核心概念
A novel variational encoder-decoder neural network approach is presented for removing interference motions from Doppler radar signals to enable accurate extraction of vital signs.
要約
The paper addresses the challenge of extracting vital sign information, such as respiration rate and heart rate, from Doppler radar signals in the presence of interfering body motions. A key problem is that the vital sign contributions and the interference motions can have overlapping Doppler bands and significant differences in power levels, making them difficult to separate using traditional signal processing techniques.
The proposed solution utilizes a variational convolutional neural network (CNN) as an interference removal system. The network is trained on a semi-experimental dataset, where the vital sign STFT (Short-Time Fourier Transform) signals are derived from real recordings and the interference motion is synthesized according to a widely applied model.
The results demonstrate that the network is capable of generalizing beyond the training dataset and can produce outputs with low reconstruction loss for unseen samples. Furthermore, the network's output is shown to enable more accurate extraction of the respiration rate compared to the original mixed signal, even in the presence of high levels of interference and additive noise.
The key highlights of the proposed approach are:
- It operates solely on single-channel complex radar signals, making it flexible for a wide range of applications.
- The probabilistic nature of the variational architecture allows the model to learn more realistic representations of the signal.
- The framework can be adjusted by controlling the datasets used for training, enabling application to various interference removal scenarios.
統計
The radar data within the dataset is supplied in quadrature format, with separate traces for I and Q signals.
The power ratio between the vital sign signal and the simulated interference motion is controlled by first normalising the power of both signals to achieve zero-mean and unit-variance, followed by scaling of the interference signal to match a given Signal-to-Interference Ratio (SIR) ratio.
引用
"The contribution presented in this work constitutes the first application of a variational encoder-decoder neural network for interference removal from vital signs radar returns."
"The probabilistic nature of the architecture allows the model to learn a more realistic representations of the signal, where a given input is matched with a distribution of possible satisfactory output samples rather than a single sample."