A Convolutional Neural Network for Classification of Stimuli Based on Stretchable Mechanical Sensor

While various commercialized wearable devices are being released for measuring physiological signals, the issue of signal distortion caused by motion artifacts (MAs) has not been fully resolved yet. Recently, for measuring physiological signals, stretchable sensors have been widely adopted by attaching them to the skin. If we have information indicating that the measured signal is affected by MA caused by stretching or pressure, we can develop various strategic algorithms for MA cancellation. In this study, we presented a convolutional neural network (CNN) architecture to classify the strain and pressure measured by stretchable sensors. We fabricated a resistive mechanical sensor composed of a percolated network of Ag nanowires and polybutadiene-based urethane. We used a stretchable sensor to measure the current change: 10%, 20%, and 30% stretching along with 0.5-, 1.0-, 1.5-, 2.0-, and 2.5-N pressing. To classify the eight classes, we used one cycle of the current change as an image, which was fed into a convolutional layer. In addition, we considered additional information regarding the signal amplitude and time width, which were concatenated with the output from the last convolutional layer. We evaluated the classification performance using cross-validation followed by a testing dataset. The testing data results showed that our proposed CNN architecture provided high-accuracy metrics of sensitivity (96.26%), specificity (99.54%), accuracy (96.92%), F1-score (96.35%), and area of under the receiver operating characteristic (0.9984). We believe that our CNN architecture can accurately identify the stimuli/deformation/distortion sources applied to stretchable sensors, such as strain and pressure.