility. To reduce the augmentation burden, we propose an end-to-end auto-augmentation contrastive learning (AutoCL) method for wearable-based HAR. AutoCL is based on a Siamese network architecture that shares the parameters of the backbone and with a generator embedded to learn auto-augmentation. AutoCL trains the generator based on the representation in the latent space to overcome the disturbances caused by noise and redundant information in raw sensor data. The architecture empirical study indicates the effectiveness of this design. Furthermore, we propose a stop-gradient design and correlation reduction strategy in AutoCL to enhance encoder representation learning. Extensive experiments based on four wide-used HAR datasets demonstrate that the proposed AutoCL method significantly improves recognition accuracy compared with other SOTA methods.

Radar-based human activity recognition (HAR) is a time-series classification challenge, which is widely applied in intelligent healthcare, ubiquitous computing, autonomous driving, etc. However, the received radar signal is much noisier and less intuitive than images, which makes it harder to capture the human posture information and achieve a high recognition accuracy. In this paper, we use a radar array to collect activity data from different subjects and propose a novel simultaneous spatial-temporal information encoding network (SST-HAR) to extract the posture features from the raw radar signals. The spatial and temporal information are extracted in two branches respectively at the same time, and a lightweight external attention (EA) module allows for the extraction of more relevant features from the fused spatial-temporal information for radar-based HAR. We evaluate the SST-HAR method through extensive experiments and achieve a mean accuracy of 97.91% on our dataset and 94.12% on a public dataset, which shows the effectiveness of our method in radar-based HAR tasks.