Our team report on a hierarchically resistive skin (HR-skin) which can overcome the aforementioned challenges, offering the ability of “one resistive signal - multiple biometrics”. By integrating resistive skins with different tiers into one single sensor, we could achieve an unconventional staircase-like resistive responses to tensile stains. This HR-skin sensor, in conjunction with Machine learning, offers the unprecedented ability of identifying five physical/psychological activities (speak, heartrate, breath rate, touch, and neck movement) from a single electrical signal.
Regarding the neural network, we initially trained two different 4-layer fully connected neural networks for classification.
The two networks have the same numbers of hidden layers and neurons, but the different parameters, because one is based-on raw signals, named time-domain neural network, and the other employed FFT responses, called frequency-domain neural network.
Next, based on two pretrained networks, we trained the third fully connected neural network, taking the input of concatenated outputs from the second-to-last layers from these two networks.
We used stft to extract the frequency features from raw signals, and used some filters for detecting two rates.
The network aims to classify 11 classes of throat activities, including
speaking yes/no/one/two, nodding/shaking/stretch, nodding+yes/nodding+no/shaking+yes/shaking+no.
Under non-cross-person experiments, the classification accuracies of the time-domain model and frequency-domain model could reach up to 76.28% ±1.25% and 89.57% ±2.26% respectively. The final classification accuracy of the fused model could reach up to 92.73% ±0.82% for 11 throat activities, and 82.68% ±1.44% on gender-balanced dataset. For personalization, accuracy can reach up tp 84.21% ±5.82% after fine-tuning the third MLP.
The folder "live demo" is only a toy model for live demonstration.
The paper "A Hierarchically Resistive Skin as Intelligent Multimodal On-Throat Wearable Biosensors" is published in Nature Nanotechnology.