Hopfield Network Enabled Embroidered Textile Electrical Resistance Tomography

Among tactile sensors, embroidered textiles showcase high versatility and easy implementation in everyday fabric. However, it is complex to implement multiple sensing units, especially if using a single textile layer, due to cross-noise and cross-talk. A solution to this problem is electric resistance tomography (ERT), an acquisition technique that can record high quantities of tactile data from a substrate with delocalized electrodes. However, an analytical reconstruction of the mechanical interaction is extremely difficult, and machine learning algorithms are commonly used to achieve localization or classification of the tactile stimuli. We propose a system composed of a tactile sensor manufactured by embroidering electrodes in a lesser conductive substrate and a Hopfield network to process the data for the classification of mechanical stimuli. Whilst the embroidered electrodes are used to perform ERT, resulting in the first reported case of textile ERT, the Hopfield network achieves good separation of labeled clusters using a significantly smaller training dataset than any other machine learning approach. The device has been tested using a set of calibrated weights. Results show that the presence of embroidered electrodes and the Hopfield network both contribute to a much higher Silhouette score, given a limited number of stored stimuli.