Our work focuses on the exploration of the internal relationships of signals in an individual sensor. In particular, we address the problem of not being able to evaluate such intrasensor relationships due to missing rich and explicit feature representation. To solve this problem, we propose GraphSensor, a graph attention network, with a shared-weight convolution feature encoder to generate the signal segments and learn the internal relationships between them. Furthermore, we enrich the representation of the features by utilizing a multi-head approach when creating the internal relationship graph. Compared with traditional multi-head approaches, we propose a more efficient convolution-based multi-head mechanism, which only requires $56\%$ of model parameters compared with the best multi-head baseline as demonstrated in the experiments. Moreover, GraphSensor is capable of achieving state-of-the-art performance in the electroencephalography dataset and improving the accuracy by $13.8\%$ compared to the best baseline in an inertial measurement unit (IMU) dataset.

This work develops a generalizable neural network, SENSORNET, for sensor feature learning across various applications. The primary challenge addressed is the poor portability of pretrained neural networks to new applications with limited sensor data. To solve this challenge, we design SensorNet, which integrates the flexibility of self-attention with multi-scale feature locality of convolution. Moreover, we invent patch-wise self-attention with stacked multi-heads to enrich the sensor feature representation. SensorNet is generalizable to pervasive applications with any number of sensor inputs, and is much smaller than the state-of-the-art self-attention and convolution hybrid baseline ($0.83$ M vs. $3.87$ M parameters) with similar performance. The experimental results show that SensorNet is able to achieve state-of-the-art performance compared with the top five models on a competition activity recognition dataset (SHL’18). Moreover, pretrained SensorNet in a large inertial measurement unit (IMU) dataset can be fine-tuned to achieve the best accuracy on a much smaller IMU dataset (up to $5\%$ improvement in WISDM) and to achieve the state-of-the-art performance on an EEG dataset (SLEEP-EDF-20), showing the strong generalizability of our approach.