Abstract:

Objective Flexible fabric pressure sensor arrays have great application potential in detecting and distinguishing the physical state of the elderly, because this technology offers high detection sensitivity, wide detection range and simple fabrication process. However, in the process of fabricating the piezoresistive flexible pressure sensor array, it is difficult to avoid the occurrence of abnormal sensing units and crosstalk, which results in the collection of inaccurate pressure data. In particular, the crosstalk phenomenon of the flexible pressure sensor array will cause serious distortion of the pressure contour generated based on the collected pressure data. The aim of this work is to devise a method for solving the problems mentioned above.

Method In order to solve these problems, a piezoresistive flexible fabric pressure sensor array system with 32×32 sensing units is first constructed. The causes for the detected abnormal values of the sensing units are then analyzed and the median filtering algorithm is adopted to process the abnormal values in the pressure distribution contour obtained by the system. Aimed at solving the crosstalk phenomenon, a U-Net convolutional neural network model is constructed and the pressure contour generated by the fabric pressure sensor array system is corrected by this machine learning method. In addition, the method for collecting the input and output data set for the model is designed.

Results The results show that the peak signal-to-noise ratio is taken as the benchmark to evaluate the quality of the processed pressure contour image in view of the detected abnormal values of the sensor units. Generally, when the PSNR of the image is greater than 30 dB, the image quality is good. In this paper, the pressure contour of three states, namely, pressure generated with no object, by a cylindrical water bottle, and by cylindrical small medicine bottle, are selected. After the three pressure contours with abnormal values are processed by the median filter algorithm, the calculated values of the peak signal-to-noise ratio are all between 30-40 dB for the three states. The actual processing results reflect the effectiveness of the median filter algorithm. For the crosstalk problem, after 4 200 iterations of training using the input and output pressure contour data sets, the root-mean-square error of the U-Net convolutional neural network model is decreased from over 100 to 7.108, reaching convergence. Under the same pressure exertion conditions by placing palm, foot, cylindrical bottle and rectangular stick on the sensor array, it is shown that the U-Net model can eliminate more effectively the influence of crosstalk by the pressure contours of the fabric pressure sensor array than by the flexible pressure sensor array without crosstalk effect, in terms of the corrected pressure contour.

Conclusion The median filtering algorithm can effectively eliminate the the abnormally detected values by a sensing unit in the flexible fabric pressure sensor array, and the peak signal-to-noise ratio of the pressure contour presented by the flexible fabric pressure sensor array is greater than 30 dB, ensuring contour images with better quality. This also help reduce the manufacturing requirements for the flexible fabric pressure sensor array. The U-Net convolutional neural network can solve the crosstalk problem of the flexible fabric pressure sensor array and eliminate the crosstalk effect while allowing the use of a simple voltage divider circuit to measure the resistance value of the sensing unit.

Key words: fabric pressure sensor array, median filtering, U-Net convolutional neural network, plantar pressure monitoring, crosstalk