Multi-channel Time Series Decomposition Network For Generalizable Sensor-Based Activity Recognition

TL;DR

A model based on time series decomposition is proposed, which helps the model learn universal features and effectively improve its generalization ability, and can be further extended to time series tasks with varying domain distributions.

Abstract

Sensor-based human activity recognition is important in daily scenarios such as smart healthcare and homes due to its non-intrusive privacy and low cost advantages, but the problem of out-of-domain generalization caused by differences in focusing individuals and operating environments can lead to significant accuracy degradation on cross-person behavior recognition due to the inconsistent distributions of training and test data. To address the above problems, this paper proposes a new method, Multi-channel Time Series Decomposition Network (MTSDNet). Firstly, MTSDNet decomposes the original signal into a combination of multiple polynomials and trigonometric functions by the trainable parameterized temporal decomposition to learn the low-rank representation of the original signal for improving the extraterritorial generalization ability of the model. Then, the different components obtained by the decomposition are classified layer by layer and the layer attention is used to aggregate components to obtain the final classification result. Extensive evaluation on DSADS, OPPORTUNITY, PAMAP2, UCIHAR and UniMib public datasets shows the advantages in predicting accuracy and stability of our method compared with other competing strategies, including the state-of-the-art ones. And the visualization is conducted to reveal MTSDNet's interpretability and layer-by-layer characteristics.

Figures (8)

• Figure 1: Illustration of the proposed method. Firstly, data is collected through sensors such as accelerometers and gyroscopes. After data preprocessing, denoising and slicing, input data is decomposed into multiple components by decomposer. Each component can be considered as a multiplication combination of preterm coefficients and constraint terms. The orange part indicates that the classifier use the preterm coefficients of each component. Finally, MTSDNet integrates classification results using a layer attention. Here it adopts the structure of the model MTSDNet-A-tsg, which is an additive model and decompose the signal to one trend, one seasonal and one general component. More details can be found in Section \ref{['Overall Structure']}.
• Figure 2: The overall structure of proposed MTSDNet according to the additive model with three layers and no decomposition term output in the last layer. The model input is the sensor data divided into many sliding windows and the output is a weighted classification results of multiple layers.
• Figure 3: The illustration of decomposer structure with one input and two outputs. The left and right sides output the decomposition term and classification results, respectively. The dashed line indicates that the data is used as parameter for the function instead of function input.
• Figure 4: Violin plot of accuracy distribution on the DSADS dataset. MTSDNet shown in blue and orange achieves higher accuracy than other baselines on almost all Parts.
• Figure 5: Violin plot of accuracy distribution of UniMib dataset. MTSDNet shown in blue and orange achieves higher accuracy than other baselines on 3 parts and also has higher stability.